Skip to content

GitLab

Commit 80444539 authored by Zhuoran Liu's avatar Zhuoran Liu Committed by pkulzc
Browse files

Add TPU SavedModel exporter and refactor OD code (#6737) 

247226201  by ronnyvotel:

    Updating the visualization tools to accept unique_ids for color coding.

--
247067830  by Zhichao Lu:

    Add box_encodings_clip_range options for the convolutional box predictor (for TPU compatibility).

--
246888475  by Zhichao Lu:

    Remove unused _update_eval_steps function.

--
246163259  by lzc:

    Add a gather op that can handle ignore indices (which are "-1"s in this case).

--
246084944  by Zhichao Lu:

    Keras based implementation for SSD + MobilenetV2 + FPN.

--
245544227  by rathodv:

    Add batch_get_targets method to target assigner module to gather any groundtruth tensors based on the results of target assigner.

--
245540854  by rathodv:

    Update target assigner to return match tensor instead of a match object.

--
245434441  by Zhichao Lu:

    Add README for tpu_exporters package.

--
245381834  by lzc:

    Internal change.

--
245298983  by Zhichao Lu:

    Add conditional_shape_resizer to config_util

--
245134666  by Zhichao Lu:

    Adds ConditionalShapeResizer to the ImageResizer proto which enables resizing only if input image height or width is is greater or smaller than a certain size. Also enables specification of resize method in resize_to_{max, min}_dimension methods.

--
245093975  by Zhichao Lu:

    Exporting SavedModel for Object Detection TPU inference. (faster-rcnn)

--
245072421  by Zhichao Lu:

    Adds a new image resizing method "resize_to_max_dimension" which resizes images only if a dimension is greater than the maximum desired value while maintaining aspect ratio.

--
244946998  by lzc:

    Internal Changes.

--
244943693  by Zhichao Lu:

    Add a custom config to mobilenet v2 that makes it more detection friendly.

--
244754158  by derekjchow:

    Internal change.

--
244699875  by Zhichao Lu:

    Add check_range=False to box_list_ops.to_normalized_coordinates when training
    for instance segmentation.  This is consistent with other calls when training
    for object detection.  There could be wrongly annotated boxes in the dataset.

--
244507425  by rathodv:

    Support bfloat16 for ssd models.

--
244399982  by Zhichao Lu:

    Exporting SavedModel for Object Detection TPU inference. (ssd)

--
244209387  by Zhichao Lu:

    Internal change.

--
243922296  by rathodv:

    Change `raw_detection_scores` to contain softmax/sigmoid scores (not logits) for `raw_ detection_boxes`.

--
243883978  by Zhichao Lu:

    Add a sample fully conv config.

--
243369455  by Zhichao Lu:

    Fix regularization loss gap in Keras and Slim.

--
243292002  by lzc:

    Internal changes.

--
243097958  by Zhichao Lu:

    Exporting SavedModel for Object Detection TPU inference. (ssd model)

--
243007177  by Zhichao Lu:

    Exporting SavedModel for Object Detection TPU inference. (ssd model)

--
242776550  by Zhichao Lu:

    Make object detection pre-processing run on GPU.  tf.map_fn() uses
    TensorArrayV3 ops, which have no int32 GPU implementation.  Cast to int64,
    then cast back to int32.

--
242723128  by Zhichao Lu:

    Using sorted dictionaries for additional heads in non_max_suppression to ensure tensor order

--
242495311  by Zhichao Lu:

    Update documentation to reflect new TFLite examples repo location

--
242230527  by Zhichao Lu:

    Fix Dropout bugs for WeightSharedConvolutionalBoxPred.

--
242226573  by Zhichao Lu:

    Create Keras-based WeightSharedConvolutionalBoxPredictor.

--
241806074  by Zhichao Lu:

    Add inference in unit tests of TFX OD template.

--
241641498  by lzc:

    Internal change.

--
241637481  by Zhichao Lu:

    matmul_crop_and_resize(): Switch to dynamic shaping, so that not all dimensions are required to be known.

--
241429980  by Zhichao Lu:

    Internal change

--
241167237  by Zhichao Lu:

    Adds a faster_rcnn_inception_resnet_v2 Keras feature extractor, and updates the model builder to construct it.

--
241088616  by Zhichao Lu:

    Make it compatible with different dtype, e.g. float32, bfloat16, etc.

--
240897364  by lzc:

    Use image_np_expanded in object_detection_tutorial notebook.

--
240890393  by Zhichao Lu:

    Disable multicore inference for OD template as its not yet compatible.

--
240352168  by Zhichao Lu:

    Make SSDResnetV1FpnFeatureExtractor not protected to allow inheritance.

--
240351470  by lzc:

    Internal change.

--
239878928  by Zhichao Lu:

    Defines Keras box predictors for Faster RCNN and RFCN

--
239872103  by Zhichao Lu:

    Delete duplicated inputs in test.

--
239714273  by Zhichao Lu:

    Adding scope variable to all class heads

--
239698643  by Zhichao Lu:

    Create FPN feature extractor for object detection.

--
239696657  by Zhichao Lu:

    Internal Change.

--
239299404  by Zhichao Lu:

    Allows the faster rcnn meta-architecture to support Keras subcomponents

--
238502595  by Zhichao Lu:

    Lay the groundwork for symmetric quantization.

--
238496885  by Zhichao Lu:

    Add flexible_grid_anchor_generator

--
238138727  by lzc:

    Remove dead code.

    _USE_C_SHAPES has been forced True in TensorFlow releases since
    TensorFlow 1.9
    (https://github.com/tensorflow/tensorflow/commit/1d74a69443f741e69f9f52cb6bc2940b4d4ae3b7)

--
238123936  by rathodv:

    Add num_matched_groundtruth summary to target assigner in SSD.

--
238103345  by ronnyvotel:

    Raising error if input file pattern does not match any files.
    Also printing the number of evaluation images for coco metrics.

--
238044081  by Zhichao Lu:

    Fix docstring to state the correct dimensionality of `class_predictions_with_background`.

--
237920279  by Zhichao Lu:

    [XLA] Rework debug flags for dumping HLO.

    The following flags (usually passed via the XLA_FLAGS envvar) are removed:

      xla_dump_computations_to
      xla_dump_executions_to
      xla_dump_ir_to
      xla_dump_optimized_hlo_proto_to
      xla_dump_per_pass_hlo_proto_to
      xla_dump_unoptimized_hlo_proto_to
      xla_generate_hlo_graph
      xla_generate_hlo_text_to
      xla_hlo_dump_as_html
      xla_hlo_graph_path
      xla_log_hlo_text

    The following new flags are added:

      xla_dump_to
      xla_dump_hlo_module_re
      xla_dump_hlo_pass_re
      xla_dump_hlo_as_text
      xla_dump_hlo_as_proto
      xla_dump_hlo_as_dot
      xla_dump_hlo_as_url
      xla_dump_hlo_as_html
      xla_dump_ir
      xla_dump_hlo_snapshots

    The default is not to dump anything at all, but as soon as some dumping flag is
    specified, we enable the following defaults (most of which can be overridden).

     * dump to stdout (overridden by --xla_dump_to)
     * dump HLO modules at the very beginning and end of the optimization pipeline
     * don't dump between any HLO passes (overridden by --xla_dump_hlo_pass_re)
     * dump all HLO modules (overridden by --xla_dump_hlo_module_re)
     * dump in textual format (overridden by
       --xla_dump_hlo_as_{text,proto,dot,url,html}).

    For example, to dump optimized and unoptimized HLO text and protos to /tmp/foo,
    pass

      --xla_dump_to=/tmp/foo --xla_dump_hlo_as_text --xla_dump_hlo_as_proto

    For details on these flags' meanings, see xla.proto.

    The intent of this change is to make dumping both simpler to use and more
    powerful.

    For example:

     * Previously there was no way to dump the HLO module during the pass pipeline
       in HLO text format; the only option was --dump_per_pass_hlo_proto_to, which
       dumped in proto format.

       Now this is --xla_dump_pass_re=.* --xla_dump_hlo_as_text.  (In fact, the
       second flag is not necessary in this case, as dumping as text is the
       default.)

     * Previously there was no way to dump HLO as a graph before and after
       compilation; the only option was --xla_generate_hlo_graph, which would dump
       before/after every pass.

       Now this is --xla_dump_hlo_as_{dot,url,html} (depending on what format you
       want the graph in).

     * Previously, there was no coordination between the filenames written by the
       various flags, so info about one module might be dumped with various
       filename prefixes.  Now the filenames are consistent and all dumps from a
       particular module are next to each other.

    If you only specify some of these flags, we try to figure out what you wanted.
    For example:

     * --xla_dump_to implies --xla_dump_hlo_as_text unless you specify some
       other --xla_dump_as_* flag.

     * --xla_dump_hlo_as_text or --xla_dump_ir implies dumping to stdout unless you
       specify a different --xla_dump_to directory.  You can explicitly dump to
       stdout with --xla_dump_to=-.

    As part of this change, I simplified the debugging code in the HLO passes for
    dumping HLO modules.  Previously, many tests explicitly VLOG'ed the HLO module
    before, after, and sometimes during the pass.  I removed these VLOGs.  If you
    want dumps before/during/after an HLO pass, use --xla_dump_pass_re=<pass_name>.

--
237510043  by lzc:

    Internal Change.

--
237469515  by Zhichao Lu:

    Parameterize model_builder.build in inputs.py.

--
237293511  by rathodv:

    Remove multiclass_scores from tensor_dict in transform_data_fn always.

--
237260333  by ronnyvotel:

    Updating faster_rcnn_meta_arch to define prediction dictionary fields that are batched.

