GitLab

* Updating python API to use CombinedNonMaxSuppresion TF operator

1. Adds a unit test to test post_processing python API
2. Currently sets clip_window to None as the kernel uses the default
   clip_window of [0,0,1,1]
3. Added use_static_shapes to the API. In old API if
   use_static_shapes is true, then it pads/clips outputs to max_total_size, if
specified. If not specified, it pads to num_classes*max_size_per_class.
 If use_static_shapes is false, it always pads/clips to max_total_size.

Update unit test to account for clipped bouding boxes

Changed the name to CombinedNonMaxSuppression based on feedback from Google

Added additional parameters to combinedNMS python function. They are currently
unused and required for networks like FasterRCNN and MaskRCNN

* Delete selected_indices from API

Because it was removed from CombinedNMS recently in the PR.

* Improve doc of function combined_non_max_suppression

* Enable CombinedNonMaxSuppression for first_stage_nms

* fix bug

* Ensure agnostic_nms is not used with combined_nms

Remove redundant arguments from combined_nms

* Fix pylint

* Add checks for unsupported args

* Fix pylint

* Move combined_non_max_suppression to batch_multiclass_non_max_suppression

Also rename combined_nms to use_combined_nms

* Delete combined_nms for first_stage_nms because it does not work

* Revert "Delete combined_nms for first_stage_nms because it does not work"

This reverts commit 2a3cc5145f17cee630a67ddedd20e90c2920fa9f.

* Use nmsed_additional_fields.get to avoid error

* Merge combined_non_max_suppression with main nms function

* Rename combined_nms for first stage nms

* Improve  docs

* Use assertListEqual for numpy arrays

* Fix pylint errors

* End comments with period

257914648  by lzc:

    Internal changes

--
257525973  by Zhichao Lu:

    Fixes bug that silently prevents checkpoints from loading when training w/ eager + functions. Also sets up scripts to run training.

--
257296614  by Zhichao Lu:

    Adding detection_features to model outputs

--
257234565  by Zhichao Lu:

    Fix wrong order of `classes_with_max_scores` in class-agnostic NMS caused by
    sorting in partitioned-NMS.

--
257232002  by ronnyvotel:

    Supporting `filter_nonoverlapping` option in np_box_list_ops.clip_to_window().

--
257198282  by Zhichao Lu:

    Adding the focal loss and l1 loss from the Objects as Points paper.

--
257089535  by Zhichao Lu:

    Create Keras based ssd + resnetv1 + fpn.

--
257087407  by Zhichao Lu:

    Make object_detection/data_decoders Python3-compatible.

--
257004582  by Zhichao Lu:

    Updates _decode_raw_data_into_masks_and_boxes to the latest binary masks-to-string encoding format.

--
257002124  by Zhichao Lu:

    Make object_detection/utils Python3-compatible, except json_utils.

    The patching trick used in json_utils is not going to work in Python 3.

--
256795056  by lzc:

    Add a detection_anchor_indices field to detection outputs.

--
256477542  by Zhichao Lu:

    Make object_detection/core Python3-compatible.

--
256387593  by Zhichao Lu:

    Edit class_id_function_approximations builder to skip class ids not present in label map.

--
256259039  by Zhichao Lu:

    Move NMS to TPU for FasterRCNN.

--
256071360  by rathodv:

    When multiclass_scores is empty, add one-hot encoding of groundtruth_classes as multiclass scores so that data_augmentation ops that expect the presence of multiclass_scores don't have to individually handle this case.

    Also copy input tensor_dict to out_tensor_dict first to avoid inplace modification.

--
256023645  by Zhichao Lu:

    Adds the first WIP iterations of TensorFlow v2 eager + functions style custom training & evaluation loops.

--
255980623  by Zhichao Lu:

    Adds a new data augmentation operation "remap_labels" which remaps a set of labels to a new label.

--
255753259  by Zhichao Lu:

    Announcement of the released evaluation tutorial for Open Images Challenge
    2019.

--
255698776  by lzc:

    Fix rewrite_nn_resize_op function which was broken by tf forward compatibility movement.

--
255623150  by Zhichao Lu:

    Add Keras-based ResnetV1 models.

--
255504992  by Zhichao Lu:

    Fixing the typo in specifying label expansion for ground truth segmentation
    file.