--

PiperOrigin-RevId: 247226201
parent c4f34e58
Hide whitespace changes
Inline Side-by-side
Showing with 3394 additions and 297 deletions
research/object_detection/core/target_assigner_test.py
View file @ 80444539
......@@ -713,9 +713,6 @@ class BatchTargetAssignerTest(test_case.TestCase):
groundtruth_boxlist2 = np.array([[0, 0.25123152, 1, 1],
[0.015789, 0.0985, 0.55789, 0.3842]],
dtype=np.float32)
class_targets1 = np.array([[0, 1, 0, 0]], dtype=np.float32)
class_targets2 = np.array([[0, 0, 0, 1],
[0, 0, 1, 0]], dtype=np.float32)
class_targets1 = np.array([[[0, 1, 1],
[1, 1, 0]]], dtype=np.float32)
class_targets2 = np.array([[[0, 1, 1],
......@@ -821,6 +818,63 @@ class BatchTargetAssignerTest(test_case.TestCase):
self.assertAllClose(reg_weights_out, exp_reg_weights)
class BatchGetTargetsTest(test_case.TestCase):
def test_scalar_targets(self):
batch_match = np.array([[1, 0, 1],
[-2, -1, 1]], dtype=np.int32)
groundtruth_tensors_list = np.array([[11, 12], [13, 14]], dtype=np.int32)
groundtruth_weights_list = np.array([[1.0, 1.0], [1.0, 0.5]],
dtype=np.float32)
unmatched_value = np.array(99, dtype=np.int32)
unmatched_weight = np.array(0.0, dtype=np.float32)
def graph_fn(batch_match, groundtruth_tensors_list,
groundtruth_weights_list, unmatched_value, unmatched_weight):
targets, weights = targetassigner.batch_get_targets(
batch_match, tf.unstack(groundtruth_tensors_list),
tf.unstack(groundtruth_weights_list),
unmatched_value, unmatched_weight)
return (targets, weights)
(targets_np, weights_np) = self.execute(graph_fn, [
batch_match, groundtruth_tensors_list, groundtruth_weights_list,
unmatched_value, unmatched_weight
])
self.assertAllEqual([[12, 11, 12],
[99, 99, 14]], targets_np)
self.assertAllClose([[1.0, 1.0, 1.0],
[0.0, 0.0, 0.5]], weights_np)
def test_1d_targets(self):
batch_match = np.array([[1, 0, 1],
[-2, -1, 1]], dtype=np.int32)
groundtruth_tensors_list = np.array([[[11, 12], [12, 13]],
[[13, 14], [14, 15]]],
dtype=np.float32)
groundtruth_weights_list = np.array([[1.0, 1.0], [1.0, 0.5]],
dtype=np.float32)
unmatched_value = np.array([99, 99], dtype=np.float32)
unmatched_weight = np.array(0.0, dtype=np.float32)
def graph_fn(batch_match, groundtruth_tensors_list,
groundtruth_weights_list, unmatched_value, unmatched_weight):
targets, weights = targetassigner.batch_get_targets(
batch_match, tf.unstack(groundtruth_tensors_list),
tf.unstack(groundtruth_weights_list),
unmatched_value, unmatched_weight)
return (targets, weights)
(targets_np, weights_np) = self.execute(graph_fn, [
batch_match, groundtruth_tensors_list, groundtruth_weights_list,
unmatched_value, unmatched_weight
])
self.assertAllClose([[[12, 13], [11, 12], [12, 13]],
[[99, 99], [99, 99], [14, 15]]], targets_np)
self.assertAllClose([[1.0, 1.0, 1.0],
[0.0, 0.0, 0.5]], weights_np)
class BatchTargetAssignConfidencesTest(test_case.TestCase):
def _get_target_assigner(self):
......
research/object_detection/data_decoders/tf_example_decoder_test.py
View file @ 80444539
......@@ -18,7 +18,6 @@ import os
import numpy as np
import tensorflow as tf
from tensorflow.python.framework import test_util
from object_detection.core import standard_fields as fields
from object_detection.data_decoders import tf_example_decoder
from object_detection.protos import input_reader_pb2
......@@ -257,7 +256,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
self.assertAllEqual(expected_boxes,
tensor_dict[fields.InputDataFields.groundtruth_boxes])
@test_util.enable_c_shapes
def testDecodeKeypoint(self):
image_tensor = np.random.randint(256, size=(4, 5, 3)).astype(np.uint8)
encoded_jpeg = self._EncodeImage(image_tensor)
......@@ -346,7 +344,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
self.assertAllClose(tensor_dict[fields.InputDataFields.groundtruth_weights],
np.ones(2, dtype=np.float32))
@test_util.enable_c_shapes
def testDecodeObjectLabel(self):
image_tensor = np.random.randint(256, size=(4, 5, 3)).astype(np.uint8)
encoded_jpeg = self._EncodeImage(image_tensor)
......@@ -669,7 +666,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
self.assertAllEqual([3, 1],
tensor_dict[fields.InputDataFields.groundtruth_classes])
@test_util.enable_c_shapes
def testDecodeObjectArea(self):
image_tensor = np.random.randint(256, size=(4, 5, 3)).astype(np.uint8)
encoded_jpeg = self._EncodeImage(image_tensor)
......@@ -696,7 +692,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
self.assertAllEqual(object_area,
tensor_dict[fields.InputDataFields.groundtruth_area])
@test_util.enable_c_shapes
def testDecodeObjectIsCrowd(self):
image_tensor = np.random.randint(256, size=(4, 5, 3)).astype(np.uint8)
encoded_jpeg = self._EncodeImage(image_tensor)
......@@ -725,7 +720,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
[bool(item) for item in object_is_crowd],
tensor_dict[fields.InputDataFields.groundtruth_is_crowd])
@test_util.enable_c_shapes
def testDecodeObjectDifficult(self):
image_tensor = np.random.randint(256, size=(4, 5, 3)).astype(np.uint8)
encoded_jpeg = self._EncodeImage(image_tensor)
......@@ -754,7 +748,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
[bool(item) for item in object_difficult],
tensor_dict[fields.InputDataFields.groundtruth_difficult])
@test_util.enable_c_shapes
def testDecodeObjectGroupOf(self):
image_tensor = np.random.randint(256, size=(4, 5, 3)).astype(np.uint8)
encoded_jpeg = self._EncodeImage(image_tensor)
......@@ -809,7 +802,6 @@ class TfExampleDecoderTest(tf.test.TestCase):
self.assertAllEqual(object_weights,
tensor_dict[fields.InputDataFields.groundtruth_weights])
@test_util.enable_c_shapes
def testDecodeInstanceSegmentation(self):
num_instances = 4
image_height = 5
......
research/object_detection/dockerfiles/android/README.md
View file @ 80444539
......@@ -14,7 +14,7 @@ A couple words of warning:
the container. When running through the tutorial,
**do not close the container**.
2. To be able to deploy the [Android app](
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/examples/android/app)
https://github.com/tensorflow/examples/tree/master/lite/examples/object_detection/android)
(which you will build at the end of the tutorial),
you will need to kill any instances of `adb` running on the host machine. You
can accomplish this by closing all instances of Android Studio, and then
......
research/object_detection/g3doc/running_on_mobile_tensorflowlite.md
View file @ 80444539
......@@ -94,43 +94,41 @@ bazel run --config=opt tensorflow/lite/toco:toco -- \
# Running our model on Android
To run our TensorFlow Lite model on device, we will need to install the Android
NDK and SDK. The current recommended Android NDK version is 14b and can be found
on the [NDK
Archives](https://developer.android.com/ndk/downloads/older_releases.html#ndk-14b-downloads)
page. Android SDK and build tools can be [downloaded
separately](https://developer.android.com/tools/revisions/build-tools.html) or
used as part of [Android
Studio](https://developer.android.com/studio/index.html). To build the
TensorFlow Lite Android demo, build tools require API >= 23 (but it will run on
devices with API >= 21). Additional details are available on the [TensorFlow
Lite Android App
page](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/java/demo/README.md).
To run our TensorFlow Lite model on device, we will use Android Studio to build
and run the TensorFlow Lite detection example with the new model. The example is
found in the
[TensorFlow examples repository](https://github.com/tensorflow/examples) under
`/lite/examples/object_detection`. The example can be built with
[Android Studio](https://developer.android.com/studio/index.html), and requires
the
[Android SDK with build tools](https://developer.android.com/tools/revisions/build-tools.html)
that support API >= 21. Additional details are available on the
[TensorFlow Lite example page](https://github.com/tensorflow/examples/tree/master/lite/examples/object_detection/android).
Next we need to point the app to our new detect.tflite file and give it the
names of our new labels. Specifically, we will copy our TensorFlow Lite
flatbuffer to the app assets directory with the following command:
```shell
mkdir $TF_EXAMPLES/lite/examples/object_detection/android/app/src/main/assets
cp /tmp/tflite/detect.tflite \
//tensorflow/lite/examples/android/app/src/main/assets
$TF_EXAMPLES/lite/examples/object_detection/android/app/src/main/assets
```
You will also need to copy your new labelmap labels_list.txt to the assets
You will also need to copy your new labelmap labelmap.txt to the assets
directory.
We will now edit the BUILD file to point to this new model. First, open the
BUILD file tensorflow/lite/examples/android/BUILD. Then find the assets
section, and replace the line “@tflite_mobilenet_ssd_quant//:detect.tflite”
(which by default points to a COCO pretrained model) with the path to your new
TFLite model
“//tensorflow/lite/examples/android/app/src/main/assets:detect.tflite”.
Finally, change the last line in assets section to use the new label map as
well.
We will also need to tell our app to use the new label map. In order to do this,
open up the
tensorflow/lite/examples/android/app/src/main/java/org/tensorflow/demo/DetectorActivity.java
We will now edit the gradle build file to use these assets. First, open the
`build.gradle` file
`$TF_EXAMPLES/lite/examples/object_detection/android/app/build.gradle`. Comment
out the model download script to avoid your assets being overwritten: `// apply
from:'download_model.gradle'` ```
If your model is named `detect.tflite`, and your labels file `labelmap.txt`, the
example will use them automatically as long as they've been properly copied into
the base assets directory. If you need to use a custom path or filename, open up
the
$TF_EXAMPLES/lite/examples/object_detection/android/app/src/main/java/org/tensorflow/demo/DetectorActivity.java
file in a text editor and find the definition of TF_OD_API_LABELS_FILE. Update
this path to point to your new label map file:
"file:///android_asset/labels_list.txt". Note that if your model is quantized,
......@@ -144,20 +142,6 @@ DetectorActivity.java should now look as follows for a quantized model:
private static final String TF_OD_API_LABELS_FILE = "file:///android_asset/labels_list.txt";
```
Once you’ve copied the TensorFlow Lite file and edited your BUILD and
DetectorActivity.java files, you can build the demo app, run this bazel command
from the tensorflow directory:
```shell
bazel build -c opt --config=android_arm{,64} --cxxopt='--std=c++11'
"//tensorflow/lite/examples/android:tflite_demo"
```
Now install the demo on a
[debug-enabled](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/android#install)
Android phone via [Android Debug
Bridge](https://developer.android.com/studio/command-line/adb) (adb):
```shell
adb install bazel-bin/tensorflow/lite/examples/android/tflite_demo.apk
```
Once you’ve copied the TensorFlow Lite model and edited the gradle build script
to not use the downloaded assets, you can build and deploy the app using the
usual Android Studio build process.
research/object_detection/g3doc/tpu_exporters.md 0 → 100644
View file @ 80444539
# Object Detection TPU Inference Exporter
This package contains SavedModel Exporter for TPU Inference of object detection
models.
## Usage
This Exporter is intended for users who have trained models with CPUs / GPUs,
but would like to use them for inference on TPU without changing their code or
re-training their models.
Users are assumed to have:
+ `PIPELINE_CONFIG`: A pipeline_pb2.TrainEvalPipelineConfig config file;
+ `CHECKPOINT`: A model checkpoint trained on any device;
and need to correctly set:
+ `EXPORT_DIR`: Path to export SavedModel;
+ `INPUT_PLACEHOLDER`: Name of input placeholder in model's signature_def_map;
+ `INPUT_TYPE`: Type of input node, which can be one of 'image_tensor',
'encoded_image_string_tensor', or 'tf_example';
+ `USE_BFLOAT16`: Whether to use bfloat16 instead of float32 on TPU.
The model can be exported with:
```
python object_detection/tpu_exporters/export_saved_model_tpu.py \
--pipeline_config_file=<PIPELINE_CONFIG> \
--ckpt_path=<CHECKPOINT> \
--export_dir=<EXPORT_DIR> \
--input_placeholder_name=<INPUT_PLACEHOLDER> \
--input_type=<INPUT_TYPE> \
--use_bfloat16=<USE_BFLOAT16>
```
research/object_detection/inputs.py
View file @ 80444539
......@@ -43,6 +43,7 @@ SERVING_FED_EXAMPLE_KEY = 'serialized_example'
# A map of names to methods that help build the input pipeline.
INPUT_BUILDER_UTIL_MAP = {
'dataset_build': dataset_builder.build,
'model_build': model_builder.build,
}
......@@ -152,7 +153,7 @@ def transform_input_data(tensor_dict,
if use_multiclass_scores:
tensor_dict[fields.InputDataFields.groundtruth_classes] = tensor_dict[
fields.InputDataFields.multiclass_scores]
tensor_dict.pop(fields.InputDataFields.multiclass_scores, None)
tensor_dict.pop(fields.InputDataFields.multiclass_scores, None)
if fields.InputDataFields.groundtruth_confidences in tensor_dict:
groundtruth_confidences = tensor_dict[
......@@ -498,11 +499,13 @@ def create_train_input_fn(train_config, train_input_config,
data_augmentation_fn = functools.partial(
augment_input_data,
data_augmentation_options=data_augmentation_options)
model = model_builder.build(model_config, is_training=True)
model_preprocess_fn = INPUT_BUILDER_UTIL_MAP['model_build'](
model_config, is_training=True).preprocess
image_resizer_config = config_util.get_image_resizer_config(model_config)
image_resizer_fn = image_resizer_builder.build(image_resizer_config)
transform_data_fn = functools.partial(
transform_input_data, model_preprocess_fn=model.preprocess,
transform_input_data, model_preprocess_fn=model_preprocess_fn,
image_resizer_fn=image_resizer_fn,
num_classes=config_util.get_number_of_classes(model_config),
data_augmentation_fn=data_augmentation_fn,
......@@ -593,12 +596,14 @@ def create_eval_input_fn(eval_config, eval_input_config, model_config):
def transform_and_pad_input_data_fn(tensor_dict):
"""Combines transform and pad operation."""
num_classes = config_util.get_number_of_classes(model_config)
model = model_builder.build(model_config, is_training=False)
model_preprocess_fn = INPUT_BUILDER_UTIL_MAP['model_build'](
model_config, is_training=False).preprocess
image_resizer_config = config_util.get_image_resizer_config(model_config)
image_resizer_fn = image_resizer_builder.build(image_resizer_config)
transform_data_fn = functools.partial(
transform_input_data, model_preprocess_fn=model.preprocess,
transform_input_data, model_preprocess_fn=model_preprocess_fn,
image_resizer_fn=image_resizer_fn,
num_classes=num_classes,
data_augmentation_fn=None,
......@@ -643,12 +648,14 @@ def create_predict_input_fn(model_config, predict_input_config):
example = tf.placeholder(dtype=tf.string, shape=[], name='tf_example')
num_classes = config_util.get_number_of_classes(model_config)
model = model_builder.build(model_config, is_training=False)
model_preprocess_fn = INPUT_BUILDER_UTIL_MAP['model_build'](
model_config, is_training=False).preprocess
image_resizer_config = config_util.get_image_resizer_config(model_config)
image_resizer_fn = image_resizer_builder.build(image_resizer_config)
transform_fn = functools.partial(
transform_input_data, model_preprocess_fn=model.preprocess,
transform_input_data, model_preprocess_fn=model_preprocess_fn,
image_resizer_fn=image_resizer_fn,
num_classes=num_classes,
data_augmentation_fn=None)
......
research/object_detection/meta_architectures/faster_rcnn_meta_arch.py
View file @ 80444539
......@@ -98,6 +98,7 @@ import tensorflow as tf
from object_detection.anchor_generators import grid_anchor_generator
from object_detection.builders import box_predictor_builder
from object_detection.builders import hyperparams_builder
from object_detection.core import box_list
from object_detection.core import box_list_ops
from object_detection.core import box_predictor
......@@ -110,6 +111,8 @@ from object_detection.utils import shape_utils
slim = tf.contrib.slim
_UNINITIALIZED_FEATURE_EXTRACTOR = '__uninitialized__'
class FasterRCNNFeatureExtractor(object):
"""Faster R-CNN Feature Extractor definition."""
......@@ -216,6 +219,71 @@ class FasterRCNNFeatureExtractor(object):
return variables_to_restore
class FasterRCNNKerasFeatureExtractor(object):
"""Keras-based Faster R-CNN Feature Extractor definition."""
def __init__(self,
is_training,
first_stage_features_stride,
batch_norm_trainable=False,
weight_decay=0.0):
"""Constructor.
Args:
is_training: A boolean indicating whether the training version of the
computation graph should be constructed.
first_stage_features_stride: Output stride of extracted RPN feature map.
batch_norm_trainable: Whether to update batch norm parameters during
training or not. When training with a relative large batch size
(e.g. 8), it could be desirable to enable batch norm update.
weight_decay: float weight decay for feature extractor (default: 0.0).
"""
self._is_training = is_training
self._first_stage_features_stride = first_stage_features_stride
self._train_batch_norm = (batch_norm_trainable and is_training)
self._weight_decay = weight_decay
@abc.abstractmethod
def preprocess(self, resized_inputs):
"""Feature-extractor specific preprocessing (minus image resizing)."""
pass
@abc.abstractmethod
def get_proposal_feature_extractor_model(self, name):
"""Get model that extracts first stage RPN features, to be overridden."""
pass
@abc.abstractmethod
def get_box_classifier_feature_extractor_model(self, name):
"""Get model that extracts second stage box classifier features."""
pass
def restore_from_classification_checkpoint_fn(
self,
first_stage_feature_extractor_scope,
second_stage_feature_extractor_scope):
"""Returns a map of variables to load from a foreign checkpoint.
Args:
first_stage_feature_extractor_scope: A scope name for the first stage
feature extractor.
second_stage_feature_extractor_scope: A scope name for the second stage
feature extractor.
Returns:
A dict mapping variable names (to load from a checkpoint) to variables in
the model graph.
"""
variables_to_restore = {}
for variable in tf.global_variables():
for scope_name in [first_stage_feature_extractor_scope,
second_stage_feature_extractor_scope]:
if variable.op.name.startswith(scope_name):
var_name = variable.op.name.replace(scope_name + '/', '')
variables_to_restore[var_name] = variable
return variables_to_restore
class FasterRCNNMetaArch(model.DetectionModel):
"""Faster R-CNN Meta-architecture definition."""
......@@ -256,7 +324,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
add_summaries=True,
clip_anchors_to_image=False,
use_static_shapes=False,
resize_masks=True):
resize_masks=True,
freeze_batchnorm=False):
"""FasterRCNNMetaArch Constructor.
Args:
......@@ -296,9 +365,12 @@ class FasterRCNNMetaArch(model.DetectionModel):
denser resolutions. The atrous rate is used to compensate for the
denser feature maps by using an effectively larger receptive field.
(This should typically be set to 1).
first_stage_box_predictor_arg_scope_fn: A function to construct tf-slim
arg_scope for conv2d, separable_conv2d and fully_connected ops for the
RPN box predictor.
first_stage_box_predictor_arg_scope_fn: Either a
Keras layer hyperparams object or a function to construct tf-slim
arg_scope for conv2d, separable_conv2d and fully_connected ops. Used
for the RPN box predictor. If it is a keras hyperparams object the
RPN box predictor will be a Keras model. If it is a function to
construct an arg scope it will be a tf-slim box predictor.
first_stage_box_predictor_kernel_size: Kernel size to use for the
convolution op just prior to RPN box predictions.
first_stage_box_predictor_depth: Output depth for the convolution op
......@@ -378,6 +450,10 @@ class FasterRCNNMetaArch(model.DetectionModel):
guarantees.
resize_masks: Indicates whether the masks presend in the groundtruth
should be resized in the model with `image_resizer_fn`
freeze_batchnorm: Whether to freeze batch norm parameters in the first
stage box predictor during training or not. When training with a small
batch size (e.g. 1), it is desirable to freeze batch norm update and
use pretrained batch norm params.
Raises:
ValueError: If `second_stage_batch_size` > `first_stage_max_proposals` at
......@@ -398,6 +474,19 @@ class FasterRCNNMetaArch(model.DetectionModel):
self._image_resizer_fn = image_resizer_fn
self._resize_masks = resize_masks
self._feature_extractor = feature_extractor
if isinstance(feature_extractor, FasterRCNNKerasFeatureExtractor):
# We delay building the feature extractor until it is used,
# to avoid creating the variables when a model is built just for data
# preprocessing. (This prevents a subtle bug where variable names are
# mismatched across workers, causing only one worker to be able to train)
self._feature_extractor_for_proposal_features = (
_UNINITIALIZED_FEATURE_EXTRACTOR)
self._feature_extractor_for_box_classifier_features = (
_UNINITIALIZED_FEATURE_EXTRACTOR)
else:
self._feature_extractor_for_proposal_features = None
self._feature_extractor_for_box_classifier_features = None
self._number_of_stages = number_of_stages
self._proposal_target_assigner = first_stage_target_assigner
......@@ -408,25 +497,82 @@ class FasterRCNNMetaArch(model.DetectionModel):
# (First stage) Region proposal network parameters
self._first_stage_anchor_generator = first_stage_anchor_generator
self._first_stage_atrous_rate = first_stage_atrous_rate
self._first_stage_box_predictor_arg_scope_fn = (
first_stage_box_predictor_arg_scope_fn)
self._first_stage_box_predictor_depth = first_stage_box_predictor_depth
self._first_stage_box_predictor_kernel_size = (
first_stage_box_predictor_kernel_size)
self._first_stage_box_predictor_depth = first_stage_box_predictor_depth
self._first_stage_minibatch_size = first_stage_minibatch_size
self._first_stage_sampler = first_stage_sampler
self._first_stage_box_predictor = (
box_predictor_builder.build_convolutional_box_predictor(
is_training=self._is_training,
num_classes=1,
conv_hyperparams_fn=self._first_stage_box_predictor_arg_scope_fn,
use_dropout=False,
dropout_keep_prob=1.0,
box_code_size=self._box_coder.code_size,
kernel_size=1,
num_layers_before_predictor=0,
min_depth=0,
max_depth=0))
if isinstance(first_stage_box_predictor_arg_scope_fn,
hyperparams_builder.KerasLayerHyperparams):
num_anchors_per_location = (
self._first_stage_anchor_generator.num_anchors_per_location())
if len(num_anchors_per_location) != 1:
raise ValueError('anchor_generator is expected to generate anchors '
'corresponding to a single feature map.')
conv_hyperparams = (
first_stage_box_predictor_arg_scope_fn)
self._first_stage_box_predictor_first_conv = (
tf.keras.Sequential([
tf.keras.layers.Conv2D(
self._first_stage_box_predictor_depth,
kernel_size=[self._first_stage_box_predictor_kernel_size,
self._first_stage_box_predictor_kernel_size],
dilation_rate=self._first_stage_atrous_rate,
padding='SAME',
name='RPNConv',
**conv_hyperparams.params()),
conv_hyperparams.build_batch_norm(
(self._is_training and not freeze_batchnorm),
name='RPNBatchNorm'),
tf.keras.layers.Lambda(
tf.nn.relu6,
name='RPNActivation')
], name='FirstStageRPNFeatures'))
self._first_stage_box_predictor = (
box_predictor_builder.build_convolutional_keras_box_predictor(
is_training=self._is_training,
num_classes=1,
conv_hyperparams=conv_hyperparams,
freeze_batchnorm=freeze_batchnorm,
inplace_batchnorm_update=False,
num_predictions_per_location_list=num_anchors_per_location,
use_dropout=False,
dropout_keep_prob=1.0,
box_code_size=self._box_coder.code_size,
kernel_size=1,
num_layers_before_predictor=0,
min_depth=0,
max_depth=0,
name=self.first_stage_box_predictor_scope))
else:
self._first_stage_box_predictor_arg_scope_fn = (
first_stage_box_predictor_arg_scope_fn)
def rpn_box_predictor_feature_extractor(rpn_features_to_crop):
with slim.arg_scope(self._first_stage_box_predictor_arg_scope_fn()):
reuse = tf.get_variable_scope().reuse
return slim.conv2d(
rpn_features_to_crop,
self._first_stage_box_predictor_depth,
kernel_size=[self._first_stage_box_predictor_kernel_size,
self._first_stage_box_predictor_kernel_size],
rate=self._first_stage_atrous_rate,
activation_fn=tf.nn.relu6,
scope='Conv',
reuse=reuse)
self._first_stage_box_predictor_first_conv = (
rpn_box_predictor_feature_extractor)
self._first_stage_box_predictor = (
box_predictor_builder.build_convolutional_box_predictor(
is_training=self._is_training,
num_classes=1,
conv_hyperparams_fn=self._first_stage_box_predictor_arg_scope_fn,
use_dropout=False,
dropout_keep_prob=1.0,
box_code_size=self._box_coder.code_size,
kernel_size=1,
num_layers_before_predictor=0,
min_depth=0,
max_depth=0))
self._first_stage_nms_fn = first_stage_non_max_suppression_fn
self._first_stage_max_proposals = first_stage_max_proposals
......@@ -444,6 +590,12 @@ class FasterRCNNMetaArch(model.DetectionModel):
self._initial_crop_size = initial_crop_size
self._maxpool_kernel_size = maxpool_kernel_size
self._maxpool_stride = maxpool_stride
# If max pooling is to be used, build the layer
if maxpool_kernel_size:
self._maxpool_layer = tf.keras.layers.MaxPooling2D(
[self._maxpool_kernel_size, self._maxpool_kernel_size],
strides=self._maxpool_stride,
name='MaxPool2D')
self._mask_rcnn_box_predictor = second_stage_mask_rcnn_box_predictor
......@@ -469,6 +621,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
if self._number_of_stages <= 0 or self._number_of_stages > 3:
raise ValueError('Number of stages should be a value in {1, 2, 3}.')
self._batched_prediction_tensor_names = []
@property
def first_stage_feature_extractor_scope(self):
......@@ -510,6 +663,13 @@ class FasterRCNNMetaArch(model.DetectionModel):
raise RuntimeError('anchors should be a BoxList object, but is not.')
return self._anchors
@property
def batched_prediction_tensor_names(self):
if not self._batched_prediction_tensor_names:
raise RuntimeError('Must call predict() method to get batched prediction '
'tensor names.')
return self._batched_prediction_tensor_names
def preprocess(self, inputs):
"""Feature-extractor specific preprocessing.
......@@ -567,6 +727,17 @@ class FasterRCNNMetaArch(model.DetectionModel):
clip_heights, clip_widths], axis=1))
return clip_window
def _proposal_postprocess(self, rpn_box_encodings,
rpn_objectness_predictions_with_background, anchors,
image_shape, true_image_shapes):
"""Wraps over FasterRCNNMetaArch._postprocess_rpn()."""
image_shape_2d = self._image_batch_shape_2d(image_shape)
proposal_boxes_normalized, _, num_proposals, _, _ = \
self._postprocess_rpn(
rpn_box_encodings, rpn_objectness_predictions_with_background,
anchors, image_shape_2d, true_image_shapes)
return proposal_boxes_normalized, num_proposals
def predict(self, preprocessed_inputs, true_image_shapes):
"""Predicts unpostprocessed tensors from input tensor.
......@@ -643,6 +814,60 @@ class FasterRCNNMetaArch(model.DetectionModel):
Raises:
ValueError: If `predict` is called before `preprocess`.
"""
prediction_dict = self._predict_first_stage(preprocessed_inputs)
if self._number_of_stages >= 2:
# If mixed-precision training on TPU is enabled, rpn_box_encodings and
# rpn_objectness_predictions_with_background are bfloat16 tensors.
# Considered prediction results, they need to be casted to float32
# tensors for correct postprocess_rpn computation in predict_second_stage.
prediction_dict.update(
self._predict_second_stage(
tf.to_float(prediction_dict['rpn_box_encodings']),
tf.to_float(
prediction_dict['rpn_objectness_predictions_with_background']
), prediction_dict['rpn_features_to_crop'],
prediction_dict['anchors'], prediction_dict['image_shape'],
true_image_shapes))
if self._number_of_stages == 3:
prediction_dict = self._predict_third_stage(prediction_dict,
true_image_shapes)
self._batched_prediction_tensor_names = [
x for x in prediction_dict if x not in ('image_shape', 'anchors')
]
return prediction_dict
def _predict_first_stage(self, preprocessed_inputs):
"""First stage of prediction.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
prediction_dict: a dictionary holding "raw" prediction tensors:
1) rpn_box_predictor_features: A 4-D float32 tensor with shape
[batch_size, height, width, depth] to be used for predicting proposal
boxes and corresponding objectness scores.
2) rpn_features_to_crop: A 4-D float32 tensor with shape
[batch_size, height, width, depth] representing image features to crop
using the proposal boxes predicted by the RPN.
3) image_shape: a 1-D tensor of shape [4] representing the input
image shape.
4) rpn_box_encodings: 3-D float tensor of shape
[batch_size, num_anchors, self._box_coder.code_size] containing
predicted boxes.
5) rpn_objectness_predictions_with_background: 3-D float tensor of shape
[batch_size, num_anchors, 2] containing class
predictions (logits) for each of the anchors. Note that this
tensor *includes* background class predictions (at class index 0).
6) anchors: A 2-D tensor of shape [num_anchors, 4] representing anchors
for the first stage RPN (in absolute coordinates). Note that
`num_anchors` can differ depending on whether the model is created in
training or inference mode.
"""
(rpn_box_predictor_features, rpn_features_to_crop, anchors_boxlist,
image_shape) = self._extract_rpn_feature_maps(preprocessed_inputs)
(rpn_box_encodings, rpn_objectness_predictions_with_background
......@@ -667,30 +892,19 @@ class FasterRCNNMetaArch(model.DetectionModel):
self._anchors = anchors_boxlist
prediction_dict = {
'rpn_box_predictor_features': rpn_box_predictor_features,
'rpn_features_to_crop': rpn_features_to_crop,
'image_shape': image_shape,
'rpn_box_encodings': rpn_box_encodings,
'rpn_box_predictor_features':
rpn_box_predictor_features,
'rpn_features_to_crop':
rpn_features_to_crop,
'image_shape':
image_shape,
'rpn_box_encodings':
rpn_box_encodings,
'rpn_objectness_predictions_with_background':
rpn_objectness_predictions_with_background,
'anchors': self._anchors.get()
rpn_objectness_predictions_with_background,
'anchors':
anchors_boxlist.data['boxes'],
}
if self._number_of_stages >= 2:
# If mixed-precision training on TPU is enabled, rpn_box_encodings and
# rpn_objectness_predictions_with_background are bfloat16 tensors.
# Considered prediction results, they need to be casted to float32
# tensors for correct postprocess_rpn computation in predict_second_stage.
prediction_dict.update(self._predict_second_stage(
tf.to_float(rpn_box_encodings),
tf.to_float(rpn_objectness_predictions_with_background),
rpn_features_to_crop,
self._anchors.get(), image_shape, true_image_shapes))
if self._number_of_stages == 3:
prediction_dict = self._predict_third_stage(
prediction_dict, true_image_shapes)
return prediction_dict
def _image_batch_shape_2d(self, image_batch_shape_1d):
......@@ -769,11 +983,55 @@ class FasterRCNNMetaArch(model.DetectionModel):
6) box_classifier_features: a 4-D float32 or bfloat16 tensor
representing the features for each proposal.
"""
image_shape_2d = self._image_batch_shape_2d(image_shape)
proposal_boxes_normalized, _, num_proposals, _, _ = self._postprocess_rpn(
rpn_box_encodings, rpn_objectness_predictions_with_background,
anchors, image_shape_2d, true_image_shapes)
proposal_boxes_normalized, num_proposals = self._proposal_postprocess(