--
255470768  by Zhichao Lu:

    1. Fixing Python bug with parsed arguments.
    2. Adding capability to parse relevant columns from CSV header.
    3. Fixing bug with duplicated labels expansion.

--
255462432  by Zhichao Lu:

    Adds a new data augmentation operation "drop_label_probabilistically" which drops a given label with the given probability. This supports experiments on training in the presence of label noise.

--
255441632  by rathodv:

    Fallback on groundtruth classes when multiclass_scores tensor is empty.

--
255434899  by Zhichao Lu:

    Ensuring evaluation binary can run even with big files by synchronizing
    processing of ground truth and predictions: in this way, ground truth is not stored but immediatly
    used for evaluation. In case gt of object masks, this allows to run
    evaluations on relatively large sets.

--
255337855  by lzc:

    Internal change.

--
255308908  by Zhichao Lu:

    Add comment to clarify usage of calibration parameters proto.

--
255266371  by Zhichao Lu:

    Ensuring correct processing of the case, when no groundtruth masks are provided
    for an image.

--
255236648  by Zhichao Lu:

    Refactor model_builder in faster_rcnn.py to a util_map, so that it's possible to be overwritten.

--
255093285  by Zhichao Lu:

    Updating capability to subsample data during evaluation

--
255081222  by rathodv:

    Convert groundtruth masks to be of type float32 before its used in the loss function.

    When using mixed precision training, masks are represented using bfloat16 tensors in the input pipeline for performance reasons. We need to convert them to float32 before using it in the loss function.

--
254788436  by Zhichao Lu:

    Add forward_compatible to non_max_suppression_with_scores to make it is
    compatible with older tensorflow version.

--
254442362  by Zhichao Lu:

    Add num_layer field to ssd feature extractor proto.

--
253911582  by jonathanhuang:

    Plumbs Soft-NMS options (using the new tf.image.non_max_suppression_with_scores op) into the TF Object Detection API.  It adds a `soft_nms_sigma` field to the postprocessing proto file and plumbs this through to both the multiclass and class_agnostic versions of NMS. Note that there is no effect on behavior of NMS when soft_nms_sigma=0 (which it is set to by default).

    See also "Soft-NMS -- Improving Object Detection With One Line of Code" by Bodla et al (https://arxiv.org/abs/1704.04503)

--
253703949  by Zhichao Lu:

    Internal test fixes.

--
253151266  by Zhichao Lu:

    Fix the op type check for FusedBatchNorm, given that we introduced
    FusedBatchNormV3 in a previous change.

--
252718956  by Zhichao Lu:

    Customize activation function to enable relu6 instead of relu for saliency
    prediction model seastarization

--
252158593  by Zhichao Lu:

    Make object_detection/core Python3-compatible.

--
252150717  by Zhichao Lu:

    Make object_detection/core Python3-compatible.

--
251967048  by Zhichao Lu:

    Make GraphRewriter proto extensible.

--
251950039  by Zhichao Lu:

    Remove experimental_export_device_assignment from TPUEstimator.export_savedmodel(), so as to remove rewrite_for_inference().

    As a replacement, export_savedmodel() V2 API supports device_assignment where user call tpu.rewrite in model_fn and pass in device_assigment there.

--
251890697  by rathodv:

    Updated docstring to include new output nodes.

--
251662894  by Zhichao Lu:

    Add autoaugment augmentation option to objection detection api codebase. This
    is an available option in preprocessor.py.

    The intended usage of autoaugment is to be done along with random flipping and
    cropping for best results.

--
251532908  by Zhichao Lu:

    Add TrainingDataType enum to track whether class-specific or agnostic data was used to fit the calibration function.

    This is useful, since classes with few observations may require a calibration function fit on all classes.

--
251511339  by Zhichao Lu:

    Add multiclass isotonic regression to the calibration builder.

--
251317769  by pengchong:

    Internal Change.

--
250729989  by Zhichao Lu:

    Fixing bug in gt statistics count in case of mask and box annotations.

--
250729627  by Zhichao Lu:

    Label expansion for segmentation.

--
250724905  by Zhichao Lu:

    Fix use_depthwise in fpn and test it with fpnlite on ssd + mobilenet v2.

--
250670379  by Zhichao Lu:

    Internal change

250630364  by lzc:

    Fix detection_model_zoo footnotes

--
250560654  by Zhichao Lu:

    Fix static shape issue in matmul_crop_and_resize.