rpn_box_encodings, rpn_objectness_predictions_with_background, anchors,
image_shape, true_image_shapes)
prediction_dict = self._box_prediction(rpn_features_to_crop,
proposal_boxes_normalized,
image_shape)
prediction_dict['num_proposals'] = num_proposals
return prediction_dict
def _box_prediction(self, rpn_features_to_crop, proposal_boxes_normalized,
image_shape):
"""Predicts the output tensors from second stage of Faster R-CNN.
Args:
rpn_features_to_crop: A 4-D float32 or bfloat16 tensor with shape
[batch_size, height, width, depth] representing image features to crop
using the proposal boxes predicted by the RPN.
proposal_boxes_normalized: A float tensor with shape [batch_size,
max_num_proposals, 4] representing the (potentially zero padded)
proposal boxes for all images in the batch. These boxes are represented
as normalized coordinates.
image_shape: A 1D int32 tensors of size [4] containing the image shape.
Returns:
prediction_dict: a dictionary holding "raw" prediction tensors:
1) refined_box_encodings: a 3-D tensor with shape
[total_num_proposals, num_classes, self._box_coder.code_size]
representing predicted (final) refined box encodings, where
total_num_proposals=batch_size*self._max_num_proposals. If using a
shared box across classes the shape will instead be
[total_num_proposals, 1, self._box_coder.code_size].
2) class_predictions_with_background: a 3-D tensor with shape
[total_num_proposals, num_classes + 1] containing class
predictions (logits) for each of the anchors, where
total_num_proposals=batch_size*self._max_num_proposals.
Note that this tensor *includes* background class predictions
(at class index 0).
3) proposal_boxes: A float32 tensor of shape
[batch_size, self.max_num_proposals, 4] representing
decoded proposal bounding boxes in absolute coordinates.
4) proposal_boxes_normalized: A float32 tensor of shape
[batch_size, self.max_num_proposals, 4] representing decoded proposal
bounding boxes in normalized coordinates. Can be used to override the
boxes proposed by the RPN, thus enabling one to extract features and
get box classification and prediction for externally selected areas
of the image.
5) box_classifier_features: a 4-D float32 or bfloat16 tensor
representing the features for each proposal.
"""
# If mixed-precision training on TPU is enabled, the dtype of
# rpn_features_to_crop is bfloat16, otherwise it is float32. tf.cast is
# used to match the dtype of proposal_boxes_normalized to that of
......@@ -783,10 +1041,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
rpn_features_to_crop,
tf.cast(proposal_boxes_normalized, rpn_features_to_crop.dtype)))
box_classifier_features = (
self._feature_extractor.extract_box_classifier_features(
flattened_proposal_feature_maps,
scope=self.second_stage_feature_extractor_scope))
box_classifier_features = self._extract_box_classifier_features(
flattened_proposal_feature_maps)
if self._mask_rcnn_box_predictor.is_keras_model:
box_predictions = self._mask_rcnn_box_predictor(
......@@ -813,7 +1069,6 @@ class FasterRCNNMetaArch(model.DetectionModel):
'refined_box_encodings': refined_box_encodings,
'class_predictions_with_background':
class_predictions_with_background,
'num_proposals': num_proposals,
'proposal_boxes': absolute_proposal_boxes,
'box_classifier_features': box_classifier_features,
'proposal_boxes_normalized': proposal_boxes_normalized,
......@@ -821,6 +1076,24 @@ class FasterRCNNMetaArch(model.DetectionModel):
return prediction_dict
def _extract_box_classifier_features(self, flattened_feature_maps):
if self._feature_extractor_for_box_classifier_features == (
_UNINITIALIZED_FEATURE_EXTRACTOR):
self._feature_extractor_for_box_classifier_features = (
self._feature_extractor.get_box_classifier_feature_extractor_model(
name=self.second_stage_feature_extractor_scope))
if self._feature_extractor_for_box_classifier_features:
box_classifier_features = (
self._feature_extractor_for_box_classifier_features(
flattened_feature_maps))
else:
box_classifier_features = (
self._feature_extractor.extract_box_classifier_features(
flattened_feature_maps,
scope=self.second_stage_feature_extractor_scope))
return box_classifier_features
def _predict_third_stage(self, prediction_dict, image_shapes):
"""Predicts non-box, non-class outputs using refined detections.
......@@ -896,10 +1169,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
flattened_detected_feature_maps = (
self._compute_second_stage_input_feature_maps(
rpn_features_to_crop, detection_boxes))
curr_box_classifier_features = (
self._feature_extractor.extract_box_classifier_features(
flattened_detected_feature_maps,
scope=self.second_stage_feature_extractor_scope))
curr_box_classifier_features = self._extract_box_classifier_features(
flattened_detected_feature_maps)
if self._mask_rcnn_box_predictor.is_keras_model:
mask_predictions = self._mask_rcnn_box_predictor(
......@@ -972,29 +1243,35 @@ class FasterRCNNMetaArch(model.DetectionModel):
"""
image_shape = tf.shape(preprocessed_inputs)
rpn_features_to_crop, self.endpoints = (
self._feature_extractor.extract_proposal_features(
preprocessed_inputs,
scope=self.first_stage_feature_extractor_scope))
rpn_features_to_crop, self.endpoints = self._extract_proposal_features(
preprocessed_inputs)
feature_map_shape = tf.shape(rpn_features_to_crop)
anchors = box_list_ops.concatenate(
self._first_stage_anchor_generator.generate([(feature_map_shape[1],
feature_map_shape[2])]))
with slim.arg_scope(self._first_stage_box_predictor_arg_scope_fn()):
kernel_size = self._first_stage_box_predictor_kernel_size
reuse = tf.get_variable_scope().reuse
rpn_box_predictor_features = slim.conv2d(
rpn_features_to_crop,
self._first_stage_box_predictor_depth,
kernel_size=[kernel_size, kernel_size],
rate=self._first_stage_atrous_rate,
activation_fn=tf.nn.relu6,
scope='Conv',
reuse=reuse)
rpn_box_predictor_features = (
self._first_stage_box_predictor_first_conv(rpn_features_to_crop))
return (rpn_box_predictor_features, rpn_features_to_crop,
anchors, image_shape)
def _extract_proposal_features(self, preprocessed_inputs):
if self._feature_extractor_for_proposal_features == (
_UNINITIALIZED_FEATURE_EXTRACTOR):
self._feature_extractor_for_proposal_features = (
self._feature_extractor.get_proposal_feature_extractor_model(
name=self.first_stage_feature_extractor_scope))
if self._feature_extractor_for_proposal_features:
proposal_features = (
self._feature_extractor_for_proposal_features(preprocessed_inputs),
{})
else:
proposal_features = (
self._feature_extractor.extract_proposal_features(
preprocessed_inputs,
scope=self.first_stage_feature_extractor_scope))
return proposal_features
def _predict_rpn_proposals(self, rpn_box_predictor_features):
"""Adds box predictors to RPN feature map to predict proposals.
......@@ -1141,6 +1418,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
detection_classes: [batch, max_detections]
(this entry is only created if rpn_mode=False)
num_detections: [batch]
raw_detection_boxes: [batch, max_detections, 4]
raw_detection_scores: [batch, max_detections, num_classes + 1]
Raises:
ValueError: If `predict` is called before `preprocess`.
......@@ -1210,10 +1489,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
flattened_detected_feature_maps = (
self._compute_second_stage_input_feature_maps(
rpn_features_to_crop, detection_boxes))
detection_features_unpooled = (
self._feature_extractor.extract_box_classifier_features(
flattened_detected_feature_maps,
scope=self.second_stage_feature_extractor_scope))
detection_features_unpooled = self._extract_box_classifier_features(
flattened_detected_feature_maps)
batch_size = tf.shape(detection_boxes)[0]
max_detection = tf.shape(detection_boxes)[1]
......@@ -1273,9 +1550,9 @@ class FasterRCNNMetaArch(model.DetectionModel):
the batch.
raw_detection_boxes: [batch, total_detections, 4] tensor with decoded
proposal boxes before Non-Max Suppression.
raw_detection_score: [batch, total_detections,
num_classes_with_background] tensor of class score logits for
raw proposal boxes.
raw_detection_scores: [batch, total_detections,
num_classes_with_background] tensor of multi-class scores for raw
proposal boxes.
"""
rpn_box_encodings_batch = tf.expand_dims(rpn_box_encodings_batch, axis=2)
rpn_encodings_shape = shape_utils.combined_static_and_dynamic_shape(
......@@ -1285,8 +1562,9 @@ class FasterRCNNMetaArch(model.DetectionModel):
proposal_boxes = self._batch_decode_boxes(rpn_box_encodings_batch,
tiled_anchor_boxes)
raw_proposal_boxes = tf.squeeze(proposal_boxes, axis=2)
rpn_objectness_softmax_without_background = tf.nn.softmax(
rpn_objectness_predictions_with_background_batch)[:, :, 1]
rpn_objectness_softmax = tf.nn.softmax(
rpn_objectness_predictions_with_background_batch)
rpn_objectness_softmax_without_background = rpn_objectness_softmax[:, :, 1]
clip_window = self._compute_clip_window(image_shapes)
(proposal_boxes, proposal_scores, _, _, _,
num_proposals) = self._first_stage_nms_fn(
......@@ -1319,8 +1597,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
elems=[raw_proposal_boxes, image_shapes],
dtype=tf.float32)
return (normalized_proposal_boxes, proposal_scores, num_proposals,
raw_normalized_proposal_boxes,
rpn_objectness_predictions_with_background_batch)
raw_normalized_proposal_boxes, rpn_objectness_softmax)
def _sample_box_classifier_batch(
self,
......@@ -1547,10 +1824,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
self._crop_and_resize_fn(
features_to_crop, proposal_boxes_normalized,
[self._initial_crop_size, self._initial_crop_size]))
return slim.max_pool2d(
cropped_regions,
[self._maxpool_kernel_size, self._maxpool_kernel_size],
stride=self._maxpool_stride)
return self._maxpool_layer(cropped_regions)
def _postprocess_box_classifier(self,
refined_box_encodings,
......@@ -1567,7 +1841,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
representing predicted (final) refined box encodings. If using a shared
box across classes the shape will instead be
[total_num_padded_proposals, 1, 4]
class_predictions_with_background: a 3-D tensor float with shape
class_predictions_with_background: a 2-D tensor float with shape
[total_num_padded_proposals, num_classes + 1] containing class
predictions (logits) for each of the proposals. Note that this tensor
*includes* background class predictions (at class index 0).
......@@ -1594,8 +1868,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
masks.
`raw_detection_boxes`: [batch, total_detections, 4] tensor with decoded
detection boxes before Non-Max Suppression.
`raw_detection_score`: [batch, total_detections,
num_classes_with_background] tensor of multi-class score logits for
`raw_detection_scores`: [batch, total_detections,
num_classes_with_background] tensor of multi-class scores for
raw detection boxes.
"""
refined_box_encodings_batch = tf.reshape(
......@@ -1679,7 +1953,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
fields.DetectionResultFields.raw_detection_boxes:
raw_normalized_detection_boxes,
fields.DetectionResultFields.raw_detection_scores:
class_predictions_with_background_batch
class_predictions_with_background_batch_normalized
}
if nmsed_masks is not None:
detections[fields.DetectionResultFields.detection_masks] = nmsed_masks
......@@ -2072,7 +2346,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
# Use normalized proposals to crop mask targets from image masks.
flat_normalized_proposals = box_list_ops.to_normalized_coordinates(
box_list.BoxList(tf.reshape(proposal_boxes, [-1, 4])),
image_shape[1], image_shape[2]).get()
image_shape[1], image_shape[2], check_range=False).get()
flat_cropped_gt_mask = self._crop_and_resize_fn(
tf.expand_dims(flat_gt_masks, -1),
......@@ -2261,7 +2535,32 @@ class FasterRCNNMetaArch(model.DetectionModel):
Returns:
A list of regularization loss tensors.
"""
return tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
all_losses = []
slim_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
# Copy the slim losses to avoid modifying the collection
if slim_losses:
all_losses.extend(slim_losses)
# TODO(kaftan): Possibly raise an error if the feature extractors are
# uninitialized in Keras.
if self._feature_extractor_for_proposal_features:
if (self._feature_extractor_for_proposal_features !=
_UNINITIALIZED_FEATURE_EXTRACTOR):
all_losses.extend(self._feature_extractor_for_proposal_features.losses)
if isinstance(self._first_stage_box_predictor_first_conv,
tf.keras.Model):
all_losses.extend(
self._first_stage_box_predictor_first_conv.losses)
if self._first_stage_box_predictor.is_keras_model:
all_losses.extend(self._first_stage_box_predictor.losses)
if self._feature_extractor_for_box_classifier_features:
if (self._feature_extractor_for_box_classifier_features !=
_UNINITIALIZED_FEATURE_EXTRACTOR):
all_losses.extend(
self._feature_extractor_for_box_classifier_features.losses)
if self._mask_rcnn_box_predictor:
if self._mask_rcnn_box_predictor.is_keras_model:
all_losses.extend(self._mask_rcnn_box_predictor.losses)
return all_losses
def restore_map(self,
fine_tune_checkpoint_type='detection',
......@@ -2318,4 +2617,55 @@ class FasterRCNNMetaArch(model.DetectionModel):
Returns:
A list of update operators.
"""
return tf.get_collection(tf.GraphKeys.UPDATE_OPS)
update_ops = []
slim_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# Copy the slim ops to avoid modifying the collection
if slim_update_ops:
update_ops.extend(slim_update_ops)
# Passing None to get_updates_for grabs updates that should always be
# executed and don't depend on any model inputs in the graph.
# (E.g. if there was some count that should be incremented every time a
# model is run).
#
# Passing inputs grabs updates that are transitively computed from the
# model inputs being passed in.
# (E.g. a batchnorm update depends on the observed inputs)
if self._feature_extractor_for_proposal_features:
if (self._feature_extractor_for_proposal_features !=
_UNINITIALIZED_FEATURE_EXTRACTOR):
update_ops.extend(
self._feature_extractor_for_proposal_features.get_updates_for(None))
update_ops.extend(
self._feature_extractor_for_proposal_features.get_updates_for(
self._feature_extractor_for_proposal_features.inputs))
if isinstance(self._first_stage_box_predictor_first_conv,
tf.keras.Model):
update_ops.extend(
self._first_stage_box_predictor_first_conv.get_updates_for(
None))
update_ops.extend(
self._first_stage_box_predictor_first_conv.get_updates_for(
self._first_stage_box_predictor_first_conv.inputs))
if self._first_stage_box_predictor.is_keras_model:
update_ops.extend(
self._first_stage_box_predictor.get_updates_for(None))
update_ops.extend(
self._first_stage_box_predictor.get_updates_for(
self._first_stage_box_predictor.inputs))
if self._feature_extractor_for_box_classifier_features:
if (self._feature_extractor_for_box_classifier_features !=
_UNINITIALIZED_FEATURE_EXTRACTOR):
update_ops.extend(
self._feature_extractor_for_box_classifier_features.get_updates_for(
None))
update_ops.extend(
self._feature_extractor_for_box_classifier_features.get_updates_for(
self._feature_extractor_for_box_classifier_features.inputs))
if self._mask_rcnn_box_predictor:
if self._mask_rcnn_box_predictor.is_keras_model:
update_ops.extend(
self._mask_rcnn_box_predictor.get_updates_for(None))
update_ops.extend(
self._mask_rcnn_box_predictor.get_updates_for(
self._mask_rcnn_box_predictor.inputs))
return update_ops
research/object_detection/meta_architectures/faster_rcnn_meta_arch_test.py
View file @ 80444539
......@@ -26,9 +26,15 @@ class FasterRCNNMetaArchTest(
faster_rcnn_meta_arch_test_lib.FasterRCNNMetaArchTestBase,
parameterized.TestCase):
def test_postprocess_second_stage_only_inference_mode_with_masks(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_postprocess_second_stage_only_inference_mode_with_masks(
self, use_keras=False):
model = self._build_model(
is_training=False, number_of_stages=2, second_stage_batch_size=6)
is_training=False, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
batch_size = 2
total_num_padded_proposals = batch_size * model.max_num_proposals
......@@ -85,9 +91,15 @@ class FasterRCNNMetaArchTest(
self.assertTrue(np.amax(detections_out['detection_masks'] <= 1.0))
self.assertTrue(np.amin(detections_out['detection_masks'] >= 0.0))
def test_postprocess_second_stage_only_inference_mode_with_calibration(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_postprocess_second_stage_only_inference_mode_with_calibration(
self, use_keras=False):
model = self._build_model(
is_training=False, number_of_stages=2, second_stage_batch_size=6,
is_training=False, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6,
calibration_mapping_value=0.5)
batch_size = 2
......@@ -147,9 +159,15 @@ class FasterRCNNMetaArchTest(
self.assertTrue(np.amax(detections_out['detection_masks'] <= 1.0))
self.assertTrue(np.amin(detections_out['detection_masks'] >= 0.0))
def test_postprocess_second_stage_only_inference_mode_with_shared_boxes(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_postprocess_second_stage_only_inference_mode_with_shared_boxes(
self, use_keras=False):
model = self._build_model(
is_training=False, number_of_stages=2, second_stage_batch_size=6)
is_training=False, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
batch_size = 2
total_num_padded_proposals = batch_size * model.max_num_proposals
......@@ -188,11 +206,13 @@ class FasterRCNNMetaArchTest(
self.assertAllClose(detections_out['num_detections'], [5, 4])
@parameterized.parameters(
{'masks_are_class_agnostic': False},
{'masks_are_class_agnostic': True},
{'masks_are_class_agnostic': False, 'use_keras': True},
{'masks_are_class_agnostic': True, 'use_keras': True},
{'masks_are_class_agnostic': False, 'use_keras': False},
{'masks_are_class_agnostic': True, 'use_keras': False},
)
def test_predict_correct_shapes_in_inference_mode_three_stages_with_masks(
self, masks_are_class_agnostic):
self, masks_are_class_agnostic, use_keras):
batch_size = 2
image_size = 10
max_num_proposals = 8
......@@ -232,6 +252,7 @@ class FasterRCNNMetaArchTest(
with test_graph.as_default():
model = self._build_model(
is_training=False,
use_keras=use_keras,
number_of_stages=3,
second_stage_batch_size=2,
predict_masks=True,
......@@ -267,15 +288,18 @@ class FasterRCNNMetaArchTest(
[10, num_classes, 14, 14])
@parameterized.parameters(
{'masks_are_class_agnostic': False},
{'masks_are_class_agnostic': True},
{'masks_are_class_agnostic': False, 'use_keras': True},
{'masks_are_class_agnostic': True, 'use_keras': True},
{'masks_are_class_agnostic': False, 'use_keras': False},
{'masks_are_class_agnostic': True, 'use_keras': False},
)
def test_predict_gives_correct_shapes_in_train_mode_both_stages_with_masks(
self, masks_are_class_agnostic):
self, masks_are_class_agnostic, use_keras):
test_graph = tf.Graph()
with test_graph.as_default():
model = self._build_model(
is_training=True,
use_keras=use_keras,
number_of_stages=3,
second_stage_batch_size=7,
predict_masks=True,
......@@ -348,7 +372,11 @@ class FasterRCNNMetaArchTest(
tensor_dict_out['rpn_objectness_predictions_with_background'].shape,
(2, num_anchors_out, 2))
def test_postprocess_third_stage_only_inference_mode(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_postprocess_third_stage_only_inference_mode(self, use_keras=False):
num_proposals_shapes = [(2), (None)]
refined_box_encodings_shapes = [(16, 2, 4), (None, 2, 4)]
class_predictions_with_background_shapes = [(16, 3), (None, 3)]
......@@ -364,7 +392,7 @@ class FasterRCNNMetaArchTest(
tf_graph = tf.Graph()
with tf_graph.as_default():
model = self._build_model(
is_training=False, number_of_stages=3,
is_training=False, use_keras=use_keras, number_of_stages=3,
second_stage_batch_size=6, predict_masks=True)
total_num_padded_proposals = batch_size * model.max_num_proposals
proposal_boxes = np.array(
......
research/object_detection/meta_architectures/faster_rcnn_meta_arch_test_lib.py
View file @ 80444539
......@@ -43,7 +43,7 @@ BOX_CODE_SIZE = 4
class FakeFasterRCNNFeatureExtractor(
faster_rcnn_meta_arch.FasterRCNNFeatureExtractor):
"""Fake feature extracture to use in tests."""
"""Fake feature extractor to use in tests."""
def __init__(self):
super(FakeFasterRCNNFeatureExtractor, self).__init__(
......@@ -67,6 +67,42 @@ class FakeFasterRCNNFeatureExtractor(
num_outputs=3, kernel_size=1, scope='layer2')
class FakeFasterRCNNKerasFeatureExtractor(
faster_rcnn_meta_arch.FasterRCNNKerasFeatureExtractor):
"""Fake feature extractor to use in tests."""
def __init__(self):
super(FakeFasterRCNNKerasFeatureExtractor, self).__init__(
is_training=False,
first_stage_features_stride=32,
weight_decay=0.0)
def preprocess(self, resized_inputs):
return tf.identity(resized_inputs)
def get_proposal_feature_extractor_model(self, name):
class ProposalFeatureExtractor(tf.keras.Model):
"""Dummy proposal feature extraction."""
def __init__(self, name):
super(ProposalFeatureExtractor, self).__init__(name=name)
self.conv = None
def build(self, input_shape):
self.conv = tf.keras.layers.Conv2D(
3, kernel_size=1, padding='SAME', name='layer1')
def call(self, inputs):
return self.conv(inputs)
return ProposalFeatureExtractor(name=name)
def get_box_classifier_feature_extractor_model(self, name):
return tf.keras.Sequential([tf.keras.layers.Conv2D(
3, kernel_size=1, padding='SAME', name=name + '_layer2')])
class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
"""Base class to test Faster R-CNN and R-FCN meta architectures."""
......@@ -77,6 +113,11 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
text_format.Merge(hyperparams_text_proto, hyperparams)
return hyperparams_builder.build(hyperparams, is_training=is_training)
def _build_keras_layer_hyperparams(self, hyperparams_text_proto):
hyperparams = hyperparams_pb2.Hyperparams()
text_format.Merge(hyperparams_text_proto, hyperparams)
return hyperparams_builder.KerasLayerHyperparams(hyperparams)
def _get_second_stage_box_predictor_text_proto(self):
box_predictor_text_proto = """
mask_rcnn_box_predictor {
......@@ -127,7 +168,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
return box_predictor_text_proto
def _get_second_stage_box_predictor(self, num_classes, is_training,
predict_masks, masks_are_class_agnostic):
predict_masks, masks_are_class_agnostic,
use_keras=False):
box_predictor_proto = box_predictor_pb2.BoxPredictor()
text_format.Merge(self._get_second_stage_box_predictor_text_proto(),
box_predictor_proto)
......@@ -137,13 +179,23 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
masks_are_class_agnostic),
box_predictor_proto)
return box_predictor_builder.build(
hyperparams_builder.build,
box_predictor_proto,
num_classes=num_classes,
is_training=is_training)
if use_keras:
return box_predictor_builder.build_keras(
hyperparams_builder.KerasLayerHyperparams,
inplace_batchnorm_update=False,
freeze_batchnorm=False,
box_predictor_config=box_predictor_proto,
num_classes=num_classes,
num_predictions_per_location_list=None,
is_training=is_training)
else:
return box_predictor_builder.build(
hyperparams_builder.build,
box_predictor_proto,
num_classes=num_classes,
is_training=is_training)
def _get_model(self, box_predictor, **common_kwargs):
def _get_model(self, box_predictor, keras_model=False, **common_kwargs):
return faster_rcnn_meta_arch.FasterRCNNMetaArch(
initial_crop_size=3,
maxpool_kernel_size=1,
......@@ -155,6 +207,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
is_training,
number_of_stages,
second_stage_batch_size,
use_keras=False,
first_stage_max_proposals=8,
num_classes=2,
hard_mining=False,
......@@ -204,7 +257,10 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
'proposal',
use_matmul_gather=use_matmul_gather_in_matcher)
fake_feature_extractor = FakeFasterRCNNFeatureExtractor()
if use_keras:
fake_feature_extractor = FakeFasterRCNNKerasFeatureExtractor()
else:
fake_feature_extractor = FakeFasterRCNNFeatureExtractor()
first_stage_box_predictor_hyperparams_text_proto = """
op: CONV
......@@ -220,9 +276,14 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
}
}
"""
first_stage_box_predictor_arg_scope_fn = (
self._build_arg_scope_with_hyperparams(
first_stage_box_predictor_hyperparams_text_proto, is_training))
if use_keras:
first_stage_box_predictor_arg_scope_fn = (
self._build_keras_layer_hyperparams(
first_stage_box_predictor_hyperparams_text_proto))
else:
first_stage_box_predictor_arg_scope_fn = (
self._build_arg_scope_with_hyperparams(
first_stage_box_predictor_hyperparams_text_proto, is_training))
first_stage_box_predictor_kernel_size = 3
first_stage_atrous_rate = 1
......@@ -349,15 +410,18 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self._get_second_stage_box_predictor(
num_classes=num_classes,
is_training=is_training,
use_keras=use_keras,
predict_masks=predict_masks,
masks_are_class_agnostic=masks_are_class_agnostic), **common_kwargs)
@parameterized.parameters(
{'use_static_shapes': False},
{'use_static_shapes': True}
{'use_static_shapes': False, 'use_keras': True},
{'use_static_shapes': False, 'use_keras': False},
{'use_static_shapes': True, 'use_keras': True},
{'use_static_shapes': True, 'use_keras': False},
)
def test_predict_gives_correct_shapes_in_inference_mode_first_stage_only(
self, use_static_shapes=False):
self, use_static_shapes=False, use_keras=False):
batch_size = 2
height = 10
width = 12
......@@ -367,6 +431,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
"""Function to construct tf graph for the test."""
model = self._build_model(
is_training=False,
use_keras=use_keras,
number_of_stages=1,
second_stage_batch_size=2,
clip_anchors_to_image=use_static_shapes,
......@@ -423,11 +488,44 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertTrue(np.all(np.less_equal(anchors[:, 2], height)))
self.assertTrue(np.all(np.less_equal(anchors[:, 3], width)))
def test_predict_gives_valid_anchors_in_training_mode_first_stage_only(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_regularization_losses(
self, use_keras=False):
test_graph = tf.Graph()
with test_graph.as_default():
model = self._build_model(
is_training=True, number_of_stages=1, second_stage_batch_size=2)
is_training=True, use_keras=use_keras,
number_of_stages=1, second_stage_batch_size=2,)
batch_size = 2
height = 10
width = 12
input_image_shape = (batch_size, height, width, 3)
_, true_image_shapes = model.preprocess(tf.zeros(input_image_shape))
preprocessed_inputs = tf.placeholder(
dtype=tf.float32, shape=(batch_size, None, None, 3))
model.predict(preprocessed_inputs, true_image_shapes)
reg_losses = tf.math.add_n(model.regularization_losses())
init_op = tf.global_variables_initializer()
with self.test_session(graph=test_graph) as sess:
sess.run(init_op)
self.assertGreaterEqual(sess.run(reg_losses), 0)
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_predict_gives_valid_anchors_in_training_mode_first_stage_only(
self, use_keras=False):
test_graph = tf.Graph()
with test_graph.as_default():
model = self._build_model(
is_training=True, use_keras=use_keras,
number_of_stages=1, second_stage_batch_size=2,)
batch_size = 2
height = 10
width = 12
......@@ -478,11 +576,13 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
(batch_size, num_anchors_out, 2))
@parameterized.parameters(
{'use_static_shapes': False},
{'use_static_shapes': True}
{'use_static_shapes': False, 'use_keras': True},
{'use_static_shapes': False, 'use_keras': False},
{'use_static_shapes': True, 'use_keras': True},
{'use_static_shapes': True, 'use_keras': False},
)
def test_predict_correct_shapes_in_inference_mode_two_stages(
self, use_static_shapes=False):
self, use_static_shapes=False, use_keras=False):
def compare_results(results, expected_output_shapes):
"""Checks if the shape of the predictions are as expected."""
......@@ -528,6 +628,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
"""Function to construct tf graph for the test."""
model = self._build_model(
is_training=False,
use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=2,
predict_masks=False,
......@@ -584,6 +685,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
with test_graph.as_default():
model = self._build_model(
is_training=False,
use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=2,
predict_masks=False)
......@@ -603,12 +705,15 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllEqual(tensor_dict_out[key].shape, expected_shapes[key])
@parameterized.parameters(
{'use_static_shapes': False},
{'use_static_shapes': True}
{'use_static_shapes': False, 'use_keras': True},
{'use_static_shapes': False, 'use_keras': False},
{'use_static_shapes': True, 'use_keras': True},
{'use_static_shapes': True, 'use_keras': False},
)
def test_predict_gives_correct_shapes_in_train_mode_both_stages(
self,
use_static_shapes=False):
use_static_shapes=False,
use_keras=False):
batch_size = 2
image_size = 10
max_num_proposals = 7
......@@ -619,6 +724,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
"""Function to construct tf graph for the test."""
model = self._build_model(
is_training=True,
use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=7,
predict_masks=False,
......@@ -632,6 +738,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
groundtruth_classes_list=tf.unstack(gt_classes),
groundtruth_weights_list=tf.unstack(gt_weights))
result_tensor_dict = model.predict(preprocessed_inputs, true_image_shapes)
updates = model.updates()
return (result_tensor_dict['refined_box_encodings'],
result_tensor_dict['class_predictions_with_background'],
result_tensor_dict['proposal_boxes'],
......@@ -641,6 +748,7 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
result_tensor_dict['rpn_objectness_predictions_with_background'],
result_tensor_dict['rpn_features_to_crop'],
result_tensor_dict['rpn_box_predictor_features'],
updates
)
image_shape = (batch_size, image_size, image_size, 3)
......@@ -699,13 +807,27 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
expected_shapes['rpn_box_predictor_features'])
@parameterized.parameters(
{'use_static_shapes': False, 'pad_to_max_dimension': None},
{'use_static_shapes': True, 'pad_to_max_dimension': None},
{'use_static_shapes': False, 'pad_to_max_dimension': 56},
{'use_static_shapes': True, 'pad_to_max_dimension': 56}
{'use_static_shapes': False, 'pad_to_max_dimension': None,
'use_keras': True},
{'use_static_shapes': True, 'pad_to_max_dimension': None,
'use_keras': True},
{'use_static_shapes': False, 'pad_to_max_dimension': 56,
'use_keras': True},
{'use_static_shapes': True, 'pad_to_max_dimension': 56,
'use_keras': True},
{'use_static_shapes': False, 'pad_to_max_dimension': None,
'use_keras': False},
{'use_static_shapes': True, 'pad_to_max_dimension': None,
'use_keras': False},
{'use_static_shapes': False, 'pad_to_max_dimension': 56,
'use_keras': False},
{'use_static_shapes': True, 'pad_to_max_dimension': 56,
'use_keras': False}
)
def test_postprocess_first_stage_only_inference_mode(
self, use_static_shapes=False, pad_to_max_dimension=None):
self, use_static_shapes=False,
pad_to_max_dimension=None,
use_keras=False):
batch_size = 2
first_stage_max_proposals = 4 if use_static_shapes else 8
......@@ -716,7 +838,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
anchors):
"""Function to construct tf graph for the test."""