--
250534857  by Zhichao Lu:

    Edit class agnostic calibration function docstring to more accurately describe the function's outputs.

--
250533277  by Zhichao Lu:

    Edit the multiclass messages to use class ids instead of labels.

--

PiperOrigin-RevId: 257914648

250447559  by Zhichao Lu:

    Update expected files format for Instance Segmentation challenge:
    - add fields ImageWidth, ImageHeight and store the values per prediction
    - as mask, store only encoded image and assume its size is ImageWidth x ImageHeight

--
250402780  by rathodv:

    Fix failing Mask R-CNN TPU convergence test.

    Cast second stage prediction tensors from bfloat16 to float32 to prevent errors in third target assignment (Mask Prediction) - Concat with different types bfloat16 and bfloat32 isn't allowed.

--
250300240  by Zhichao Lu:

    Addion Open Images Challenge 2019 object detection and instance segmentation
    support into Estimator framework.

--
249944839  by rathodv:

    Modify exporter.py to add multiclass score nodes in exported inference graphs.

--
249935201  by rathodv:

    Modify postprocess methods to preserve multiclass scores after non max suppression.

--
249878079  by Zhichao Lu:

    This CL slightly refactors some Object Detection helper functions for data creation, evaluation, and groundtruth providing.

    This will allow the eager+function custom loops to share code with the existing estimator training loops.

    Concretely we make the following changes:
    1. In input creation we separate dataset-creation into top-level helpers, and allow it to optionally accept a pre-constructed model directly instead of always creating a model from the config just for feature preprocessing.

    2. In coco evaluation we split the update_op creation into its own function, which the custom loops will call directly.

    3. In model_lib we move groundtruth providing/ datastructure munging into a helper function

    4. For now we put an escape hatch in `_summarize_target_assignment` when executing in tf v2.0 behavior because the summary apis used only work w/ tf 1.x

--
249673507  by rathodv:

    Use explicit casts instead of tf.to_float and tf.to_int32 to avoid warnings.

--
249656006  by Zhichao Lu:

    Add named "raw_keypoint_locations" node that corresponds with the "raw_box_locations" node.

--
249651674  by rathodv:

    Keep proposal boxes in float format. MatMulCropAndResize can handle the type even when feature themselves are bfloat16s.

--
249568633  by rathodv:

    Support q > 1 in class agnostic NMS.
    Break post_processing_test.py into 3 separate files to avoid linter errors.

--
249535530  by rathodv:

    Update some deprecated arguments to tf ops.

--
249368223  by rathodv:

    Modify MatMulCropAndResize to use MultiLevelRoIAlign method and move the tests to spatial_transform_ops.py module.

    This cl establishes that CropAndResize and RoIAlign are equivalent and only differ in the sampling point grid within the boxes. CropAndResize uses a uniform size x size point grid such that the corner points exactly overlap box corners, while RoiAlign divides boxes into size x size cells and uses their centers as sampling points. In this cl, we switch MatMulCropAndResize to use the MultiLevelRoIAlign implementation with `align_corner` option as MultiLevelRoIAlign implementation is more memory efficient on TPU when compared to the original MatMulCropAndResize.

--
249337338  by chowdhery:

    Add class-agnostic non-max-suppression in post_processing

--
249139196  by Zhichao Lu:

    Fix positional argument bug in export_tflite_ssd_graph

--
249120219  by Zhichao Lu:

    Add evaluator for computing precision limited to a given recall range.

--
249030593  by Zhichao Lu:

    Evaluation util to run segmentation and detection challenge evaluation.

--
248554358  by Zhichao Lu:

    This change contains the auxiliary changes required for TF 2.0 style training with eager+functions+dist strat loops, but not the loops themselves.

    It includes:
    - Updates to shape usage to support both tensorshape v1 and tensorshape v2
    - A fix to FreezableBatchNorm to not override the `training` arg in call when `None` was passed to the constructor (Not an issue in the estimator loops but it was in the custom loops)
    - Puts some constants in init_scope so they work in eager + functions
    - Makes learning rate schedules return a callable in eager mode (required so they update when the global_step changes)
    - Makes DetectionModel a tf.module so it tracks variables (e.g. ones nested in layers)
    - Removes some references to `op.name` for some losses and replaces it w/ explicit names
    - A small part of the change to allow the coco evaluation metrics to work in eager mode

--
248271226  by rathodv:

    Add MultiLevel RoIAlign op.