model = self._build_model(
is_training=False, number_of_stages=1, second_stage_batch_size=6,
is_training=False, use_keras=use_keras,
number_of_stages=1, second_stage_batch_size=6,
use_matmul_crop_and_resize=use_static_shapes,
clip_anchors_to_image=use_static_shapes,
use_static_shapes=use_static_shapes,
......@@ -779,8 +902,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
[0.5, 0., 1., 0.5], [0.5, 0.5, 1., 1.]],
[[0., 0., 0.5, 0.5], [0., 0.5, 0.5, 1.],
[0.5, 0., 1., 0.5], [0.5, 0.5, 1., 1.]]]
expected_raw_scores = [[[-10., 13.], [10., -10.], [10., -11.], [-10., 12.]],
[[10., -10.], [-10., 13.], [-10., 12.], [10., -11.]]]
expected_raw_scores = [[[0., 1.], [1., 0.], [1., 0.], [0., 1.]],
[[1., 0.], [0., 1.], [0., 1.], [1., 0.]]]
self.assertAllClose(results[0], expected_num_proposals)
for indx, num_proposals in enumerate(expected_num_proposals):
......@@ -792,9 +915,11 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllClose(results[4], expected_raw_scores)
def _test_postprocess_first_stage_only_train_mode(self,
use_keras=False,
pad_to_max_dimension=None):
model = self._build_model(
is_training=True, number_of_stages=1, second_stage_batch_size=2,
is_training=True, use_keras=use_keras,
number_of_stages=1, second_stage_batch_size=2,
pad_to_max_dimension=pad_to_max_dimension)
batch_size = 2
anchors = tf.constant(
......@@ -847,9 +972,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
[0.5, 0., 1., 0.5], [0.5, 0.5, 1., 1.]],
[[0., 0., 0.5, 0.5], [0., 0.5, 0.5, 1.],
[0.5, 0., 1., 0.5], [0.5, 0.5, 1., 1.]]]
expected_raw_scores = [[[-10., 13.], [-10., 12.], [-10., 11.], [-10., 10.]],
[[-10., 13.], [-10., 12.], [-10., 11.], [-10., 10.]]]
expected_raw_scores = [[[0., 1.], [0., 1.], [0., 1.], [0., 1.]],
[[0., 1.], [0., 1.], [0., 1.], [0., 1.]]]
expected_output_keys = set([
'detection_boxes', 'detection_scores', 'num_detections',
'raw_detection_boxes', 'raw_detection_scores'
......@@ -872,20 +996,44 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllClose(proposals_out['raw_detection_scores'],
expected_raw_scores)
def test_postprocess_first_stage_only_train_mode(self):
self._test_postprocess_first_stage_only_train_mode()
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_postprocess_first_stage_only_train_mode(self, use_keras=False):
self._test_postprocess_first_stage_only_train_mode(use_keras=use_keras)
def test_postprocess_first_stage_only_train_mode_padded_image(self):
self._test_postprocess_first_stage_only_train_mode(pad_to_max_dimension=56)
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_postprocess_first_stage_only_train_mode_padded_image(
self, use_keras=False):
self._test_postprocess_first_stage_only_train_mode(pad_to_max_dimension=56,
use_keras=use_keras)
@parameterized.parameters(
{'use_static_shapes': False, 'pad_to_max_dimension': None},
{'use_static_shapes': True, 'pad_to_max_dimension': None},
{'use_static_shapes': False, 'pad_to_max_dimension': 56},
{'use_static_shapes': True, 'pad_to_max_dimension': 56}
{'use_static_shapes': False, 'pad_to_max_dimension': None,
'use_keras': True},
{'use_static_shapes': True, 'pad_to_max_dimension': None,
'use_keras': True},
{'use_static_shapes': False, 'pad_to_max_dimension': 56,
'use_keras': True},
{'use_static_shapes': True, 'pad_to_max_dimension': 56,
'use_keras': True},
{'use_static_shapes': False, 'pad_to_max_dimension': None,
'use_keras': False},
{'use_static_shapes': True, 'pad_to_max_dimension': None,
'use_keras': False},
{'use_static_shapes': False, 'pad_to_max_dimension': 56,
'use_keras': False},
{'use_static_shapes': True, 'pad_to_max_dimension': 56,
'use_keras': False}
)
def test_postprocess_second_stage_only_inference_mode(
self, use_static_shapes=False, pad_to_max_dimension=None):
self, use_static_shapes=False,
pad_to_max_dimension=None,
use_keras=False):
batch_size = 2
num_classes = 2
image_shape = np.array((2, 36, 48, 3), dtype=np.int32)
......@@ -899,7 +1047,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
proposal_boxes):
"""Function to construct tf graph for the test."""
model = self._build_model(
is_training=False, number_of_stages=2,
is_training=False, use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=6,
use_matmul_crop_and_resize=use_static_shapes,
clip_anchors_to_image=use_static_shapes,
......@@ -972,21 +1121,31 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
if not use_static_shapes:
self.assertAllEqual(results[1].shape, [2, 5, 4])
def test_preprocess_preserves_input_shapes(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_preprocess_preserves_input_shapes(self, use_keras=False):
image_shapes = [(3, None, None, 3),
(None, 10, 10, 3),
(None, None, None, 3)]
for image_shape in image_shapes:
model = self._build_model(
is_training=False, number_of_stages=2, second_stage_batch_size=6)
is_training=False, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
image_placeholder = tf.placeholder(tf.float32, shape=image_shape)
preprocessed_inputs, _ = model.preprocess(image_placeholder)
self.assertAllEqual(preprocessed_inputs.shape.as_list(), image_shape)
# TODO(rathodv): Split test into two - with and without masks.
def test_loss_first_stage_only_mode(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_loss_first_stage_only_mode(self, use_keras=False):
model = self._build_model(
is_training=True, number_of_stages=1, second_stage_batch_size=6)
is_training=True, use_keras=use_keras,
number_of_stages=1, second_stage_batch_size=6)
batch_size = 2
anchors = tf.constant(
[[0, 0, 16, 16],
......@@ -1038,9 +1197,14 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
loss_dict_out)
# TODO(rathodv): Split test into two - with and without masks.
def test_loss_full(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_loss_full(self, use_keras=False):
model = self._build_model(
is_training=True, number_of_stages=2, second_stage_batch_size=6)
is_training=True, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
batch_size = 3
anchors = tf.constant(
[[0, 0, 16, 16],
......@@ -1153,9 +1317,14 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
'Loss/BoxClassifierLoss/classification_loss'], 0)
self.assertAllClose(loss_dict_out['Loss/BoxClassifierLoss/mask_loss'], 0)
def test_loss_full_zero_padded_proposals(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_loss_full_zero_padded_proposals(self, use_keras=False):
model = self._build_model(
is_training=True, number_of_stages=2, second_stage_batch_size=6)
is_training=True, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
batch_size = 1
anchors = tf.constant(
[[0, 0, 16, 16],
......@@ -1243,9 +1412,14 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
'Loss/BoxClassifierLoss/classification_loss'], 0)
self.assertAllClose(loss_dict_out['Loss/BoxClassifierLoss/mask_loss'], 0)
def test_loss_full_multiple_label_groundtruth(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_loss_full_multiple_label_groundtruth(self, use_keras=False):
model = self._build_model(
is_training=True, number_of_stages=2, second_stage_batch_size=6,
is_training=True, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6,
softmax_second_stage_classification_loss=False)
batch_size = 1
anchors = tf.constant(
......@@ -1342,14 +1516,17 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllClose(loss_dict_out['Loss/BoxClassifierLoss/mask_loss'], 0)
@parameterized.parameters(
{'use_static_shapes': False, 'shared_boxes': False},
{'use_static_shapes': False, 'shared_boxes': True},
{'use_static_shapes': True, 'shared_boxes': False},
{'use_static_shapes': True, 'shared_boxes': True},
{'use_static_shapes': False, 'shared_boxes': False, 'use_keras': True},
{'use_static_shapes': False, 'shared_boxes': True, 'use_keras': True},
{'use_static_shapes': True, 'shared_boxes': False, 'use_keras': True},
{'use_static_shapes': True, 'shared_boxes': True, 'use_keras': True},
{'use_static_shapes': False, 'shared_boxes': False, 'use_keras': False},
{'use_static_shapes': False, 'shared_boxes': True, 'use_keras': False},
{'use_static_shapes': True, 'shared_boxes': False, 'use_keras': False},
{'use_static_shapes': True, 'shared_boxes': True, 'use_keras': False},
)
def test_loss_full_zero_padded_proposals_nonzero_loss_with_two_images(
self, use_static_shapes=False, shared_boxes=False):
self, use_static_shapes=False, shared_boxes=False, use_keras=False):
batch_size = 2
first_stage_max_proposals = 8
second_stage_batch_size = 6
......@@ -1361,7 +1538,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
groundtruth_classes):
"""Function to construct tf graph for the test."""
model = self._build_model(
is_training=True, number_of_stages=2,
is_training=True, use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=second_stage_batch_size,
first_stage_max_proposals=first_stage_max_proposals,
num_classes=num_classes,
......@@ -1476,8 +1654,13 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllClose(results[2], exp_loc_loss, rtol=1e-4, atol=1e-4)
self.assertAllClose(results[3], 0.0)
def test_loss_with_hard_mining(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_loss_with_hard_mining(self, use_keras=False):
model = self._build_model(is_training=True,
use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=None,
first_stage_max_proposals=6,
......@@ -1564,8 +1747,13 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllClose(loss_dict_out[
'Loss/BoxClassifierLoss/classification_loss'], 0)
def test_loss_with_hard_mining_and_losses_mask(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_loss_with_hard_mining_and_losses_mask(self, use_keras=False):
model = self._build_model(is_training=True,
use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=None,
first_stage_max_proposals=6,
......@@ -1678,7 +1866,11 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertAllClose(loss_dict_out[
'Loss/BoxClassifierLoss/classification_loss'], 0)
def test_restore_map_for_classification_ckpt(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_restore_map_for_classification_ckpt(self, use_keras=False):
# Define mock tensorflow classification graph and save variables.
test_graph_classification = tf.Graph()
with test_graph_classification.as_default():
......@@ -1699,7 +1891,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
test_graph_detection = tf.Graph()
with test_graph_detection.as_default():
model = self._build_model(
is_training=False, number_of_stages=2, second_stage_batch_size=6)
is_training=False, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
inputs_shape = (2, 20, 20, 3)
inputs = tf.to_float(tf.random_uniform(
......@@ -1716,12 +1909,17 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertNotIn(model.first_stage_feature_extractor_scope, var)
self.assertNotIn(model.second_stage_feature_extractor_scope, var)
def test_restore_map_for_detection_ckpt(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_restore_map_for_detection_ckpt(self, use_keras=False):
# Define first detection graph and save variables.
test_graph_detection1 = tf.Graph()
with test_graph_detection1.as_default():
model = self._build_model(
is_training=False, number_of_stages=2, second_stage_batch_size=6)
is_training=False, use_keras=use_keras,
number_of_stages=2, second_stage_batch_size=6)
inputs_shape = (2, 20, 20, 3)
inputs = tf.to_float(tf.random_uniform(
inputs_shape, minval=0, maxval=255, dtype=tf.int32))
......@@ -1739,7 +1937,8 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
# Define second detection graph and restore variables.
test_graph_detection2 = tf.Graph()
with test_graph_detection2.as_default():
model2 = self._build_model(is_training=False, number_of_stages=2,
model2 = self._build_model(is_training=False, use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=6, num_classes=42)
inputs_shape2 = (2, 20, 20, 3)
......@@ -1760,11 +1959,16 @@ class FasterRCNNMetaArchTestBase(test_case.TestCase, parameterized.TestCase):
self.assertNotIn(model2.first_stage_feature_extractor_scope, var)
self.assertNotIn(model2.second_stage_feature_extractor_scope, var)
def test_load_all_det_checkpoint_vars(self):
@parameterized.parameters(
{'use_keras': True},
{'use_keras': False}
)
def test_load_all_det_checkpoint_vars(self, use_keras=False):
test_graph_detection = tf.Graph()
with test_graph_detection.as_default():
model = self._build_model(
is_training=False,
use_keras=use_keras,
number_of_stages=2,
second_stage_batch_size=6,
num_classes=42)
......
research/object_detection/meta_architectures/rfcn_meta_arch.py
View file @ 80444539
......@@ -81,7 +81,8 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
add_summaries=True,
clip_anchors_to_image=False,
use_static_shapes=False,
resize_masks=False):
resize_masks=False,
freeze_batchnorm=False):
"""RFCNMetaArch Constructor.
Args:
......@@ -110,9 +111,12 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
denser resolutions. The atrous rate is used to compensate for the
denser feature maps by using an effectively larger receptive field.
(This should typically be set to 1).
first_stage_box_predictor_arg_scope_fn: A function to generate tf-slim
arg_scope for conv2d, separable_conv2d and fully_connected ops for the
RPN box predictor.
first_stage_box_predictor_arg_scope_fn: Either a
Keras layer hyperparams object or a function to construct tf-slim
arg_scope for conv2d, separable_conv2d and fully_connected ops. Used
for the RPN box predictor. If it is a keras hyperparams object the
RPN box predictor will be a Keras model. If it is a function to
construct an arg scope it will be a tf-slim box predictor.
first_stage_box_predictor_kernel_size: Kernel size to use for the
convolution op just prior to RPN box predictions.
first_stage_box_predictor_depth: Output depth for the convolution op
......@@ -180,6 +184,10 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
guarantees.
resize_masks: Indicates whether the masks presend in the groundtruth
should be resized in the model with `image_resizer_fn`
freeze_batchnorm: Whether to freeze batch norm parameters during
training or not. When training with a small batch size (e.g. 1), it is
desirable to freeze batch norm update and use pretrained batch norm
params.
Raises:
ValueError: If `second_stage_batch_size` > `first_stage_max_proposals`
......@@ -225,7 +233,8 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
add_summaries,
clip_anchors_to_image,
use_static_shapes,
resize_masks)
resize_masks,
freeze_batchnorm=freeze_batchnorm)
self._rfcn_box_predictor = second_stage_rfcn_box_predictor
......@@ -293,9 +302,7 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
anchors, image_shape_2d, true_image_shapes)
box_classifier_features = (
self._feature_extractor.extract_box_classifier_features(
rpn_features,
scope=self.second_stage_feature_extractor_scope))
self._extract_box_classifier_features(rpn_features))
if self._rfcn_box_predictor.is_keras_model:
box_predictions = self._rfcn_box_predictor(
......@@ -329,3 +336,37 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
'proposal_boxes_normalized': proposal_boxes_normalized,
}
return prediction_dict
def regularization_losses(self):
"""Returns a list of regularization losses for this model.
Returns a list of regularization losses for this model that the estimator
needs to use during training/optimization.
Returns:
A list of regularization loss tensors.
"""
reg_losses = super(RFCNMetaArch, self).regularization_losses()
if self._rfcn_box_predictor.is_keras_model:
reg_losses.extend(self._rfcn_box_predictor.losses)
return reg_losses
def updates(self):
"""Returns a list of update operators for this model.
Returns a list of update operators for this model that must be executed at
each training step. The estimator's train op needs to have a control
dependency on these updates.
Returns:
A list of update operators.
"""
update_ops = super(RFCNMetaArch, self).updates()
if self._rfcn_box_predictor.is_keras_model:
update_ops.extend(
self._rfcn_box_predictor.get_updates_for(None))
update_ops.extend(
self._rfcn_box_predictor.get_updates_for(
self._rfcn_box_predictor.inputs))
return update_ops
research/object_detection/meta_architectures/ssd_meta_arch.py
View file @ 80444539
......@@ -22,6 +22,7 @@ import tensorflow as tf
from object_detection.core import box_list
from object_detection.core import box_list_ops
from object_detection.core import matcher
from object_detection.core import model
from object_detection.core import standard_fields as fields
from object_detection.core import target_assigner
......@@ -663,8 +664,8 @@ class SSDMetaArch(model.DetectionModel):
raw_detection_boxes: [batch, total_detections, 4] tensor with decoded
detection boxes before Non-Max Suppression.
raw_detection_score: [batch, total_detections,
num_classes_with_background] tensor of multi-class score logits for
raw detection boxes.
num_classes_with_background] tensor of multi-class scores for raw
detection boxes.
Raises:
ValueError: if prediction_dict does not contain `box_encodings` or
`class_predictions_with_background` fields.
......@@ -672,17 +673,23 @@ class SSDMetaArch(model.DetectionModel):
if ('box_encodings' not in prediction_dict or
'class_predictions_with_background' not in prediction_dict):
raise ValueError('prediction_dict does not contain expected entries.')
if 'anchors' not in prediction_dict:
prediction_dict['anchors'] = self.anchors.get()
with tf.name_scope('Postprocessor'):
preprocessed_images = prediction_dict['preprocessed_inputs']
box_encodings = prediction_dict['box_encodings']
box_encodings = tf.identity(box_encodings, 'raw_box_encodings')
class_predictions = prediction_dict['class_predictions_with_background']
detection_boxes, detection_keypoints = self._batch_decode(box_encodings)
class_predictions_with_background = (
prediction_dict['class_predictions_with_background'])
detection_boxes, detection_keypoints = self._batch_decode(
box_encodings, prediction_dict['anchors'])
detection_boxes = tf.identity(detection_boxes, 'raw_box_locations')
detection_boxes = tf.expand_dims(detection_boxes, axis=2)
detection_scores = self._score_conversion_fn(class_predictions)
detection_scores = tf.identity(detection_scores, 'raw_box_scores')
detection_scores_with_background = self._score_conversion_fn(
class_predictions_with_background)
detection_scores = tf.identity(detection_scores_with_background,
'raw_box_scores')
if self._add_background_class or self._explicit_background_class:
detection_scores = tf.slice(detection_scores, [0, 0, 1], [-1, -1, -1])
additional_fields = None
......@@ -720,7 +727,7 @@ class SSDMetaArch(model.DetectionModel):
fields.DetectionResultFields.raw_detection_boxes:
tf.squeeze(detection_boxes, axis=2),
fields.DetectionResultFields.raw_detection_scores:
class_predictions
detection_scores_with_background
}
if (nmsed_additional_fields is not None and
fields.BoxListFields.keypoints in nmsed_additional_fields):
......@@ -767,10 +774,11 @@ class SSDMetaArch(model.DetectionModel):
if self.groundtruth_has_field(fields.BoxListFields.confidences):
confidences = self.groundtruth_lists(fields.BoxListFields.confidences)
(batch_cls_targets, batch_cls_weights, batch_reg_targets,
batch_reg_weights, match_list) = self._assign_targets(
batch_reg_weights, batch_match) = self._assign_targets(
self.groundtruth_lists(fields.BoxListFields.boxes),
self.groundtruth_lists(fields.BoxListFields.classes),
keypoints, weights, confidences)
match_list = [matcher.Match(match) for match in tf.unstack(batch_match)]
if self._add_summaries:
self._summarize_target_assignment(
self.groundtruth_lists(fields.BoxListFields.boxes), match_list)
......@@ -1017,25 +1025,31 @@ class SSDMetaArch(model.DetectionModel):
with rows of the Match objects corresponding to groundtruth boxes
and columns corresponding to anchors.
"""
num_boxes_per_image = tf.stack(
[tf.shape(x)[0] for x in groundtruth_boxes_list])
pos_anchors_per_image = tf.stack(
[match.num_matched_columns() for match in match_list])
neg_anchors_per_image = tf.stack(
[match.num_unmatched_columns() for match in match_list])
ignored_anchors_per_image = tf.stack(
[match.num_ignored_columns() for match in match_list])
avg_num_gt_boxes = tf.reduce_mean(tf.to_float(tf.stack(
[tf.shape(x)[0] for x in groundtruth_boxes_list])))
avg_num_matched_gt_boxes = tf.reduce_mean(tf.to_float(tf.stack(
[match.num_matched_rows() for match in match_list])))
avg_pos_anchors = tf.reduce_mean(tf.to_float(tf.stack(
[match.num_matched_columns() for match in match_list])))
avg_neg_anchors = tf.reduce_mean(tf.to_float(tf.stack(
[match.num_unmatched_columns() for match in match_list])))
avg_ignored_anchors = tf.reduce_mean(tf.to_float(tf.stack(
[match.num_ignored_columns() for match in match_list])))
# TODO(rathodv): Add a test for these summaries.
tf.summary.scalar('AvgNumGroundtruthBoxesPerImage',
tf.reduce_mean(tf.to_float(num_boxes_per_image)),
avg_num_gt_boxes,
family='TargetAssignment')
tf.summary.scalar('AvgNumGroundtruthBoxesMatchedPerImage',
avg_num_matched_gt_boxes,
family='TargetAssignment')
tf.summary.scalar('AvgNumPositiveAnchorsPerImage',
tf.reduce_mean(tf.to_float(pos_anchors_per_image)),
avg_pos_anchors,
family='TargetAssignment')
tf.summary.scalar('AvgNumNegativeAnchorsPerImage',
tf.reduce_mean(tf.to_float(neg_anchors_per_image)),
avg_neg_anchors,
family='TargetAssignment')
tf.summary.scalar('AvgNumIgnoredAnchorsPerImage',
tf.reduce_mean(tf.to_float(ignored_anchors_per_image)),
avg_ignored_anchors,
family='TargetAssignment')
def _apply_hard_mining(self, location_losses, cls_losses, prediction_dict,
......@@ -1054,6 +1068,7 @@ class SSDMetaArch(model.DetectionModel):
[batch_size, num_anchors, num_classes+1] containing class predictions
(logits) for each of the anchors. Note that this tensor *includes*
background class predictions.
3) anchors: (optional) 2-D float tensor of shape [num_anchors, 4].
match_list: a list of matcher.Match objects encoding the match between
anchors and groundtruth boxes for each image of the batch,
with rows of the Match objects corresponding to groundtruth boxes
......@@ -1069,7 +1084,10 @@ class SSDMetaArch(model.DetectionModel):
if self._add_background_class:
class_predictions = tf.slice(class_predictions, [0, 0, 1], [-1, -1, -1])
decoded_boxes, _ = self._batch_decode(prediction_dict['box_encodings'])
if 'anchors' not in prediction_dict:
prediction_dict['anchors'] = self.anchors.get()
decoded_boxes, _ = self._batch_decode(prediction_dict['box_encodings'],
prediction_dict['anchors'])
decoded_box_tensors_list = tf.unstack(decoded_boxes)
class_prediction_list = tf.unstack(class_predictions)
decoded_boxlist_list = []
......@@ -1084,12 +1102,13 @@ class SSDMetaArch(model.DetectionModel):
decoded_boxlist_list=decoded_boxlist_list,
match_list=match_list)
def _batch_decode(self, box_encodings):
def _batch_decode(self, box_encodings, anchors):
"""Decodes a batch of box encodings with respect to the anchors.
Args:
box_encodings: A float32 tensor of shape
[batch_size, num_anchors, box_code_size] containing box encodings.
anchors: A tensor of shape [num_anchors, 4].
Returns:
decoded_boxes: A float32 tensor of shape
......@@ -1101,8 +1120,7 @@ class SSDMetaArch(model.DetectionModel):
combined_shape = shape_utils.combined_static_and_dynamic_shape(
box_encodings)
batch_size = combined_shape[0]
tiled_anchor_boxes = tf.tile(
tf.expand_dims(self.anchors.get(), 0), [batch_size, 1, 1])
tiled_anchor_boxes = tf.tile(tf.expand_dims(anchors, 0), [batch_size, 1, 1])
tiled_anchors_boxlist = box_list.BoxList(
tf.reshape(tiled_anchor_boxes, [-1, 4]))
decoded_boxes = self._box_coder.decode(
......
research/object_detection/meta_architectures/ssd_meta_arch_test.py
View file @ 80444539
......@@ -311,8 +311,8 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
[0.5, 0., 1., 0.5], [1., 1., 1.5, 1.5]],
[[0., 0., 0.5, 0.5], [0., 0.5, 0.5, 1.],
[0.5, 0., 1., 0.5], [1., 1., 1.5, 1.5]]]
raw_detection_scores = [[[0, 0], [0, 0], [0, 0], [0, 0]],
[[0, 0], [0, 0], [0, 0], [0, 0]]]
raw_detection_scores = [[[0.5, 0.5], [0.5, 0.5], [0.5, 0.5], [0.5, 0.5]],
[[0.5, 0.5], [0.5, 0.5], [0.5, 0.5], [0.5, 0.5]]]
for input_shape in input_shapes:
tf_graph = tf.Graph()
......
research/object_detection/metrics/coco_evaluation.py
View file @ 80444539
......@@ -197,6 +197,7 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
'PerformanceByCategory' is included in the output regardless of
all_metrics_per_category.
"""
tf.logging.info('Performing evaluation on %d images.', len(self._image_ids))
groundtruth_dict = {
'annotations': self._groundtruth_list,
'images': [{'id': image_id} for image_id in self._image_ids],
......
research/object_detection/model_lib.py
View file @ 80444539
......@@ -24,6 +24,7 @@ import os
import tensorflow as tf
from tensorflow.python.util import function_utils
from object_detection import eval_util
from object_detection import exporter as exporter_lib
from object_detection import inputs
......@@ -33,6 +34,7 @@ from object_detection.builders import optimizer_builder
from object_detection.core import standard_fields as fields
from object_detection.utils import config_util
from object_detection.utils import label_map_util
from object_detection.utils import ops
from object_detection.utils import shape_utils
from object_detection.utils import variables_helper
from object_detection.utils import visualization_utils as vis_utils
......@@ -279,9 +281,7 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False,
prediction_dict = detection_model.predict(
preprocessed_images,
features[fields.InputDataFields.true_image_shape])
for k, v in prediction_dict.items():
if v.dtype == tf.bfloat16:
prediction_dict[k] = tf.cast(v, tf.float32)
prediction_dict = ops.bfloat16_to_float32_nested(prediction_dict)
else:
prediction_dict = detection_model.predict(
preprocessed_images,
......@@ -338,6 +338,9 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False,
losses = [loss_tensor for loss_tensor in losses_dict.values()]
if train_config.add_regularization_loss:
regularization_losses = detection_model.regularization_losses()
if use_tpu and train_config.use_bfloat16:
regularization_losses = ops.bfloat16_to_float32_nested(
regularization_losses)
if regularization_losses:
regularization_loss = tf.add_n(
regularization_losses, name='regularization_loss')
......@@ -650,6 +653,11 @@ def create_estimator_and_inputs(run_config,
model_fn = model_fn_creator(detection_model_fn, configs, hparams, use_tpu,
postprocess_on_cpu)
if use_tpu_estimator:
# Multicore inference disabled due to b/129367127
tpu_estimator_args = function_utils.fn_args(tf.contrib.tpu.TPUEstimator)
kwargs = {}
if 'experimental_export_device_assignment' in tpu_estimator_args:
kwargs['experimental_export_device_assignment'] = True
estimator = tf.contrib.tpu.TPUEstimator(
model_fn=model_fn,
train_batch_size=train_config.batch_size,
......@@ -659,7 +667,8 @@ def create_estimator_and_inputs(run_config,
config=run_config,
export_to_tpu=export_to_tpu,
eval_on_tpu=False, # Eval runs on CPU, so disable eval on TPU
params=params if params else {})
params=params if params else {},
**kwargs)
else:
estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config)
......
research/object_detection/models/faster_rcnn_inception_resnet_v2_keras_feature_extractor.py 0 → 100644
View file @ 80444539
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Inception Resnet v2 Faster R-CNN implementation in Keras.
See "Inception-v4, Inception-ResNet and the Impact of Residual Connections on
Learning" by Szegedy et al. (https://arxiv.org/abs/1602.07261)
as well as
"Speed/accuracy trade-offs for modern convolutional object detectors" by
Huang et al. (https://arxiv.org/abs/1611.10012)
"""
# Skip pylint for this file because it times out
# pylint: skip-file
import tensorflow as tf
from object_detection.meta_architectures import faster_rcnn_meta_arch
from object_detection.models.keras_models import inception_resnet_v2
from object_detection.utils import model_util
class FasterRCNNInceptionResnetV2KerasFeatureExtractor(
faster_rcnn_meta_arch.FasterRCNNKerasFeatureExtractor):
"""Faster R-CNN with Inception Resnet v2 feature extractor implementation."""
def __init__(self,
is_training,
first_stage_features_stride,
batch_norm_trainable=False,
weight_decay=0.0):
"""Constructor.
Args:
is_training: See base class.
first_stage_features_stride: See base class.
batch_norm_trainable: See base class.
weight_decay: See base class.
Raises:
ValueError: If `first_stage_features_stride` is not 8 or 16.
"""
if first_stage_features_stride != 8 and first_stage_features_stride != 16:
raise ValueError('`first_stage_features_stride` must be 8 or 16.')
super(FasterRCNNInceptionResnetV2KerasFeatureExtractor, self).__init__(
is_training, first_stage_features_stride, batch_norm_trainable,
weight_decay)
def preprocess(self, resized_inputs):
"""Faster R-CNN with Inception Resnet v2 preprocessing.
Maps pixel values to the range [-1, 1].
Args:
resized_inputs: A [batch, height_in, width_in, channels] float32 tensor
representing a batch of images with values between 0 and 255.0.
Returns:
preprocessed_inputs: A [batch, height_out, width_out, channels] float32
tensor representing a batch of images.
"""
return (2.0 / 255.0) * resized_inputs - 1.0
def get_proposal_feature_extractor_model(self, name=None):
"""Returns a model that extracts first stage RPN features.
Extracts features using the first half of the Inception Resnet v2 network.
We construct the network in `align_feature_maps=True` mode, which means
that all VALID paddings in the network are changed to SAME padding so that
the feature maps are aligned.
Args:
name: A scope name to construct all variables within.
Returns:
A Keras model that takes preprocessed_inputs:
A [batch, height, width, channels] float32 tensor
representing a batch of images.
And returns rpn_feature_map:
A tensor with shape [batch, height, width, depth]
"""
with tf.name_scope(name):
with tf.name_scope('InceptionResnetV2'):
model = inception_resnet_v2.inception_resnet_v2(
self._train_batch_norm,
output_stride=self._first_stage_features_stride,
align_feature_maps=True,
weight_decay=self._weight_decay,
weights=None,
include_top=False)
proposal_features = model.get_layer(
name='block17_20_ac').output
return tf.keras.Model(
inputs=model.inputs,
outputs=proposal_features)
def get_box_classifier_feature_extractor_model(self, name=None):
"""Returns a model that extracts second stage box classifier features.
This function reconstructs the "second half" of the Inception ResNet v2
network after the part defined in `get_proposal_feature_extractor_model`.
Args:
name: A scope name to construct all variables within.
Returns:
A Keras model that takes proposal_feature_maps:
A 4-D float tensor with shape
[batch_size * self.max_num_proposals, crop_height, crop_width, depth]
representing the feature map cropped to each proposal.
And returns proposal_classifier_features:
A 4-D float tensor with shape
[batch_size * self.max_num_proposals, height, width, depth]
representing box classifier features for each proposal.
"""
with tf.name_scope(name):
with tf.name_scope('InceptionResnetV2'):
model = inception_resnet_v2.inception_resnet_v2(
self._train_batch_norm,
output_stride=16,
align_feature_maps=False,
weight_decay=self._weight_decay,
weights=None,
include_top=False)
proposal_feature_maps = model.get_layer(
name='block17_20_ac').output
proposal_classifier_features = model.get_layer(
name='conv_7b_ac').output
return model_util.extract_submodel(
model=model,
inputs=proposal_feature_maps,
outputs=proposal_classifier_features)
def restore_from_classification_checkpoint_fn(
self,
first_stage_feature_extractor_scope,
second_stage_feature_extractor_scope):
"""Returns a map of variables to load from a foreign checkpoint.
This uses a hard-coded conversion to load into Keras from a slim-trained
inception_resnet_v2 checkpoint.
Note that this overrides the default implementation in
faster_rcnn_meta_arch.FasterRCNNKerasFeatureExtractor which does not work
for InceptionResnetV2 checkpoints.
Args:
first_stage_feature_extractor_scope: A scope name for the first stage
feature extractor.
second_stage_feature_extractor_scope: A scope name for the second stage
feature extractor.
Returns:
A dict mapping variable names (to load from a checkpoint) to variables in
the model graph.
"""
keras_to_slim_name_mapping = {
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d/kernel': 'InceptionResnetV2/Conv2d_1a_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm/beta': 'InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm/moving_mean': 'InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm/moving_variance': 'InceptionResnetV2/Conv2d_1a_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_1/kernel': 'InceptionResnetV2/Conv2d_2a_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_1/beta': 'InceptionResnetV2/Conv2d_2a_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_1/moving_mean': 'InceptionResnetV2/Conv2d_2a_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_1/moving_variance': 'InceptionResnetV2/Conv2d_2a_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_2/kernel': 'InceptionResnetV2/Conv2d_2b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_2/beta': 'InceptionResnetV2/Conv2d_2b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_2/moving_mean': 'InceptionResnetV2/Conv2d_2b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_2/moving_variance': 'InceptionResnetV2/Conv2d_2b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_3/kernel': 'InceptionResnetV2/Conv2d_3b_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_3/beta': 'InceptionResnetV2/Conv2d_3b_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_3/moving_mean': 'InceptionResnetV2/Conv2d_3b_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_3/moving_variance': 'InceptionResnetV2/Conv2d_3b_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_4/kernel': 'InceptionResnetV2/Conv2d_4a_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_4/beta': 'InceptionResnetV2/Conv2d_4a_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_4/moving_mean': 'InceptionResnetV2/Conv2d_4a_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_4/moving_variance': 'InceptionResnetV2/Conv2d_4a_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_5/kernel': 'InceptionResnetV2/Mixed_5b/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_5/beta': 'InceptionResnetV2/Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_5/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_5/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_6/kernel': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_6/beta': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_6/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_6/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_7/kernel': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0b_5x5/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_7/beta': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_7/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_7/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_1/Conv2d_0b_5x5/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_8/kernel': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_8/beta': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_8/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_8/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_9/kernel': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_9/beta': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_9/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_9/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_10/kernel': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_10/beta': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_10/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_10/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_11/kernel': 'InceptionResnetV2/Mixed_5b/Branch_3/Conv2d_0b_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_11/beta': 'InceptionResnetV2/Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_11/moving_mean': 'InceptionResnetV2/Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_11/moving_variance': 'InceptionResnetV2/Mixed_5b/Branch_3/Conv2d_0b_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_12/kernel': 'InceptionResnetV2/Repeat/block35_1/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_12/beta': 'InceptionResnetV2/Repeat/block35_1/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_12/moving_mean': 'InceptionResnetV2/Repeat/block35_1/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_12/moving_variance': 'InceptionResnetV2/Repeat/block35_1/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_13/kernel': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_13/beta': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_13/moving_mean': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_13/moving_variance': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_14/kernel': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_14/beta': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_14/moving_mean': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_14/moving_variance': 'InceptionResnetV2/Repeat/block35_1/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_15/kernel': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_15/beta': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_15/moving_mean': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_15/moving_variance': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_16/kernel': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_16/beta': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_16/moving_mean': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_16/moving_variance': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_17/kernel': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_17/beta': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_17/moving_mean': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_17/moving_variance': 'InceptionResnetV2/Repeat/block35_1/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_1_conv/kernel': 'InceptionResnetV2/Repeat/block35_1/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_1_conv/bias': 'InceptionResnetV2/Repeat/block35_1/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_18/kernel': 'InceptionResnetV2/Repeat/block35_2/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_18/beta': 'InceptionResnetV2/Repeat/block35_2/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_18/moving_mean': 'InceptionResnetV2/Repeat/block35_2/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_18/moving_variance': 'InceptionResnetV2/Repeat/block35_2/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_19/kernel': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_19/beta': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_19/moving_mean': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_19/moving_variance': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_20/kernel': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_20/beta': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_20/moving_mean': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_20/moving_variance': 'InceptionResnetV2/Repeat/block35_2/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_21/kernel': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_21/beta': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_21/moving_mean': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_21/moving_variance': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_22/kernel': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_22/beta': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_22/moving_mean': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_22/moving_variance': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_23/kernel': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_23/beta': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_23/moving_mean': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_23/moving_variance': 'InceptionResnetV2/Repeat/block35_2/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_2_conv/kernel': 'InceptionResnetV2/Repeat/block35_2/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_2_conv/bias': 'InceptionResnetV2/Repeat/block35_2/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_24/kernel': 'InceptionResnetV2/Repeat/block35_3/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_24/beta': 'InceptionResnetV2/Repeat/block35_3/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_24/moving_mean': 'InceptionResnetV2/Repeat/block35_3/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_24/moving_variance': 'InceptionResnetV2/Repeat/block35_3/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_25/kernel': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_25/beta': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_25/moving_mean': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_25/moving_variance': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_26/kernel': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_26/beta': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_26/moving_mean': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_26/moving_variance': 'InceptionResnetV2/Repeat/block35_3/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_27/kernel': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_27/beta': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_27/moving_mean': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_27/moving_variance': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_28/kernel': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_28/beta': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_28/moving_mean': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_28/moving_variance': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_29/kernel': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_29/beta': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_29/moving_mean': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_29/moving_variance': 'InceptionResnetV2/Repeat/block35_3/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_3_conv/kernel': 'InceptionResnetV2/Repeat/block35_3/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_3_conv/bias': 'InceptionResnetV2/Repeat/block35_3/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_30/kernel': 'InceptionResnetV2/Repeat/block35_4/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_30/beta': 'InceptionResnetV2/Repeat/block35_4/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_30/moving_mean': 'InceptionResnetV2/Repeat/block35_4/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_30/moving_variance': 'InceptionResnetV2/Repeat/block35_4/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_31/kernel': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_31/beta': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_31/moving_mean': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_31/moving_variance': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_32/kernel': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_32/beta': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_32/moving_mean': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_32/moving_variance': 'InceptionResnetV2/Repeat/block35_4/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_33/kernel': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_33/beta': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_33/moving_mean': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_33/moving_variance': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_34/kernel': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_34/beta': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_34/moving_mean': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_34/moving_variance': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_35/kernel': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_35/beta': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_35/moving_mean': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_35/moving_variance': 'InceptionResnetV2/Repeat/block35_4/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_4_conv/kernel': 'InceptionResnetV2/Repeat/block35_4/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_4_conv/bias': 'InceptionResnetV2/Repeat/block35_4/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_36/kernel': 'InceptionResnetV2/Repeat/block35_5/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_36/beta': 'InceptionResnetV2/Repeat/block35_5/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_36/moving_mean': 'InceptionResnetV2/Repeat/block35_5/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_36/moving_variance': 'InceptionResnetV2/Repeat/block35_5/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_37/kernel': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_37/beta': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_37/moving_mean': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_37/moving_variance': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_38/kernel': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_38/beta': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_38/moving_mean': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_38/moving_variance': 'InceptionResnetV2/Repeat/block35_5/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_39/kernel': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_39/beta': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_39/moving_mean': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_39/moving_variance': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_40/kernel': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_40/beta': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_40/moving_mean': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_40/moving_variance': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_41/kernel': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_41/beta': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_41/moving_mean': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_41/moving_variance': 'InceptionResnetV2/Repeat/block35_5/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_5_conv/kernel': 'InceptionResnetV2/Repeat/block35_5/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_5_conv/bias': 'InceptionResnetV2/Repeat/block35_5/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_42/kernel': 'InceptionResnetV2/Repeat/block35_6/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_42/beta': 'InceptionResnetV2/Repeat/block35_6/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_42/moving_mean': 'InceptionResnetV2/Repeat/block35_6/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_42/moving_variance': 'InceptionResnetV2/Repeat/block35_6/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_43/kernel': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_43/beta': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_43/moving_mean': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_43/moving_variance': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_44/kernel': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_44/beta': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_44/moving_mean': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_44/moving_variance': 'InceptionResnetV2/Repeat/block35_6/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_45/kernel': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_45/beta': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_45/moving_mean': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_45/moving_variance': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_46/kernel': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_46/beta': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_46/moving_mean': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_46/moving_variance': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_47/kernel': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_47/beta': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_47/moving_mean': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_47/moving_variance': 'InceptionResnetV2/Repeat/block35_6/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_6_conv/kernel': 'InceptionResnetV2/Repeat/block35_6/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_6_conv/bias': 'InceptionResnetV2/Repeat/block35_6/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_48/kernel': 'InceptionResnetV2/Repeat/block35_7/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_48/beta': 'InceptionResnetV2/Repeat/block35_7/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_48/moving_mean': 'InceptionResnetV2/Repeat/block35_7/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_48/moving_variance': 'InceptionResnetV2/Repeat/block35_7/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_49/kernel': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_49/beta': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_49/moving_mean': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_49/moving_variance': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_50/kernel': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_50/beta': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_50/moving_mean': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_50/moving_variance': 'InceptionResnetV2/Repeat/block35_7/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_51/kernel': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_51/beta': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_51/moving_mean': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_51/moving_variance': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_52/kernel': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_52/beta': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_52/moving_mean': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_52/moving_variance': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_53/kernel': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_53/beta': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_53/moving_mean': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_53/moving_variance': 'InceptionResnetV2/Repeat/block35_7/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_7_conv/kernel': 'InceptionResnetV2/Repeat/block35_7/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_7_conv/bias': 'InceptionResnetV2/Repeat/block35_7/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_54/kernel': 'InceptionResnetV2/Repeat/block35_8/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_54/beta': 'InceptionResnetV2/Repeat/block35_8/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_54/moving_mean': 'InceptionResnetV2/Repeat/block35_8/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_54/moving_variance': 'InceptionResnetV2/Repeat/block35_8/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_55/kernel': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_55/beta': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_55/moving_mean': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_55/moving_variance': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_56/kernel': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_56/beta': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_56/moving_mean': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_56/moving_variance': 'InceptionResnetV2/Repeat/block35_8/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_57/kernel': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_57/beta': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_57/moving_mean': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_57/moving_variance': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_58/kernel': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_58/beta': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_58/moving_mean': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_58/moving_variance': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_59/kernel': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_59/beta': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_59/moving_mean': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_59/moving_variance': 'InceptionResnetV2/Repeat/block35_8/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_8_conv/kernel': 'InceptionResnetV2/Repeat/block35_8/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_8_conv/bias': 'InceptionResnetV2/Repeat/block35_8/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_60/kernel': 'InceptionResnetV2/Repeat/block35_9/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_60/beta': 'InceptionResnetV2/Repeat/block35_9/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_60/moving_mean': 'InceptionResnetV2/Repeat/block35_9/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_60/moving_variance': 'InceptionResnetV2/Repeat/block35_9/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_61/kernel': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_61/beta': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_61/moving_mean': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_61/moving_variance': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_62/kernel': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_62/beta': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_62/moving_mean': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_62/moving_variance': 'InceptionResnetV2/Repeat/block35_9/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_63/kernel': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_63/beta': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_63/moving_mean': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_63/moving_variance': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_64/kernel': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_64/beta': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_64/moving_mean': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_64/moving_variance': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_65/kernel': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_65/beta': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_65/moving_mean': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_65/moving_variance': 'InceptionResnetV2/Repeat/block35_9/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_9_conv/kernel': 'InceptionResnetV2/Repeat/block35_9/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_9_conv/bias': 'InceptionResnetV2/Repeat/block35_9/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_66/kernel': 'InceptionResnetV2/Repeat/block35_10/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_66/beta': 'InceptionResnetV2/Repeat/block35_10/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_66/moving_mean': 'InceptionResnetV2/Repeat/block35_10/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_66/moving_variance': 'InceptionResnetV2/Repeat/block35_10/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_67/kernel': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_67/beta': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_67/moving_mean': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_67/moving_variance': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_68/kernel': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_68/beta': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_68/moving_mean': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_68/moving_variance': 'InceptionResnetV2/Repeat/block35_10/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_69/kernel': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_69/beta': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_69/moving_mean': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_69/moving_variance': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_70/kernel': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_70/beta': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_70/moving_mean': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_70/moving_variance': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_71/kernel': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0c_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_71/beta': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0c_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_71/moving_mean': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_71/moving_variance': 'InceptionResnetV2/Repeat/block35_10/Branch_2/Conv2d_0c_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_10_conv/kernel': 'InceptionResnetV2/Repeat/block35_10/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block35_10_conv/bias': 'InceptionResnetV2/Repeat/block35_10/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_72/kernel': 'InceptionResnetV2/Mixed_6a/Branch_0/Conv2d_1a_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_72/beta': 'InceptionResnetV2/Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_72/moving_mean': 'InceptionResnetV2/Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_72/moving_variance': 'InceptionResnetV2/Mixed_6a/Branch_0/Conv2d_1a_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_73/kernel': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_73/beta': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_73/moving_mean': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_73/moving_variance': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_74/kernel': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0b_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_74/beta': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_74/moving_mean': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_74/moving_variance': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_0b_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_75/kernel': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_1a_3x3/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_75/beta': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_75/moving_mean': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_75/moving_variance': 'InceptionResnetV2/Mixed_6a/Branch_1/Conv2d_1a_3x3/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_76/kernel': 'InceptionResnetV2/Repeat_1/block17_1/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_76/beta': 'InceptionResnetV2/Repeat_1/block17_1/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_76/moving_mean': 'InceptionResnetV2/Repeat_1/block17_1/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_76/moving_variance': 'InceptionResnetV2/Repeat_1/block17_1/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_77/kernel': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_77/beta': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_77/moving_mean': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_77/moving_variance': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_78/kernel': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_78/beta': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_78/moving_mean': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_78/moving_variance': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_79/kernel': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_79/beta': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_79/moving_mean': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_79/moving_variance': 'InceptionResnetV2/Repeat_1/block17_1/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_1_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_1/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_1_conv/bias': 'InceptionResnetV2/Repeat_1/block17_1/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_80/kernel': 'InceptionResnetV2/Repeat_1/block17_2/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_80/beta': 'InceptionResnetV2/Repeat_1/block17_2/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_80/moving_mean': 'InceptionResnetV2/Repeat_1/block17_2/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_80/moving_variance': 'InceptionResnetV2/Repeat_1/block17_2/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_81/kernel': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_81/beta': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_81/moving_mean': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_81/moving_variance': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_82/kernel': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_82/beta': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_82/moving_mean': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_82/moving_variance': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_83/kernel': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_83/beta': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_83/moving_mean': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_83/moving_variance': 'InceptionResnetV2/Repeat_1/block17_2/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_2_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_2/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_2_conv/bias': 'InceptionResnetV2/Repeat_1/block17_2/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_84/kernel': 'InceptionResnetV2/Repeat_1/block17_3/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_84/beta': 'InceptionResnetV2/Repeat_1/block17_3/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_84/moving_mean': 'InceptionResnetV2/Repeat_1/block17_3/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_84/moving_variance': 'InceptionResnetV2/Repeat_1/block17_3/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_85/kernel': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_85/beta': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_85/moving_mean': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_85/moving_variance': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_86/kernel': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_86/beta': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_86/moving_mean': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_86/moving_variance': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_87/kernel': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_87/beta': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_87/moving_mean': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_87/moving_variance': 