--
248229103  by rathodv:

    Add functions to 1. pad features maps 2. ravel 5-D indices

--
248206769  by rathodv:

    Add utilities needed to introduce RoI Align op.

--
248177733  by pengchong:

    Internal changes

--
247742582  by Zhichao Lu:

    Open Images Challenge 2019 instance segmentation metric: part 2

--
247525401  by Zhichao Lu:

    Update comments on max_class_per_detection.

--
247520753  by rathodv:

    Add multilevel crop and resize operation that builds on top of matmul_crop_and_resize.

--
247391600  by Zhichao Lu:

    Open Images Challenge 2019 instance segmentation metric

--
247325813  by chowdhery:

    Quantized MobileNet v2 SSD FPNLite config with depth multiplier 0.75

--

PiperOrigin-RevId: 250447559

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 Zh...

np.uint8 to np.int64 for detection class label index overflow when using OIDv4 labels

The evaluation crashes in python3 if iteritems() is used instead of items()

236813471  by lzc:

    Internal change.

--
236507310  by lzc:

    Fix preprocess.random_resize_method config type issue. The target height and width will be passed as "size" to tf.image.resize_images which only accepts integer.

--
236409989  by Zhichao Lu:

    Config export_to_tpu from function parameter instead of HParams for TPU inference.

--
236403186  by Zhichao Lu:

    Make graph file names optional arguments.

--
236237072  by Zhichao Lu:

    Minor bugfix for keyword args.

--
236209602  by Zhichao Lu:

    Add support for PartitionedVariable to get_variables_available_in_checkpoint.

--
235828658  by Zhichao Lu:

    Automatically stop evaluation jobs when training is finished.

--
235817964  by Zhichao Lu:

    Add an optional process_metrics_fn callback to eval_util, it gets called
    with evaluation results once each evaluation is complete.

--
235788721  by lzc:

    Fix yml file tf runtime...

This import from google3.third_party doesn't work.

This file was inadvertantly renamed in #6071.

As discussed here: https://github.com/tensorflow/tensorflow/issues/11312, this change allows this script to work in Python 3.

This file was re-located below the tensorflow/lite/g3doc/convert
Signed-off-by: MyungSung Kwak <yesmung@gmail.com>

* Remove duplicate elif case in _check_and_convert_legacy_input_config_key

* Remove duplicate eval_input_path case.

Fixes the variable name for the tf proto example function.

223075771  by lzc:

    Bring in external fixes.

--
222919755  by ronnyvotel:

    Bug fix in faster r-cnn model builder. Was previously using `inplace_batchnorm_update` for `reuse_weights`.

--
222885680  by Zhichao Lu:

    Use the result_dict_for_batched_example in models_lib
    Also fixes the visualization size on when eval is on GPU

--
222883648  by Zhichao Lu:

    Fix _unmatched_class_label for the _add_background_class == False case in ssd_meta_arch.py.

--
222836663  by Zhichao Lu:

    Adding support for visualizing grayscale images. Without this change, the images are black-red instead of grayscale.

--
222501978  by Zhichao Lu:

    Fix a bug that caused convert_to_grayscale flag not to be respected.

--
222432846  by richardmunoz:

    Fix mapping of groundtruth_confidences from shape [num_boxes] to [num_boxes, num_classes] when the input contains the groundtruth_confidences field.

--
221725755  by richardmunoz:

    Internal change.

--
221458536  by Zhichao Lu:

    Fix saver defer build bug in object detection train codepath.

--
221391590  by Zhichao Lu:

    Add support for group normalization in the object detection API. Just adding MobileNet-v1 SSD currently. This may serve as a road map for other models that wish to support group normalization as an option.

--
221367993  by Zhichao Lu:

    Bug fixes (1) Make RandomPadImage work, (2) Fix keep_checkpoint_every_n_hours.

--
221266403  by rathodv:

    Use detection boxes as proposals to compute correct mask loss in eval jobs.

--
220845934  by lzc:

    Internal change.