'InceptionResnetV2/Repeat_1/block17_3/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_3_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_3/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_3_conv/bias': 'InceptionResnetV2/Repeat_1/block17_3/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_88/kernel': 'InceptionResnetV2/Repeat_1/block17_4/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_88/beta': 'InceptionResnetV2/Repeat_1/block17_4/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_88/moving_mean': 'InceptionResnetV2/Repeat_1/block17_4/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_88/moving_variance': 'InceptionResnetV2/Repeat_1/block17_4/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_89/kernel': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_89/beta': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_89/moving_mean': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_89/moving_variance': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_90/kernel': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_90/beta': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_90/moving_mean': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_90/moving_variance': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_91/kernel': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_91/beta': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_91/moving_mean': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_91/moving_variance': 'InceptionResnetV2/Repeat_1/block17_4/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_4_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_4/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_4_conv/bias': 'InceptionResnetV2/Repeat_1/block17_4/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_92/kernel': 'InceptionResnetV2/Repeat_1/block17_5/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_92/beta': 'InceptionResnetV2/Repeat_1/block17_5/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_92/moving_mean': 'InceptionResnetV2/Repeat_1/block17_5/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_92/moving_variance': 'InceptionResnetV2/Repeat_1/block17_5/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_93/kernel': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_93/beta': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_93/moving_mean': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_93/moving_variance': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_94/kernel': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_94/beta': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_94/moving_mean': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_94/moving_variance': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_95/kernel': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_95/beta': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_95/moving_mean': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_95/moving_variance': 'InceptionResnetV2/Repeat_1/block17_5/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_5_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_5/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_5_conv/bias': 'InceptionResnetV2/Repeat_1/block17_5/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_96/kernel': 'InceptionResnetV2/Repeat_1/block17_6/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_96/beta': 'InceptionResnetV2/Repeat_1/block17_6/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_96/moving_mean': 'InceptionResnetV2/Repeat_1/block17_6/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_96/moving_variance': 'InceptionResnetV2/Repeat_1/block17_6/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_97/kernel': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_97/beta': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_97/moving_mean': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_97/moving_variance': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_98/kernel': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_98/beta': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_98/moving_mean': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_98/moving_variance': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_99/kernel': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_99/beta': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_99/moving_mean': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_99/moving_variance': 'InceptionResnetV2/Repeat_1/block17_6/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_6_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_6/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_6_conv/bias': 'InceptionResnetV2/Repeat_1/block17_6/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_100/kernel': 'InceptionResnetV2/Repeat_1/block17_7/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_100/beta': 'InceptionResnetV2/Repeat_1/block17_7/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_100/moving_mean': 'InceptionResnetV2/Repeat_1/block17_7/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_100/moving_variance': 'InceptionResnetV2/Repeat_1/block17_7/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_101/kernel': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_101/beta': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_101/moving_mean': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_101/moving_variance': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_102/kernel': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_102/beta': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_102/moving_mean': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_102/moving_variance': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_103/kernel': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_103/beta': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_103/moving_mean': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_103/moving_variance': 'InceptionResnetV2/Repeat_1/block17_7/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_7_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_7/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_7_conv/bias': 'InceptionResnetV2/Repeat_1/block17_7/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_104/kernel': 'InceptionResnetV2/Repeat_1/block17_8/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_104/beta': 'InceptionResnetV2/Repeat_1/block17_8/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_104/moving_mean': 'InceptionResnetV2/Repeat_1/block17_8/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_104/moving_variance': 'InceptionResnetV2/Repeat_1/block17_8/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_105/kernel': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_105/beta': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_105/moving_mean': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_105/moving_variance': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_106/kernel': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_106/beta': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_106/moving_mean': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_106/moving_variance': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_107/kernel': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_107/beta': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_107/moving_mean': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_107/moving_variance': 'InceptionResnetV2/Repeat_1/block17_8/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_8_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_8/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_8_conv/bias': 'InceptionResnetV2/Repeat_1/block17_8/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_108/kernel': 'InceptionResnetV2/Repeat_1/block17_9/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_108/beta': 'InceptionResnetV2/Repeat_1/block17_9/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_108/moving_mean': 'InceptionResnetV2/Repeat_1/block17_9/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_108/moving_variance': 'InceptionResnetV2/Repeat_1/block17_9/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_109/kernel': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_109/beta': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_109/moving_mean': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_109/moving_variance': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_110/kernel': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_110/beta': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_110/moving_mean': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_110/moving_variance': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_111/kernel': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_111/beta': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_111/moving_mean': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_111/moving_variance': 'InceptionResnetV2/Repeat_1/block17_9/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_9_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_9/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_9_conv/bias': 'InceptionResnetV2/Repeat_1/block17_9/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_112/kernel': 'InceptionResnetV2/Repeat_1/block17_10/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_112/beta': 'InceptionResnetV2/Repeat_1/block17_10/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_112/moving_mean': 'InceptionResnetV2/Repeat_1/block17_10/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_112/moving_variance': 'InceptionResnetV2/Repeat_1/block17_10/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_113/kernel': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_113/beta': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_113/moving_mean': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_113/moving_variance': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_114/kernel': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_114/beta': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_114/moving_mean': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_114/moving_variance': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_115/kernel': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_115/beta': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_115/moving_mean': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_115/moving_variance': 'InceptionResnetV2/Repeat_1/block17_10/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_10_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_10/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_10_conv/bias': 'InceptionResnetV2/Repeat_1/block17_10/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_116/kernel': 'InceptionResnetV2/Repeat_1/block17_11/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_116/beta': 'InceptionResnetV2/Repeat_1/block17_11/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_116/moving_mean': 'InceptionResnetV2/Repeat_1/block17_11/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_116/moving_variance': 'InceptionResnetV2/Repeat_1/block17_11/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_117/kernel': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_117/beta': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_117/moving_mean': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_117/moving_variance': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_118/kernel': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_118/beta': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_118/moving_mean': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_118/moving_variance': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_119/kernel': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_119/beta': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_119/moving_mean': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_119/moving_variance': 'InceptionResnetV2/Repeat_1/block17_11/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_11_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_11/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_11_conv/bias': 'InceptionResnetV2/Repeat_1/block17_11/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_120/kernel': 'InceptionResnetV2/Repeat_1/block17_12/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_120/beta': 'InceptionResnetV2/Repeat_1/block17_12/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_120/moving_mean': 'InceptionResnetV2/Repeat_1/block17_12/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_120/moving_variance': 'InceptionResnetV2/Repeat_1/block17_12/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_121/kernel': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_121/beta': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_121/moving_mean': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_121/moving_variance': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_122/kernel': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_122/beta': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_122/moving_mean': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_122/moving_variance': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_123/kernel': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_123/beta': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_123/moving_mean': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_123/moving_variance': 'InceptionResnetV2/Repeat_1/block17_12/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_12_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_12/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_12_conv/bias': 'InceptionResnetV2/Repeat_1/block17_12/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_124/kernel': 'InceptionResnetV2/Repeat_1/block17_13/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_124/beta': 'InceptionResnetV2/Repeat_1/block17_13/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_124/moving_mean': 'InceptionResnetV2/Repeat_1/block17_13/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_124/moving_variance': 'InceptionResnetV2/Repeat_1/block17_13/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_125/kernel': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_125/beta': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_125/moving_mean': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_125/moving_variance': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_126/kernel': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_126/beta': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_126/moving_mean': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_126/moving_variance': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_127/kernel': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_127/beta': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_127/moving_mean': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_127/moving_variance': 'InceptionResnetV2/Repeat_1/block17_13/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_13_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_13/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_13_conv/bias': 'InceptionResnetV2/Repeat_1/block17_13/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_128/kernel': 'InceptionResnetV2/Repeat_1/block17_14/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_128/beta': 'InceptionResnetV2/Repeat_1/block17_14/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_128/moving_mean': 'InceptionResnetV2/Repeat_1/block17_14/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_128/moving_variance': 'InceptionResnetV2/Repeat_1/block17_14/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_129/kernel': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_129/beta': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_129/moving_mean': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_129/moving_variance': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_130/kernel': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_130/beta': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_130/moving_mean': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_130/moving_variance': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_131/kernel': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_131/beta': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_131/moving_mean': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_131/moving_variance': 'InceptionResnetV2/Repeat_1/block17_14/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_14_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_14/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_14_conv/bias': 'InceptionResnetV2/Repeat_1/block17_14/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_132/kernel': 'InceptionResnetV2/Repeat_1/block17_15/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_132/beta': 'InceptionResnetV2/Repeat_1/block17_15/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_132/moving_mean': 'InceptionResnetV2/Repeat_1/block17_15/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_132/moving_variance': 'InceptionResnetV2/Repeat_1/block17_15/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_133/kernel': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_133/beta': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_133/moving_mean': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_133/moving_variance': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_134/kernel': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_134/beta': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_134/moving_mean': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_134/moving_variance': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_135/kernel': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_135/beta': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_135/moving_mean': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_135/moving_variance': 'InceptionResnetV2/Repeat_1/block17_15/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_15_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_15/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_15_conv/bias': 'InceptionResnetV2/Repeat_1/block17_15/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_136/kernel': 'InceptionResnetV2/Repeat_1/block17_16/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_136/beta': 'InceptionResnetV2/Repeat_1/block17_16/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_136/moving_mean': 'InceptionResnetV2/Repeat_1/block17_16/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_136/moving_variance': 'InceptionResnetV2/Repeat_1/block17_16/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_137/kernel': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_137/beta': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_137/moving_mean': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_137/moving_variance': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_138/kernel': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_138/beta': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_138/moving_mean': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_138/moving_variance': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_139/kernel': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_139/beta': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_139/moving_mean': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_139/moving_variance': 'InceptionResnetV2/Repeat_1/block17_16/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_16_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_16/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_16_conv/bias': 'InceptionResnetV2/Repeat_1/block17_16/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_140/kernel': 'InceptionResnetV2/Repeat_1/block17_17/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_140/beta': 'InceptionResnetV2/Repeat_1/block17_17/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_140/moving_mean': 'InceptionResnetV2/Repeat_1/block17_17/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_140/moving_variance': 'InceptionResnetV2/Repeat_1/block17_17/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_141/kernel': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_141/beta': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_141/moving_mean': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_141/moving_variance': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_142/kernel': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_142/beta': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_142/moving_mean': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_142/moving_variance': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_143/kernel': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_143/beta': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_143/moving_mean': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_143/moving_variance': 'InceptionResnetV2/Repeat_1/block17_17/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_17_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_17/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_17_conv/bias': 'InceptionResnetV2/Repeat_1/block17_17/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_144/kernel': 'InceptionResnetV2/Repeat_1/block17_18/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_144/beta': 'InceptionResnetV2/Repeat_1/block17_18/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_144/moving_mean': 'InceptionResnetV2/Repeat_1/block17_18/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_144/moving_variance': 'InceptionResnetV2/Repeat_1/block17_18/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_145/kernel': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_145/beta': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_145/moving_mean': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_145/moving_variance': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_146/kernel': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_146/beta': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_146/moving_mean': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_146/moving_variance': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_147/kernel': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_147/beta': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_147/moving_mean': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_147/moving_variance': 'InceptionResnetV2/Repeat_1/block17_18/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_18_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_18/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_18_conv/bias': 'InceptionResnetV2/Repeat_1/block17_18/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_148/kernel': 'InceptionResnetV2/Repeat_1/block17_19/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_148/beta': 'InceptionResnetV2/Repeat_1/block17_19/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_148/moving_mean': 'InceptionResnetV2/Repeat_1/block17_19/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_148/moving_variance': 'InceptionResnetV2/Repeat_1/block17_19/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_149/kernel': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_149/beta': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_149/moving_mean': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_149/moving_variance': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_150/kernel': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_150/beta': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_150/moving_mean': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_150/moving_variance': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_151/kernel': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_151/beta': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_151/moving_mean': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_151/moving_variance': 'InceptionResnetV2/Repeat_1/block17_19/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_19_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_19/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_19_conv/bias': 'InceptionResnetV2/Repeat_1/block17_19/Conv2d_1x1/biases',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_152/kernel': 'InceptionResnetV2/Repeat_1/block17_20/Branch_0/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_152/beta': 'InceptionResnetV2/Repeat_1/block17_20/Branch_0/Conv2d_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_152/moving_mean': 'InceptionResnetV2/Repeat_1/block17_20/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_152/moving_variance': 'InceptionResnetV2/Repeat_1/block17_20/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_153/kernel': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0a_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_153/beta': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_153/moving_mean': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_153/moving_variance': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_154/kernel': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0b_1x7/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_154/beta': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0b_1x7/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_154/moving_mean': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_154/moving_variance': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0b_1x7/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/conv2d_155/kernel': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0c_7x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_155/beta': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0c_7x1/BatchNorm/beta',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_155/moving_mean': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_mean',
'FirstStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_155/moving_variance': 'InceptionResnetV2/Repeat_1/block17_20/Branch_1/Conv2d_0c_7x1/BatchNorm/moving_variance',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_20_conv/kernel': 'InceptionResnetV2/Repeat_1/block17_20/Conv2d_1x1/weights',
'FirstStageFeatureExtractor/InceptionResnetV2/block17_20_conv/bias': 'InceptionResnetV2/Repeat_1/block17_20/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_359/kernel': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_359/beta': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_359/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_359/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_360/kernel': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_1a_3x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_360/beta': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_360/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_360/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_0/Conv2d_1a_3x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_361/kernel': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_361/beta': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_361/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_361/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_362/kernel': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_1a_3x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_362/beta': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_362/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_362/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_1/Conv2d_1a_3x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_363/kernel': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_363/beta': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_363/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_363/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_364/kernel': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0b_3x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_364/beta': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_364/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_364/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_0b_3x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_365/kernel': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_1a_3x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_365/beta': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_365/moving_mean': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_365/moving_variance': 'InceptionResnetV2/Mixed_7a/Branch_2/Conv2d_1a_3x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_366/kernel': 'InceptionResnetV2/Repeat_2/block8_1/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_366/beta': 'InceptionResnetV2/Repeat_2/block8_1/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_366/moving_mean': 'InceptionResnetV2/Repeat_2/block8_1/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_366/moving_variance': 'InceptionResnetV2/Repeat_2/block8_1/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_367/kernel': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_367/beta': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_367/moving_mean': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_367/moving_variance': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_368/kernel': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_368/beta': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_368/moving_mean': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_368/moving_variance': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_369/kernel': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_369/beta': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_369/moving_mean': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_369/moving_variance': 'InceptionResnetV2/Repeat_2/block8_1/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_1_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_1/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_1_conv/bias': 'InceptionResnetV2/Repeat_2/block8_1/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_370/kernel': 'InceptionResnetV2/Repeat_2/block8_2/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_370/beta': 'InceptionResnetV2/Repeat_2/block8_2/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_370/moving_mean': 'InceptionResnetV2/Repeat_2/block8_2/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_370/moving_variance': 'InceptionResnetV2/Repeat_2/block8_2/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_371/kernel': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_371/beta': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_371/moving_mean': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_371/moving_variance': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_372/kernel': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_372/beta': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_372/moving_mean': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_372/moving_variance': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_373/kernel': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_373/beta': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_373/moving_mean': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_373/moving_variance': 'InceptionResnetV2/Repeat_2/block8_2/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_2_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_2/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_2_conv/bias': 'InceptionResnetV2/Repeat_2/block8_2/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_374/kernel': 'InceptionResnetV2/Repeat_2/block8_3/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_374/beta': 'InceptionResnetV2/Repeat_2/block8_3/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_374/moving_mean': 'InceptionResnetV2/Repeat_2/block8_3/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_374/moving_variance': 'InceptionResnetV2/Repeat_2/block8_3/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_375/kernel': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_375/beta': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_375/moving_mean': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_375/moving_variance': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_376/kernel': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_376/beta': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_376/moving_mean': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_376/moving_variance': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_377/kernel': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_377/beta': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_377/moving_mean': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_377/moving_variance': 'InceptionResnetV2/Repeat_2/block8_3/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_3_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_3/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_3_conv/bias': 'InceptionResnetV2/Repeat_2/block8_3/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_378/kernel': 'InceptionResnetV2/Repeat_2/block8_4/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_378/beta': 'InceptionResnetV2/Repeat_2/block8_4/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_378/moving_mean': 'InceptionResnetV2/Repeat_2/block8_4/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_378/moving_variance': 'InceptionResnetV2/Repeat_2/block8_4/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_379/kernel': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_379/beta': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_379/moving_mean': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_379/moving_variance': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_380/kernel': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_380/beta': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_380/moving_mean': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_380/moving_variance': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_381/kernel': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_381/beta': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_381/moving_mean': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_381/moving_variance': 'InceptionResnetV2/Repeat_2/block8_4/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_4_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_4/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_4_conv/bias': 'InceptionResnetV2/Repeat_2/block8_4/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_382/kernel': 'InceptionResnetV2/Repeat_2/block8_5/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_382/beta': 'InceptionResnetV2/Repeat_2/block8_5/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_382/moving_mean': 'InceptionResnetV2/Repeat_2/block8_5/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_382/moving_variance': 'InceptionResnetV2/Repeat_2/block8_5/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_383/kernel': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_383/beta': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_383/moving_mean': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_383/moving_variance': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_384/kernel': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_384/beta': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_384/moving_mean': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_384/moving_variance': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_385/kernel': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_385/beta': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_385/moving_mean': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_385/moving_variance': 'InceptionResnetV2/Repeat_2/block8_5/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_5_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_5/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_5_conv/bias': 'InceptionResnetV2/Repeat_2/block8_5/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_386/kernel': 'InceptionResnetV2/Repeat_2/block8_6/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_386/beta': 'InceptionResnetV2/Repeat_2/block8_6/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_386/moving_mean': 'InceptionResnetV2/Repeat_2/block8_6/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_386/moving_variance': 'InceptionResnetV2/Repeat_2/block8_6/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_387/kernel': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_387/beta': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_387/moving_mean': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_387/moving_variance': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_388/kernel': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_388/beta': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_388/moving_mean': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_388/moving_variance': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_389/kernel': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_389/beta': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_389/moving_mean': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_389/moving_variance': 'InceptionResnetV2/Repeat_2/block8_6/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_6_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_6/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_6_conv/bias': 'InceptionResnetV2/Repeat_2/block8_6/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_390/kernel': 'InceptionResnetV2/Repeat_2/block8_7/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_390/beta': 'InceptionResnetV2/Repeat_2/block8_7/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_390/moving_mean': 'InceptionResnetV2/Repeat_2/block8_7/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_390/moving_variance': 'InceptionResnetV2/Repeat_2/block8_7/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_391/kernel': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_391/beta': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_391/moving_mean': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_391/moving_variance': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_392/kernel': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_392/beta': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_392/moving_mean': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_392/moving_variance': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_393/kernel': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_393/beta': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_393/moving_mean': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_393/moving_variance': 'InceptionResnetV2/Repeat_2/block8_7/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_7_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_7/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_7_conv/bias': 'InceptionResnetV2/Repeat_2/block8_7/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_394/kernel': 'InceptionResnetV2/Repeat_2/block8_8/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_394/beta': 'InceptionResnetV2/Repeat_2/block8_8/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_394/moving_mean': 'InceptionResnetV2/Repeat_2/block8_8/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_394/moving_variance': 'InceptionResnetV2/Repeat_2/block8_8/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_395/kernel': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_395/beta': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_395/moving_mean': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_395/moving_variance': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_396/kernel': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_396/beta': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_396/moving_mean': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_396/moving_variance': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_397/kernel': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_397/beta': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_397/moving_mean': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_397/moving_variance': 'InceptionResnetV2/Repeat_2/block8_8/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_8_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_8/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_8_conv/bias': 'InceptionResnetV2/Repeat_2/block8_8/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_398/kernel': 'InceptionResnetV2/Repeat_2/block8_9/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_398/beta': 'InceptionResnetV2/Repeat_2/block8_9/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_398/moving_mean': 'InceptionResnetV2/Repeat_2/block8_9/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_398/moving_variance': 'InceptionResnetV2/Repeat_2/block8_9/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_399/kernel': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_399/beta': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_399/moving_mean': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_399/moving_variance': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_400/kernel': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_400/beta': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_400/moving_mean': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_400/moving_variance': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_401/kernel': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_401/beta': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_401/moving_mean': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_401/moving_variance': 'InceptionResnetV2/Repeat_2/block8_9/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_9_conv/kernel': 'InceptionResnetV2/Repeat_2/block8_9/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_9_conv/bias': 'InceptionResnetV2/Repeat_2/block8_9/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_402/kernel': 'InceptionResnetV2/Block8/Branch_0/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_402/beta': 'InceptionResnetV2/Block8/Branch_0/Conv2d_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_402/moving_mean': 'InceptionResnetV2/Block8/Branch_0/Conv2d_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_402/moving_variance': 'InceptionResnetV2/Block8/Branch_0/Conv2d_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_403/kernel': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0a_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_403/beta': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0a_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_403/moving_mean': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_403/moving_variance': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0a_1x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_404/kernel': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0b_1x3/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_404/beta': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0b_1x3/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_404/moving_mean': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_404/moving_variance': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0b_1x3/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/conv2d_405/kernel': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0c_3x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_405/beta': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0c_3x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_405/moving_mean': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/freezable_batch_norm_405/moving_variance': 'InceptionResnetV2/Block8/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_10_conv/kernel': 'InceptionResnetV2/Block8/Conv2d_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/block8_10_conv/bias': 'InceptionResnetV2/Block8/Conv2d_1x1/biases',
'SecondStageFeatureExtractor/InceptionResnetV2/conv_7b/kernel': 'InceptionResnetV2/Conv2d_7b_1x1/weights',
'SecondStageFeatureExtractor/InceptionResnetV2/conv_7b_bn/beta': 'InceptionResnetV2/Conv2d_7b_1x1/BatchNorm/beta',
'SecondStageFeatureExtractor/InceptionResnetV2/conv_7b_bn/moving_mean': 'InceptionResnetV2/Conv2d_7b_1x1/BatchNorm/moving_mean',
'SecondStageFeatureExtractor/InceptionResnetV2/conv_7b_bn/moving_variance': 'InceptionResnetV2/Conv2d_7b_1x1/BatchNorm/moving_variance',
}
variables_to_restore = {}
for variable in tf.global_variables():
var_name = keras_to_slim_name_mapping.get(variable.op.name)
if var_name:
variables_to_restore[var_name] = variable
return variables_to_restore
research/object_detection/models/faster_rcnn_inception_resnet_v2_keras_feature_extractor_test.py 0 → 100644
View file @ 80444539
# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for models.faster_rcnn_inception_resnet_v2_keras_feature_extractor."""
import tensorflow as tf
from object_detection.models import faster_rcnn_inception_resnet_v2_keras_feature_extractor as frcnn_inc_res
class FasterRcnnInceptionResnetV2KerasFeatureExtractorTest(tf.test.TestCase):
def _build_feature_extractor(self, first_stage_features_stride):
return frcnn_inc_res.FasterRCNNInceptionResnetV2KerasFeatureExtractor(
is_training=False,
first_stage_features_stride=first_stage_features_stride,
batch_norm_trainable=False,
weight_decay=0.0)
def test_extract_proposal_features_returns_expected_size(self):
feature_extractor = self._build_feature_extractor(
first_stage_features_stride=16)
preprocessed_inputs = tf.random_uniform(
[1, 299, 299, 3], maxval=255, dtype=tf.float32)
rpn_feature_map = feature_extractor.get_proposal_feature_extractor_model(
name='TestScope')(preprocessed_inputs)
features_shape = tf.shape(rpn_feature_map)
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
features_shape_out = sess.run(features_shape)
self.assertAllEqual(features_shape_out, [1, 19, 19, 1088])
def test_extract_proposal_features_stride_eight(self):
feature_extractor = self._build_feature_extractor(
first_stage_features_stride=8)
preprocessed_inputs = tf.random_uniform(
[1, 224, 224, 3], maxval=255, dtype=tf.float32)
rpn_feature_map = feature_extractor.get_proposal_feature_extractor_model(
name='TestScope')(preprocessed_inputs)
features_shape = tf.shape(rpn_feature_map)
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
features_shape_out = sess.run(features_shape)
self.assertAllEqual(features_shape_out, [1, 28, 28, 1088])
def test_extract_proposal_features_half_size_input(self):
feature_extractor = self._build_feature_extractor(
first_stage_features_stride=16)
preprocessed_inputs = tf.random_uniform(
[1, 112, 112, 3], maxval=255, dtype=tf.float32)
rpn_feature_map = feature_extractor.get_proposal_feature_extractor_model(
name='TestScope')(preprocessed_inputs)
features_shape = tf.shape(rpn_feature_map)
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
features_shape_out = sess.run(features_shape)
self.assertAllEqual(features_shape_out, [1, 7, 7, 1088])
def test_extract_proposal_features_dies_on_invalid_stride(self):
with self.assertRaises(ValueError):
self._build_feature_extractor(first_stage_features_stride=99)
def test_extract_proposal_features_dies_with_incorrect_rank_inputs(self):
feature_extractor = self._build_feature_extractor(
first_stage_features_stride=16)
preprocessed_inputs = tf.random_uniform(
[224, 224, 3], maxval=255, dtype=tf.float32)
with self.assertRaises(ValueError):
feature_extractor.get_proposal_feature_extractor_model(
name='TestScope')(preprocessed_inputs)
def test_extract_box_classifier_features_returns_expected_size(self):
feature_extractor = self._build_feature_extractor(
first_stage_features_stride=16)
proposal_feature_maps = tf.random_uniform(
[2, 17, 17, 1088], maxval=255, dtype=tf.float32)
model = feature_extractor.get_box_classifier_feature_extractor_model(
name='TestScope')
proposal_classifier_features = (
model(proposal_feature_maps))
features_shape = tf.shape(proposal_classifier_features)
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
features_shape_out = sess.run(features_shape)
self.assertAllEqual(features_shape_out, [2, 8, 8, 1536])
if __name__ == '__main__':
tf.test.main()
research/object_detection/models/feature_map_generators.py
View file @ 80444539
......@@ -51,6 +51,60 @@ def get_depth_fn(depth_multiplier, min_depth):
return multiply_depth
def create_conv_block(
use_depthwise, kernel_size, padding, stride, layer_name, conv_hyperparams,
is_training, freeze_batchnorm, depth):
"""Create Keras layers for depthwise & non-depthwise convolutions.
Args:
use_depthwise: Whether to use depthwise separable conv instead of regular
conv.
kernel_size: A list of length 2: [kernel_height, kernel_width] of the
filters. Can be an int if both values are the same.
padding: One of 'VALID' or 'SAME'.
stride: A list of length 2: [stride_height, stride_width], specifying the
convolution stride. Can be an int if both strides are the same.
layer_name: String. The name of the layer.
conv_hyperparams: A `hyperparams_builder.KerasLayerHyperparams` object
containing hyperparameters for convolution ops.
is_training: Indicates whether the feature generator is in training mode.
freeze_batchnorm: Bool. Whether to freeze batch norm parameters during
training or not. When training with a small batch size (e.g. 1), it is
desirable to freeze batch norm update and use pretrained batch norm
params.
depth: Depth of output feature maps.
Returns:
A list of conv layers.
"""
layers = []
if use_depthwise:
layers.append(tf.keras.layers.SeparableConv2D(
depth,
[kernel_size, kernel_size],
depth_multiplier=1,
padding=padding,
strides=stride,
name=layer_name + '_depthwise_conv',
**conv_hyperparams.params()))
else:
layers.append(tf.keras.layers.Conv2D(
depth,
[kernel_size, kernel_size],
padding=padding,
strides=stride,
name=layer_name + '_conv',
**conv_hyperparams.params()))
layers.append(
conv_hyperparams.build_batch_norm(
training=(is_training and not freeze_batchnorm),
name=layer_name + '_batchnorm'))
layers.append(
conv_hyperparams.build_activation_layer(
name=layer_name))
return layers
class KerasMultiResolutionFeatureMaps(tf.keras.Model):
"""Generates multi resolution feature maps from input image features.
......@@ -419,6 +473,180 @@ def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
[(x, y) for (x, y) in zip(feature_map_keys, feature_maps)])
class KerasFpnTopDownFeatureMaps(tf.keras.Model):
"""Generates Keras based `top-down` feature maps for Feature Pyramid Networks.
See https://arxiv.org/abs/1612.03144 for details.
"""
def __init__(self,
num_levels,
depth,
is_training,
conv_hyperparams,
freeze_batchnorm,
use_depthwise=False,
use_explicit_padding=False,
use_bounded_activations=False,
use_native_resize_op=False,
scope=None,
name=None):
"""Constructor.
Args:
num_levels: the number of image features.
depth: depth of output feature maps.
is_training: Indicates whether the feature generator is in training mode.
conv_hyperparams: A `hyperparams_builder.KerasLayerHyperparams` object
containing hyperparameters for convolution ops.
freeze_batchnorm: Bool. Whether to freeze batch norm parameters during
training or not. When training with a small batch size (e.g. 1), it is
desirable to freeze batch norm update and use pretrained batch norm
params.
use_depthwise: whether to use depthwise separable conv instead of regular
conv.
use_explicit_padding: whether to use explicit padding.
use_bounded_activations: Whether or not to clip activations to range
[-ACTIVATION_BOUND, ACTIVATION_BOUND]. Bounded activations better lend
themselves to quantized inference.
use_native_resize_op: If True, uses tf.image.resize_nearest_neighbor op
for the upsampling process instead of reshape and broadcasting
implementation.
scope: A scope name to wrap this op under.
name: A string name scope to assign to the model. If 'None', Keras
will auto-generate one from the class name.
"""
super(KerasFpnTopDownFeatureMaps, self).__init__(name=name)
self.scope = scope if scope else 'top_down'
self.top_layers = []
self.residual_blocks = []
self.top_down_blocks = []
self.reshape_blocks = []
self.conv_layers = []
padding = 'VALID' if use_explicit_padding else 'SAME'
stride = 1
kernel_size = 3
def clip_by_value(features):
return tf.clip_by_value(features, -ACTIVATION_BOUND, ACTIVATION_BOUND)
# top layers
self.top_layers.append(tf.keras.layers.Conv2D(
depth, [1, 1], strides=stride, padding=padding,
name='projection_%d' % num_levels,
**conv_hyperparams.params(use_bias=True)))
if use_bounded_activations:
self.top_layers.append(tf.keras.layers.Lambda(
clip_by_value, name='clip_by_value'))
for level in reversed(range(num_levels - 1)):
# to generate residual from image features
residual_net = []
# to preprocess top_down (the image feature map from last layer)
top_down_net = []
# to reshape top_down according to residual if necessary
reshaped_residual = []
# to apply convolution layers to feature map
conv_net = []
# residual block
residual_net.append(tf.keras.layers.Conv2D(
depth, [1, 1], padding=padding, strides=1,
name='projection_%d' % (level + 1),
**conv_hyperparams.params(use_bias=True)))
if use_bounded_activations:
residual_net.append(tf.keras.layers.Lambda(
clip_by_value, name='clip_by_value'))
# top-down block
# TODO (b/128922690): clean-up of ops.nearest_neighbor_upsampling
if use_native_resize_op:
def resize_nearest_neighbor(image):
image_shape = image.shape.as_list()
return tf.image.resize_nearest_neighbor(
image, [image_shape[1] * 2, image_shape[2] * 2])
top_down_net.append(tf.keras.layers.Lambda(
resize_nearest_neighbor, name='nearest_neighbor_upsampling'))
else:
def nearest_neighbor_upsampling(image):
return ops.nearest_neighbor_upsampling(image, scale=2)
top_down_net.append(tf.keras.layers.Lambda(
nearest_neighbor_upsampling, name='nearest_neighbor_upsampling'))
# reshape block
if use_explicit_padding:
def reshape(inputs):
residual_shape = tf.shape(inputs[0])
return inputs[1][:, :residual_shape[1], :residual_shape[2], :]
reshaped_residual.append(
tf.keras.layers.Lambda(reshape, name='reshape'))
# down layers
if use_bounded_activations:
conv_net.append(tf.keras.layers.Lambda(
clip_by_value, name='clip_by_value'))
if use_explicit_padding:
def fixed_padding(features, kernel_size=kernel_size):
return ops.fixed_padding(features, kernel_size)
conv_net.append(tf.keras.layers.Lambda(
fixed_padding, name='fixed_padding'))
layer_name = 'smoothing_%d' % (level + 1)
conv_block = create_conv_block(
use_depthwise, kernel_size, padding, stride, layer_name,
conv_hyperparams, is_training, freeze_batchnorm, depth)
conv_net.extend(conv_block)
self.residual_blocks.append(residual_net)
self.top_down_blocks.append(top_down_net)
self.reshape_blocks.append(reshaped_residual)
self.conv_layers.append(conv_net)
def call(self, image_features):
"""Generate the multi-resolution feature maps.
Executed when calling the `.__call__` method on input.
Args:
image_features: list of tuples of (tensor_name, image_feature_tensor).
Spatial resolutions of succesive tensors must reduce exactly by a factor
of 2.
Returns:
feature_maps: an OrderedDict mapping keys (feature map names) to
tensors where each tensor has shape [batch, height_i, width_i, depth_i].
"""
output_feature_maps_list = []
output_feature_map_keys = []
with tf.name_scope(self.scope):
top_down = image_features[-1][1]
for layer in self.top_layers:
top_down = layer(top_down)
output_feature_maps_list.append(top_down)
output_feature_map_keys.append('top_down_%s' % image_features[-1][0])
num_levels = len(image_features)
for index, level in enumerate(reversed(range(num_levels - 1))):
residual = image_features[level][1]
top_down = output_feature_maps_list[-1]
for layer in self.residual_blocks[index]:
residual = layer(residual)
for layer in self.top_down_blocks[index]:
top_down = layer(top_down)
for layer in self.reshape_blocks[index]:
top_down = layer([residual, top_down])
top_down += residual
for layer in self.conv_layers[index]:
top_down = layer(top_down)
output_feature_maps_list.append(top_down)
output_feature_map_keys.append('top_down_%s' % image_features[level][0])
return collections.OrderedDict(reversed(
list(zip(output_feature_map_keys, output_feature_maps_list))))
def fpn_top_down_feature_maps(image_features,
depth,
use_depthwise=False,
......
research/object_detection/models/feature_map_generators_test.py
View file @ 80444539
......@@ -403,21 +403,101 @@ class MultiResolutionFeatureMapGeneratorTest(tf.test.TestCase):
self.assertSetEqual(expected_slim_variables, actual_variable_set)
@parameterized.parameters({'use_native_resize_op': True},
{'use_native_resize_op': False})
@parameterized.parameters({'use_native_resize_op': True, 'use_keras': False},
{'use_native_resize_op': False, 'use_keras': False},
{'use_native_resize_op': True, 'use_keras': True},
{'use_native_resize_op': False, 'use_keras': True})
class FPNFeatureMapGeneratorTest(tf.test.TestCase, parameterized.TestCase):
def test_get_expected_feature_map_shapes(self, use_native_resize_op):
def _build_conv_hyperparams(self):
conv_hyperparams = hyperparams_pb2.Hyperparams()
conv_hyperparams_text_proto = """
regularizer {
l2_regularizer {
}
}
initializer {
truncated_normal_initializer {
}
}
"""
text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams)
return hyperparams_builder.KerasLayerHyperparams(conv_hyperparams)
def _build_feature_map_generator(
self, image_features, depth, use_keras, use_bounded_activations=False,
use_native_resize_op=False, use_explicit_padding=False,
use_depthwise=False):
if use_keras:
return feature_map_generators.KerasFpnTopDownFeatureMaps(
num_levels=len(image_features),
depth=depth,
is_training=True,
conv_hyperparams=self._build_conv_hyperparams(),
freeze_batchnorm=False,
use_depthwise=use_depthwise,
use_explicit_padding=use_explicit_padding,
use_bounded_activations=use_bounded_activations,
use_native_resize_op=use_native_resize_op,
scope=None,
name='FeatureMaps',
)
else:
def feature_map_generator(image_features):
return feature_map_generators.fpn_top_down_feature_maps(
image_features=image_features,
depth=depth,
use_depthwise=use_depthwise,
use_explicit_padding=use_explicit_padding,
use_bounded_activations=use_bounded_activations,
use_native_resize_op=use_native_resize_op)
return feature_map_generator
def test_get_expected_feature_map_shapes(
self, use_native_resize_op, use_keras):
image_features = [
('block2', tf.random_uniform([4, 8, 8, 256], dtype=tf.float32)),
('block3', tf.random_uniform([4, 4, 4, 256], dtype=tf.float32)),
('block4', tf.random_uniform([4, 2, 2, 256], dtype=tf.float32)),
('block5', tf.random_uniform([4, 1, 1, 256], dtype=tf.float32))
]
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_native_resize_op=use_native_resize_op)
feature_maps = feature_map_generator(image_features)
expected_feature_map_shapes = {
'top_down_block2': (4, 8, 8, 128),
'top_down_block3': (4, 4, 4, 128),
'top_down_block4': (4, 2, 2, 128),
'top_down_block5': (4, 1, 1, 128)
}
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
out_feature_maps = sess.run(feature_maps)
out_feature_map_shapes = {key: value.shape
for key, value in out_feature_maps.items()}
self.assertDictEqual(out_feature_map_shapes, expected_feature_map_shapes)
def test_get_expected_feature_map_shapes_with_explicit_padding(
self, use_native_resize_op, use_keras):
image_features = [
('block2', tf.random_uniform([4, 8, 8, 256], dtype=tf.float32)),
('block3', tf.random_uniform([4, 4, 4, 256], dtype=tf.float32)),
('block4', tf.random_uniform([4, 2, 2, 256], dtype=tf.float32)),
('block5', tf.random_uniform([4, 1, 1, 256], dtype=tf.float32))
]
feature_maps = feature_map_generators.fpn_top_down_feature_maps(
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_explicit_padding=True,
use_native_resize_op=use_native_resize_op)
feature_maps = feature_map_generator(image_features)
expected_feature_map_shapes = {
'top_down_block2': (4, 8, 8, 128),
......@@ -434,7 +514,8 @@ class FPNFeatureMapGeneratorTest(tf.test.TestCase, parameterized.TestCase):
for key, value in out_feature_maps.items()}
self.assertDictEqual(out_feature_map_shapes, expected_feature_map_shapes)
def test_use_bounded_activations_add_operations(self, use_native_resize_op):
def test_use_bounded_activations_add_operations(
self, use_native_resize_op, use_keras):
tf_graph = tf.Graph()
with tf_graph.as_default():
image_features = [('block2',
......@@ -445,22 +526,37 @@ class FPNFeatureMapGeneratorTest(tf.test.TestCase, parameterized.TestCase):
tf.random_uniform([4, 2, 2, 256], dtype=tf.float32)),
('block5',
tf.random_uniform([4, 1, 1, 256], dtype=tf.float32))]
feature_map_generators.fpn_top_down_feature_maps(
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_bounded_activations=True,
use_native_resize_op=use_native_resize_op)
feature_map_generator(image_features)
if use_keras:
expected_added_operations = dict.fromkeys([
'FeatureMaps/top_down/clip_by_value/clip_by_value',
'FeatureMaps/top_down/clip_by_value_1/clip_by_value',
'FeatureMaps/top_down/clip_by_value_2/clip_by_value',
'FeatureMaps/top_down/clip_by_value_3/clip_by_value',
'FeatureMaps/top_down/clip_by_value_4/clip_by_value',
'FeatureMaps/top_down/clip_by_value_5/clip_by_value',
'FeatureMaps/top_down/clip_by_value_6/clip_by_value',
])
else:
expected_added_operations = dict.fromkeys([
'top_down/clip_by_value', 'top_down/clip_by_value_1',