--
220778850  by Zhichao Lu:

    Incorporating existing metrics into Estimator framework.
    Should restore:
    -oid_challenge_detection_metrics
    -pascal_voc_detection_metrics
    -weighted_pascal_voc_detection_metrics
    -pascal_voc_instance_segmentation_metrics
    -weighted_pascal_voc_instance_segmentation_metrics
    -oid_V2_detection_metrics

--
220370391  by alirezafathi:

    Adding precision and recall to the metrics.

--
220321268  by Zhichao Lu:

    Allow the option of setting max_examples_to_draw to zero.

--
220193337  by Zhichao Lu:

    This CL fixes a bug where the Keras convolutional box predictor was applying heads in the non-deterministic dict order. The consequence of this bug was that variables were created in non-deterministic orders. This in turn led different workers in a multi-gpu training setup to have slightly different graphs which had variables assigned to mismatched parameter servers. As a result, roughly half of all workers were unable to initialize and did no work, and training time was slowed down approximately 2x.

--
220136508  by huizhongc:

    Add weight equalization loss to SSD meta arch.

--
220125875  by pengchong:

    Rename label_scores to label_weights

--
219730108  by Zhichao Lu:

    Add description of detection_keypoints in postprocessed_tensors to docstring.

--
219577519  by pengchong:

    Support parsing the class confidences and training using them.

--
219547611  by lzc:

    Stop using static shapes in GPU eval jobs.

--
219536476  by Zhichao Lu:

    Migrate TensorFlow Lite out of tensorflow/contrib

    This change moves //tensorflow/contrib/lite to //tensorflow/lite in preparation
    for TensorFlow 2.0's deprecation of contrib/. If you refer to TF Lite build
    targets or headers, you will need to update them manually. If you use TF Lite
    from the TensorFlow python package, "tf.contrib.lite" now points to "tf.lite".
    Please update your imports as soon as possible.

    For more details, see https://groups.google.com/a/tensorflow.org/forum/#!topic/tflite/iIIXOTOFvwQ

    @angersson and @aselle are conducting this migration. Please contact them if
    you have any further questions.

--
219190083  by Zhichao Lu:

    Add a second expected_loss_weights function using an alternative expectation calculation compared to previous. Integrate this op into ssd_meta_arch and losses builder. Affects files that use losses_builder.build to handle the returning of an additional element.

--
218924451  by pengchong:

    Add a new way to assign training targets using groundtruth confidences.

--
218760524  by chowdhery:

    Modify export script to add option for regular NMS in TFLite post-processing op.

--

PiperOrigin-RevId: 223075771

Updated to 1.9 for consistency with running_pets.md

Another error appears when an incompatible version of Tensorboard is installed. Need to ensure Tensorboard is 1.9 as well

Updated doc as well line 211.

Line 219 runtime version updated to 1.9, causes error otherwise

`lite` is no longer under `contrib`. 

Additionally, `--config=opt` fails, see: https://github.com/tensorflow/serving/issues/517

* Internal change.

PiperOrigin-RevId: 213914693

* Add original_image_spatial_shape tensor in input dictionary to store shape of the original input image

PiperOrigin-RevId: 214018767

* Remove "groundtruth_confidences" from decoders use "groundtruth_weights" to indicate label confidence.

This also solves a bug that only surfaced now - random crop routines in core/preprocessor.py did not correctly handle "groundtruth_weight" tensors returned by the decoders.

PiperOrigin-RevId: 214091843

* Update CocoMaskEvaluator to allow for a batch of image info, rather than a single image.

PiperOrigin-RevId: 214295305

* Adding the option to be able to summarize gradients.

PiperOrigin-RevId: 214310875

* Adds FasterRCNN inference on CPU

1. Adds a flag use_static_shapes_for_eval to restrict to the ops that guarantees static shape.
2. No filtering of overlapping anchors while clipping the anchors when use_static_shapes_for_eval is set to True.
3. Adds test for faster_rcnn_meta_arch for predict and postprocess in inference mode for first and second stages.

PiperOrigin-RevId: 214329565

* Fix model_lib eval_spec_names assignment (integer->string).

PiperOrigin-RevId: 214335461

* Refactor Mask HEAD to optionally upsample after applying convolutions on ROI crops.

PiperOrigin-RevId: 214338440

* Uses final_exporter_name as exporter_name for the first eval spec for backward compatibility.

PiperOrigin-RevId: 214522032

* Add reshaped `mask_predictions` tensor to the prediction dictionary in `_predict_third_stage` method to allow computing mask loss in eval job.

PiperOrigin-RevId: 214620716

* Add support for fully conv training to fpn.

PiperOrigin-RevId: 214626274

* Fix the proprocess() function in Resnet v1 to make it work for any number of input channels.

Note: If the #channels != 3, this will simply skip the mean subtraction in preprocess() function.
PiperOrigin-RevId: 214635428

* Wrap result_dict_for_single_example in eval_util to run for batched examples.

PiperOrigin-RevId: 214678514

* Adds PNASNet-based (ImageNet model) feature extractor for SSD.

PiperOrigin-RevId: 214988331

* Update documentation

PiperOrigin-RevId: 215243502

* Correct index used to compute number of groundtruth/detection boxes in COCOMaskEvaluator.

Due to an incorrect indexing in cl/214295305 only the first detection mask and first groundtruth mask for a given image are fed to the COCO Mask evaluation library. Since groundtruth masks are arranged in no particular order, the first and highest scoring detection mask (detection masks are ordered by score) won't match the the first and only groundtruth retained in all cases. This is I think why mask evaluation metrics do not get better than ~11 mAP. Note that this code path is only active when using model_main.py binary for evaluation.

This change fixes the indices and modifies an existing test case to cover it.

PiperOrigin-RevId: 215275936

* Fixing grayscale_image_resizer to accept mask as input.

PiperOrigin-RevId: 215345836

* Add an option not to clip groundtruth boxes during preprocessing. Clipping boxes adversely affects training for partially occluded or large objects, especially for fully conv models. Clipping already occurs during postprocessing, and should not occur during training.

PiperOrigin-RevId: 215613379

* Always return recalls and precisions with length equal to the number of classes.

The previous behavior of ObjectDetectionEvaluation was somewhat dangerous: when no groundtruth boxes were present, the lists of per-class precisions and recalls were simply truncated. Unless you were aware of this phenomenon (and consulted the `num_gt_instances_per_class` vector) it was difficult to associate each metric with each class.

PiperOrigin-RevId: 215633711

* Expose the box feature node in SSD.

PiperOrigin-RevId: 215653316

* Fix ssd mobilenet v2 _CONV_DEFS overwriting issue.

PiperOrigin-RevId: 215654160

* More documentation updates

PiperOrigin-RevId: 215656580

* Add pooling + residual option in multi_resolution_feature_maps. It adds an average pooling and a residual layer between feature maps with matching depth. Designed to be used with WeightSharedBoxPredictor.

PiperOrigin-RevId: 215665619

* Only call create_modificed_mobilenet_config on init if use_depthwise is true.

PiperOrigin-RevId: 215784290

* Only call create_modificed_mobilenet_config on init if use_depthwise is true.

PiperOrigin-RevId: 215837524

* Don't prune keypoints if clip_boxes is false.

PiperOrigin-RevId: 216187642

* Makes sure "key" field exists in the result dictionary.

PiperOrigin-RevId: 216456543

* Add add_background_class parameter to allow disabling the inclusion of a background class.

PiperOrigin-RevId: 216567612

* Update expected_classification_loss_under_sampling to better account for expected sampling.

PiperOrigin-RevId: 216712287

* Let the evaluation receive a evaluation class in its constructor.

PiperOrigin-RevId: 216769374

* This CL adds model building & training support for end-to-end Keras-based SSD models. If a Keras feature extractor's name is specified in the model config (e.g. 'ssd_mobilenet_v2_keras'), the model will use that feature extractor and a corresponding Keras-based box predictor.

This CL makes sure regularization losses & batch norm updates work correctly when training models that have Keras-based components. It also updates the default hyperparameter settings of the keras-based mobilenetV2 (when not overriding hyperparams) to more closely match the legacy Slim training scope.

PiperOrigin-RevId: 216938707

* Adding the ability in the coco evaluator to indicate whether an image has been annotated. For a non-annotated image, detections and groundtruth are not supplied.

PiperOrigin-RevId: 217316342

* Release the 8k minival dataset ids for MSCOCO, used in Huang et al. "Speed/accuracy trade-offs for modern convolutional object detectors" (https://arxiv.org/abs/1611.10012)

PiperOrigin-RevId: 217549353

* Exposes weighted_