'top_down/clip_by_value_2', 'top_down/clip_by_value_3',
'top_down/clip_by_value_4', 'top_down/clip_by_value_5',
'top_down/clip_by_value_6'
])
expected_added_operations = dict.fromkeys([
'top_down/clip_by_value', 'top_down/clip_by_value_1',
'top_down/clip_by_value_2', 'top_down/clip_by_value_3',
'top_down/clip_by_value_4', 'top_down/clip_by_value_5',
'top_down/clip_by_value_6'
])
op_names = {op.name: None for op in tf_graph.get_operations()}
self.assertDictContainsSubset(expected_added_operations, op_names)
def test_use_bounded_activations_clip_value(self, use_native_resize_op):
def test_use_bounded_activations_clip_value(
self, use_native_resize_op, use_keras):
tf_graph = tf.Graph()
with tf_graph.as_default():
image_features = [
......@@ -469,18 +565,31 @@ class FPNFeatureMapGeneratorTest(tf.test.TestCase, parameterized.TestCase):
('block4', 255 * tf.ones([4, 2, 2, 256], dtype=tf.float32)),
('block5', 255 * tf.ones([4, 1, 1, 256], dtype=tf.float32))
]
feature_map_generators.fpn_top_down_feature_maps(
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_bounded_activations=True,
use_native_resize_op=use_native_resize_op)
feature_map_generator(image_features)
expected_clip_by_value_ops = [
'top_down/clip_by_value', 'top_down/clip_by_value_1',
'top_down/clip_by_value_2', 'top_down/clip_by_value_3',
'top_down/clip_by_value_4', 'top_down/clip_by_value_5',
'top_down/clip_by_value_6'
]
if use_keras:
expected_clip_by_value_ops = dict.fromkeys([
'FeatureMaps/top_down/clip_by_value/clip_by_value',
'FeatureMaps/top_down/clip_by_value_1/clip_by_value',
'FeatureMaps/top_down/clip_by_value_2/clip_by_value',
'FeatureMaps/top_down/clip_by_value_3/clip_by_value',
'FeatureMaps/top_down/clip_by_value_4/clip_by_value',
'FeatureMaps/top_down/clip_by_value_5/clip_by_value',
'FeatureMaps/top_down/clip_by_value_6/clip_by_value',
])
else:
expected_clip_by_value_ops = [
'top_down/clip_by_value', 'top_down/clip_by_value_1',
'top_down/clip_by_value_2', 'top_down/clip_by_value_3',
'top_down/clip_by_value_4', 'top_down/clip_by_value_5',
'top_down/clip_by_value_6'
]
# Gathers activation tensors before and after clip_by_value operations.
activations = {}
......@@ -522,18 +631,20 @@ class FPNFeatureMapGeneratorTest(tf.test.TestCase, parameterized.TestCase):
self.assertLessEqual(after_clipping_upper_bound, expected_upper_bound)
def test_get_expected_feature_map_shapes_with_depthwise(
self, use_native_resize_op):
self, use_native_resize_op, use_keras):
image_features = [
('block2', tf.random_uniform([4, 8, 8, 256], dtype=tf.float32)),
('block3', tf.random_uniform([4, 4, 4, 256], dtype=tf.float32)),
('block4', tf.random_uniform([4, 2, 2, 256], dtype=tf.float32)),
('block5', tf.random_uniform([4, 1, 1, 256], dtype=tf.float32))
]
feature_maps = feature_map_generators.fpn_top_down_feature_maps(
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_depthwise=True,
use_native_resize_op=use_native_resize_op)
feature_maps = feature_map_generator(image_features)
expected_feature_map_shapes = {
'top_down_block2': (4, 8, 8, 128),
......@@ -550,6 +661,131 @@ class FPNFeatureMapGeneratorTest(tf.test.TestCase, parameterized.TestCase):
for key, value in out_feature_maps.items()}
self.assertDictEqual(out_feature_map_shapes, expected_feature_map_shapes)
def test_get_expected_variable_names(
self, use_native_resize_op, use_keras):
image_features = [
('block2', tf.random_uniform([4, 8, 8, 256], dtype=tf.float32)),
('block3', tf.random_uniform([4, 4, 4, 256], dtype=tf.float32)),
('block4', tf.random_uniform([4, 2, 2, 256], dtype=tf.float32)),
('block5', tf.random_uniform([4, 1, 1, 256], dtype=tf.float32))
]
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_native_resize_op=use_native_resize_op)
feature_maps = feature_map_generator(image_features)
expected_slim_variables = set([
'projection_1/weights',
'projection_1/biases',
'projection_2/weights',
'projection_2/biases',
'projection_3/weights',
'projection_3/biases',
'projection_4/weights',
'projection_4/biases',
'smoothing_1/weights',
'smoothing_1/biases',
'smoothing_2/weights',
'smoothing_2/biases',
'smoothing_3/weights',
'smoothing_3/biases',
])
expected_keras_variables = set([
'FeatureMaps/top_down/projection_1/kernel',
'FeatureMaps/top_down/projection_1/bias',
'FeatureMaps/top_down/projection_2/kernel',
'FeatureMaps/top_down/projection_2/bias',
'FeatureMaps/top_down/projection_3/kernel',
'FeatureMaps/top_down/projection_3/bias',
'FeatureMaps/top_down/projection_4/kernel',
'FeatureMaps/top_down/projection_4/bias',
'FeatureMaps/top_down/smoothing_1_conv/kernel',
'FeatureMaps/top_down/smoothing_1_conv/bias',
'FeatureMaps/top_down/smoothing_2_conv/kernel',
'FeatureMaps/top_down/smoothing_2_conv/bias',
'FeatureMaps/top_down/smoothing_3_conv/kernel',
'FeatureMaps/top_down/smoothing_3_conv/bias'
])
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
sess.run(feature_maps)
actual_variable_set = set(
[var.op.name for var in tf.trainable_variables()])
if use_keras:
self.assertSetEqual(expected_keras_variables, actual_variable_set)
else:
self.assertSetEqual(expected_slim_variables, actual_variable_set)
def test_get_expected_variable_names_with_depthwise(
self, use_native_resize_op, use_keras):
image_features = [
('block2', tf.random_uniform([4, 8, 8, 256], dtype=tf.float32)),
('block3', tf.random_uniform([4, 4, 4, 256], dtype=tf.float32)),
('block4', tf.random_uniform([4, 2, 2, 256], dtype=tf.float32)),
('block5', tf.random_uniform([4, 1, 1, 256], dtype=tf.float32))
]
feature_map_generator = self._build_feature_map_generator(
image_features=image_features,
depth=128,
use_keras=use_keras,
use_depthwise=True,
use_native_resize_op=use_native_resize_op)
feature_maps = feature_map_generator(image_features)
expected_slim_variables = set([
'projection_1/weights',
'projection_1/biases',
'projection_2/weights',
'projection_2/biases',
'projection_3/weights',
'projection_3/biases',
'projection_4/weights',
'projection_4/biases',
'smoothing_1/depthwise_weights',
'smoothing_1/pointwise_weights',
'smoothing_1/biases',
'smoothing_2/depthwise_weights',
'smoothing_2/pointwise_weights',
'smoothing_2/biases',
'smoothing_3/depthwise_weights',
'smoothing_3/pointwise_weights',
'smoothing_3/biases',
])
expected_keras_variables = set([
'FeatureMaps/top_down/projection_1/kernel',
'FeatureMaps/top_down/projection_1/bias',
'FeatureMaps/top_down/projection_2/kernel',
'FeatureMaps/top_down/projection_2/bias',
'FeatureMaps/top_down/projection_3/kernel',
'FeatureMaps/top_down/projection_3/bias',
'FeatureMaps/top_down/projection_4/kernel',
'FeatureMaps/top_down/projection_4/bias',
'FeatureMaps/top_down/smoothing_1_depthwise_conv/depthwise_kernel',
'FeatureMaps/top_down/smoothing_1_depthwise_conv/pointwise_kernel',
'FeatureMaps/top_down/smoothing_1_depthwise_conv/bias',
'FeatureMaps/top_down/smoothing_2_depthwise_conv/depthwise_kernel',
'FeatureMaps/top_down/smoothing_2_depthwise_conv/pointwise_kernel',
'FeatureMaps/top_down/smoothing_2_depthwise_conv/bias',
'FeatureMaps/top_down/smoothing_3_depthwise_conv/depthwise_kernel',
'FeatureMaps/top_down/smoothing_3_depthwise_conv/pointwise_kernel',
'FeatureMaps/top_down/smoothing_3_depthwise_conv/bias'
])
init_op = tf.global_variables_initializer()
with self.test_session() as sess:
sess.run(init_op)
sess.run(feature_maps)
actual_variable_set = set(
[var.op.name for var in tf.trainable_variables()])
if use_keras:
self.assertSetEqual(expected_keras_variables, actual_variable_set)
else:
self.assertSetEqual(expected_slim_variables, actual_variable_set)
class GetDepthFunctionTest(tf.test.TestCase):
......
research/object_detection/models/keras_models/base_models/original_mobilenet_v2.py 0 → 100644
View file @ 80444539
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""MobileNet v2 models for Keras.
MobileNetV2 is a general architecture and can be used for multiple use cases.
Depending on the use case, it can use different input layer size and
different width factors. This allows different width models to reduce
the number of multiply-adds and thereby
reduce inference cost on mobile devices.
MobileNetV2 is very similar to the original MobileNet,
except that it uses inverted residual blocks with
bottlenecking features. It has a drastically lower
parameter count than the original MobileNet.
MobileNets support any input size greater
than 32 x 32, with larger image sizes
offering better performance.
The number of parameters and number of multiply-adds
can be modified by using the `alpha` parameter,
which increases/decreases the number of filters in each layer.
By altering the image size and `alpha` parameter,
all 22 models from the paper can be built, with ImageNet weights provided.
The paper demonstrates the performance of MobileNets using `alpha` values of
1.0 (also called 100 % MobileNet), 0.35, 0.5, 0.75, 1.0, 1.3, and 1.4
For each of these `alpha` values, weights for 5 different input image sizes
are provided (224, 192, 160, 128, and 96).
The following table describes the performance of
MobileNet on various input sizes:
------------------------------------------------------------------------
MACs stands for Multiply Adds
Classification Checkpoint| MACs (M) | Parameters (M)| Top 1 Acc| Top 5 Acc
--------------------------|------------|---------------|---------|----|-------
| [mobilenet_v2_1.4_224] | 582 | 6.06 | 75.0 | 92.5 |
| [mobilenet_v2_1.3_224] | 509 | 5.34 | 74.4 | 92.1 |
| [mobilenet_v2_1.0_224] | 300 | 3.47 | 71.8 | 91.0 |
| [mobilenet_v2_1.0_192] | 221 | 3.47 | 70.7 | 90.1 |
| [mobilenet_v2_1.0_160] | 154 | 3.47 | 68.8 | 89.0 |
| [mobilenet_v2_1.0_128] | 99 | 3.47 | 65.3 | 86.9 |
| [mobilenet_v2_1.0_96] | 56 | 3.47 | 60.3 | 83.2 |
| [mobilenet_v2_0.75_224] | 209 | 2.61 | 69.8 | 89.6 |
| [mobilenet_v2_0.75_192] | 153 | 2.61 | 68.7 | 88.9 |
| [mobilenet_v2_0.75_160] | 107 | 2.61 | 66.4 | 87.3 |
| [mobilenet_v2_0.75_128] | 69 | 2.61 | 63.2 | 85.3 |
| [mobilenet_v2_0.75_96] | 39 | 2.61 | 58.8 | 81.6 |
| [mobilenet_v2_0.5_224] | 97 | 1.95 | 65.4 | 86.4 |
| [mobilenet_v2_0.5_192] | 71 | 1.95 | 63.9 | 85.4 |
| [mobilenet_v2_0.5_160] | 50 | 1.95 | 61.0 | 83.2 |
| [mobilenet_v2_0.5_128] | 32 | 1.95 | 57.7 | 80.8 |
| [mobilenet_v2_0.5_96] | 18 | 1.95 | 51.2 | 75.8 |
| [mobilenet_v2_0.35_224] | 59 | 1.66 | 60.3 | 82.9 |
| [mobilenet_v2_0.35_192] | 43 | 1.66 | 58.2 | 81.2 |
| [mobilenet_v2_0.35_160] | 30 | 1.66 | 55.7 | 79.1 |
| [mobilenet_v2_0.35_128] | 20 | 1.66 | 50.8 | 75.0 |
| [mobilenet_v2_0.35_96] | 11 | 1.66 | 45.5 | 70.4 |
The weights for all 16 models are obtained and translated from the Tensorflow
checkpoints from TensorFlow checkpoints found at
https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/README.md
# Reference
This file contains building code for MobileNetV2, based on
[MobileNetV2: Inverted Residuals and Linear Bottlenecks]
(https://arxiv.org/abs/1801.04381)
Tests comparing this model to the existing Tensorflow model can be
found at
[mobilenet_v2_keras](https://github.com/JonathanCMitchell/mobilenet_v2_keras)
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import warnings
import numpy as np
import tensorflow as tf
Model = tf.keras.Model
Input = tf.keras.layers.Input
Activation = tf.keras.layers.Activation
BatchNormalization = tf.keras.layers.BatchNormalization
Conv2D = tf.keras.layers.Conv2D
DepthwiseConv2D = tf.keras.layers.DepthwiseConv2D
GlobalAveragePooling2D = tf.keras.layers.GlobalAveragePooling2D
Add = tf.keras.layers.Add
Dense = tf.keras.layers.Dense
K = tf.keras.Backend
def relu6(x):
return K.relu(x, max_value=6)
def _obtain_input_shape(
input_shape,
default_size,
min_size,
data_format,
require_flatten):
"""Internal utility to compute/validate an ImageNet model's input shape.
Arguments:
input_shape: either None (will return the default network input shape),
or a user-provided shape to be validated.
default_size: default input width/height for the model.
min_size: minimum input width/height accepted by the model.
data_format: image data format to use.
require_flatten: whether the model is expected to
be linked to a classifier via a Flatten layer.
Returns:
An integer shape tuple (may include None entries).
Raises:
ValueError: in case of invalid argument values.
"""
if input_shape and len(input_shape) == 3:
if data_format == 'channels_first':
if input_shape[0] not in {1, 3}:
warnings.warn(
'This model usually expects 1 or 3 input channels. '
'However, it was passed an input_shape with ' +
str(input_shape[0]) + ' input channels.')
default_shape = (input_shape[0], default_size, default_size)
else:
if input_shape[-1] not in {1, 3}:
warnings.warn(
'This model usually expects 1 or 3 input channels. '
'However, it was passed an input_shape with ' +
str(input_shape[-1]) + ' input channels.')
default_shape = (default_size, default_size, input_shape[-1])
else:
if data_format == 'channels_first':
default_shape = (3, default_size, default_size)
else:
default_shape = (default_size, default_size, 3)
if input_shape:
if data_format == 'channels_first':
if input_shape is not None:
if len(input_shape) != 3:
raise ValueError(
'`input_shape` must be a tuple of three integers.')
if ((input_shape[1] is not None and input_shape[1] < min_size) or
(input_shape[2] is not None and input_shape[2] < min_size)):
raise ValueError('Input size must be at least ' +
str(min_size) + 'x' + str(min_size) +
'; got `input_shape=' +
str(input_shape) + '`')
else:
if input_shape is not None:
if len(input_shape) != 3:
raise ValueError(
'`input_shape` must be a tuple of three integers.')
if ((input_shape[0] is not None and input_shape[0] < min_size) or
(input_shape[1] is not None and input_shape[1] < min_size)):
raise ValueError('Input size must be at least ' +
str(min_size) + 'x' + str(min_size) +
'; got `input_shape=' +
str(input_shape) + '`')
else:
if require_flatten:
input_shape = default_shape
else:
if data_format == 'channels_first':
input_shape = (3, None, None)
else:
input_shape = (None, None, 3)
if require_flatten:
if None in input_shape:
raise ValueError('If `include_top` is True, '
'you should specify a static `input_shape`. '
'Got `input_shape=' + str(input_shape) + '`')
return input_shape
def preprocess_input(x):
"""Preprocesses a numpy array encoding a batch of images.
This function applies the "Inception" preprocessing which converts
the RGB values from [0, 255] to [-1, 1]. Note that this preprocessing
function is different from `imagenet_utils.preprocess_input()`.
Arguments:
x: a 4D numpy array consists of RGB values within [0, 255].
Returns:
Preprocessed array.
"""
x /= 128.
x -= 1.
return x.astype(np.float32)
# This function is taken from the original tf repo.
# It ensures that all layers have a channel number that is divisible by 8
# It can be seen here:
# https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
def _make_divisible(v, divisor, min_value=None):
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < 0.9 * v:
new_v += divisor
return new_v
def mobilenet_v2(input_shape=None,
alpha=1.0,
include_top=True,
classes=1000):
"""Instantiates the MobileNetV2 architecture.
To load a MobileNetV2 model via `load_model`, import the custom
objects `relu6` and pass them to the `custom_objects` parameter.
E.g.
model = load_model('mobilenet.h5', custom_objects={
'relu6': mobilenet.relu6})
Arguments:
input_shape: optional shape tuple, to be specified if you would
like to use a model with an input img resolution that is not
(224, 224, 3).
It should have exactly 3 inputs channels (224, 224, 3).
You can also omit this option if you would like
to infer input_shape from an input_tensor.
If you choose to include both input_tensor and input_shape then
input_shape will be used if they match, if the shapes
do not match then we will throw an error.
E.g. `(160, 160, 3)` would be one valid value.
alpha: controls the width of the network. This is known as the
width multiplier in the MobileNetV2 paper.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
include_top: whether to include the fully-connected
layer at the top of the network.
classes: optional number of classes to classify images
into, only to be specified if `include_top` is True, and
if no `weights` argument is specified.
Returns:
A Keras model instance.
Raises:
ValueError: in case of invalid argument for `weights`,
or invalid input shape or invalid depth_multiplier, alpha,
rows when weights='imagenet'
"""
# Determine proper input shape and default size.
# If input_shape is None and no input_tensor
if input_shape is None:
default_size = 224
# If input_shape is not None, assume default size
else:
if K.image_data_format() == 'channels_first':
rows = input_shape[1]
cols = input_shape[2]
else:
rows = input_shape[0]
cols = input_shape[1]
if rows == cols and rows in [96, 128, 160, 192, 224]:
default_size = rows
else:
default_size = 224
input_shape = _obtain_input_shape(input_shape,
default_size=default_size,
min_size=32,
data_format=K.image_data_format(),
require_flatten=include_top)
if K.image_data_format() == 'channels_last':
row_axis, col_axis = (0, 1)
else:
row_axis, col_axis = (1, 2)
rows = input_shape[row_axis]
cols = input_shape[col_axis]
if K.image_data_format() != 'channels_last':
warnings.warn('The MobileNet family of models is only available '
'for the input data format "channels_last" '
'(width, height, channels). '
'However your settings specify the default '
'data format "channels_first" (channels, width, height).'
' You should set `image_data_format="channels_last"` '
'in your Keras config located at ~/.keras/keras.json. '
'The model being returned right now will expect inputs '
'to follow the "channels_last" data format.')
K.set_image_data_format('channels_last')
old_data_format = 'channels_first'
else:
old_data_format = None
img_input = Input(shape=input_shape)
first_block_filters = _make_divisible(32 * alpha, 8)
x = Conv2D(first_block_filters,
kernel_size=3,
strides=(2, 2), padding='same',
use_bias=False, name='Conv1')(img_input)
x = BatchNormalization(epsilon=1e-3, momentum=0.999, name='bn_Conv1')(x)
x = Activation(relu6, name='Conv1_relu')(x)
x = _first_inverted_res_block(x,
filters=16,
alpha=alpha,
stride=1,
block_id=0)
x = _inverted_res_block(x, filters=24, alpha=alpha, stride=2,
expansion=6, block_id=1)
x = _inverted_res_block(x, filters=24, alpha=alpha, stride=1,
expansion=6, block_id=2)
x = _inverted_res_block(x, filters=32, alpha=alpha, stride=2,
expansion=6, block_id=3)
x = _inverted_res_block(x, filters=32, alpha=alpha, stride=1,
expansion=6, block_id=4)
x = _inverted_res_block(x, filters=32, alpha=alpha, stride=1,
expansion=6, block_id=5)
x = _inverted_res_block(x, filters=64, alpha=alpha, stride=2,
expansion=6, block_id=6)
x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1,
expansion=6, block_id=7)
x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1,
expansion=6, block_id=8)
x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1,
expansion=6, block_id=9)
x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1,
expansion=6, block_id=10)
x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1,
expansion=6, block_id=11)
x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1,
expansion=6, block_id=12)
x = _inverted_res_block(x, filters=160, alpha=alpha, stride=2,
expansion=6, block_id=13)
x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1,
expansion=6, block_id=14)
x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1,
expansion=6, block_id=15)
x = _inverted_res_block(x, filters=320, alpha=alpha, stride=1,
expansion=6, block_id=16)
# no alpha applied to last conv as stated in the paper:
# if the width multiplier is greater than 1 we
# increase the number of output channels
if alpha > 1.0:
last_block_filters = _make_divisible(1280 * alpha, 8)
else:
last_block_filters = 1280
x = Conv2D(last_block_filters,
kernel_size=1,
use_bias=False,
name='Conv_1')(x)
x = BatchNormalization(epsilon=1e-3, momentum=0.999, name='Conv_1_bn')(x)
x = Activation(relu6, name='out_relu')(x)
if include_top:
x = GlobalAveragePooling2D()(x)
x = Dense(classes, activation='softmax',
use_bias=True, name='Logits')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
inputs = img_input
# Create model.
model = Model(inputs, x, name='mobilenetv2_%0.2f_%s' % (alpha, rows))
if old_data_format:
K.set_image_data_format(old_data_format)
return model
def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id):
"""Build an inverted res block."""
in_channels = int(inputs.shape[-1])
pointwise_conv_filters = int(filters * alpha)
pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
# Expand
x = Conv2D(expansion * in_channels, kernel_size=1, padding='same',
use_bias=False, activation=None,
name='mobl%d_conv_expand' % block_id)(inputs)
x = BatchNormalization(epsilon=1e-3, momentum=0.999,
name='bn%d_conv_bn_expand' %
block_id)(x)
x = Activation(relu6, name='conv_%d_relu' % block_id)(x)
# Depthwise
x = DepthwiseConv2D(kernel_size=3, strides=stride, activation=None,
use_bias=False, padding='same',
name='mobl%d_conv_depthwise' % block_id)(x)
x = BatchNormalization(epsilon=1e-3, momentum=0.999,
name='bn%d_conv_depthwise' % block_id)(x)
x = Activation(relu6, name='conv_dw_%d_relu' % block_id)(x)
# Project
x = Conv2D(pointwise_filters,
kernel_size=1, padding='same', use_bias=False, activation=None,
name='mobl%d_conv_project' % block_id)(x)
x = BatchNormalization(epsilon=1e-3, momentum=0.999,
name='bn%d_conv_bn_project' % block_id)(x)
if in_channels == pointwise_filters and stride == 1:
return Add(name='res_connect_' + str(block_id))([inputs, x])
return x
def _first_inverted_res_block(inputs,
stride,
alpha, filters, block_id):
"""Build the first inverted res block."""
in_channels = int(inputs.shape[-1])
pointwise_conv_filters = int(filters * alpha)
pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
# Depthwise
x = DepthwiseConv2D(kernel_size=3,
strides=stride, activation=None,
use_bias=False, padding='same',
name='mobl%d_conv_depthwise' %
block_id)(inputs)
x = BatchNormalization(epsilon=1e-3, momentum=0.999,
name='bn%d_conv_depthwise' %
block_id)(x)
x = Activation(relu6, name='conv_dw_%d_relu' % block_id)(x)
# Project
x = Conv2D(pointwise_filters,
kernel_size=1,
padding='same',
use_bias=False,
activation=None,
name='mobl%d_conv_project' %
block_id)(x)
x = BatchNormalization(epsilon=1e-3, momentum=0.999,
name='bn%d_conv_project' %
block_id)(x)
if in_channels == pointwise_filters and stride == 1:
return Add(name='res_connect_' + str(block_id))([inputs, x])
return x
research/object_detection/models/keras_models/inception_resnet_v2.py 0 → 100644
View file @ 80444539
# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""A wrapper around the Keras InceptionResnetV2 models for object detection."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
from object_detection.core import freezable_batch_norm
class _LayersOverride(object):
"""Alternative Keras layers interface for the Keras InceptionResNetV2."""
def __init__(self,
batchnorm_training,
output_stride=16,
align_feature_maps=False,
batchnorm_scale=False,
default_batchnorm_momentum=0.999,
default_batchnorm_epsilon=1e-3,
weight_decay=0.00004):
"""Alternative tf.keras.layers interface, for use by InceptionResNetV2.
It is used by the Keras applications kwargs injection API to
modify the Inception Resnet V2 Keras application with changes required by
the Object Detection API.
These injected interfaces make the following changes to the network:
- Supports freezing batch norm layers
- Adds support for feature map alignment (like in the Slim model)
- Adds support for changing the output stride (like in the Slim model)
- Adds support for overriding various batch norm hyperparameters
Because the Keras inception resnet v2 application does not assign explicit
names to most individual layers, the injection of output stride support
works by identifying convolution layers according to their filter counts
and pre-feature-map-alignment padding arguments.
Args:
batchnorm_training: Bool. Assigned to Batch norm layer `training` param
when constructing `freezable_batch_norm.FreezableBatchNorm` layers.
output_stride: A scalar that specifies the requested ratio of input to
output spatial resolution. Only supports 8 and 16.
align_feature_maps: When true, changes all the VALID paddings in the
network to SAME padding so that the feature maps are aligned.
batchnorm_scale: If True, uses an explicit `gamma` multiplier to scale the
activations in the batch normalization layer.
default_batchnorm_momentum: Float. Batch norm layers will be constructed
using this value as the momentum.
default_batchnorm_epsilon: small float added to variance to avoid
dividing by zero.
weight_decay: the l2 regularization weight decay for weights variables.
(gets multiplied by 0.5 to map from slim l2 regularization weight to
Keras l2 regularization weight).
"""
self._use_atrous = output_stride == 8
self._align_feature_maps = align_feature_maps
self._batchnorm_training = batchnorm_training
self._batchnorm_scale = batchnorm_scale
self._default_batchnorm_momentum = default_batchnorm_momentum
self._default_batchnorm_epsilon = default_batchnorm_epsilon
self.regularizer = tf.keras.regularizers.l2(weight_decay * 0.5)
def Conv2D(self, filters, kernel_size, **kwargs):
"""Builds a Conv2D layer according to the current Object Detection config.
Overrides the Keras InceptionResnetV2 application's convolutions with ones
that follow the spec specified by the Object Detection hyperparameters.
If feature map alignment is enabled, the padding will be forced to 'same'.
If output_stride is 8, some conv2d layers will be matched according to
their name or filter counts or pre-alignment padding parameters, and will
have the correct 'dilation rate' or 'strides' set.
Args:
filters: The number of filters to use for the convolution.
kernel_size: The kernel size to specify the height and width of the 2D
convolution window.
**kwargs: Keyword args specified by the Keras application for
constructing the convolution.
Returns:
A Keras Conv2D layer specified by the Object Detection hyperparameter
configurations.
"""
kwargs['kernel_regularizer'] = self.regularizer
kwargs['bias_regularizer'] = self.regularizer
# Because the Keras application does not set explicit names for most layers,
# (instead allowing names to auto-increment), we must match individual
# layers in the model according to their filter count, name, or
# pre-alignment mapping. This means we can only align the feature maps
# after we have applied our updates in cases where output_stride=8.
if self._use_atrous and (filters == 384):
kwargs['strides'] = 1
name = kwargs.get('name')
if self._use_atrous and (
(name and 'block17' in name) or
(filters == 128 or filters == 160 or
(filters == 192 and kwargs.get('padding', '').lower() != 'valid'))):
kwargs['dilation_rate'] = 2
if self._align_feature_maps:
kwargs['padding'] = 'same'
return tf.keras.layers.Conv2D(filters, kernel_size, **kwargs)
def MaxPooling2D(self, pool_size, strides, **kwargs):
"""Builds a pooling layer according to the current Object Detection config.
Overrides the Keras InceptionResnetV2 application's MaxPooling2D layers with
ones that follow the spec specified by the Object Detection hyperparameters.
If feature map alignment is enabled, the padding will be forced to 'same'.
If output_stride is 8, some pooling layers will be matched according to
their pre-alignment padding parameters, and will have their 'strides'
argument overridden.
Args:
pool_size: The pool size specified by the Keras application.
strides: The strides specified by the unwrapped Keras application.
**kwargs: Keyword args specified by the Keras application for
constructing the max pooling layer.
Returns:
A MaxPool2D layer specified by the Object Detection hyperparameter
configurations.
"""
if self._use_atrous and kwargs.get('padding', '').lower() == 'valid':
strides = 1
if self._align_feature_maps:
kwargs['padding'] = 'same'
return tf.keras.layers.MaxPool2D(pool_size, strides=strides, **kwargs)
# We alias MaxPool2D because Keras has that alias
MaxPool2D = MaxPooling2D # pylint: disable=invalid-name
def BatchNormalization(self, **kwargs):
"""Builds a normalization layer.
Overrides the Keras application batch norm with the norm specified by the
Object Detection configuration.
Args:
**kwargs: Keyword arguments from the `layers.BatchNormalization` calls in
the Keras application.
Returns:
A normalization layer specified by the Object Detection hyperparameter
configurations.
"""
kwargs['scale'] = self._batchnorm_scale
return freezable_batch_norm.FreezableBatchNorm(
training=self._batchnorm_training,
epsilon=self._default_batchnorm_epsilon,
momentum=self._default_batchnorm_momentum,
**kwargs)
# Forward all non-overridden methods to the keras layers
def __getattr__(self, item):
return getattr(tf.keras.layers, item)
# pylint: disable=invalid-name
def inception_resnet_v2(
batchnorm_training,
output_stride=16,
align_feature_maps=False,
batchnorm_scale=False,
weight_decay=0.00004,
default_batchnorm_momentum=0.9997,
default_batchnorm_epsilon=0.001,
**kwargs):
"""Instantiates the InceptionResnetV2 architecture.
(Modified for object detection)
This wraps the InceptionResnetV2 tensorflow Keras application, but uses the
Keras application's kwargs-based monkey-patching API to override the Keras
architecture with the following changes:
- Supports freezing batch norm layers with FreezableBatchNorms
- Adds support for feature map alignment (like in the Slim model)
- Adds support for changing the output stride (like in the Slim model)
- Changes the default batchnorm momentum to 0.9997
- Adds support for overriding various batchnorm hyperparameters
Args:
batchnorm_training: Bool. Assigned to Batch norm layer `training` param
when constructing `freezable_batch_norm.FreezableBatchNorm` layers.
output_stride: A scalar that specifies the requested ratio of input to
output spatial resolution. Only supports 8 and 16.
align_feature_maps: When true, changes all the VALID paddings in the
network to SAME padding so that the feature maps are aligned.
batchnorm_scale: If True, uses an explicit `gamma` multiplier to scale the
activations in the batch normalization layer.
weight_decay: the l2 regularization weight decay for weights variables.
(gets multiplied by 0.5 to map from slim l2 regularization weight to
Keras l2 regularization weight).
default_batchnorm_momentum: Float. Batch norm layers will be constructed
using this value as the momentum.
default_batchnorm_epsilon: small float added to variance to avoid
dividing by zero.
**kwargs: Keyword arguments forwarded directly to the
`tf.keras.applications.InceptionResNetV2` method that constructs the
Keras model.
Returns:
A Keras model instance.
"""
if output_stride != 8 and output_stride != 16:
raise ValueError('output_stride must be 8 or 16.')
layers_override = _LayersOverride(
batchnorm_training,
output_stride,
align_feature_maps=align_feature_maps,
batchnorm_scale=batchnorm_scale,
default_batchnorm_momentum=default_batchnorm_momentum,
default_batchnorm_epsilon=default_batchnorm_epsilon,
weight_decay=weight_decay)
return tf.keras.applications.InceptionResNetV2(
layers=layers_override, **kwargs)
# pylint: enable=invalid-name
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment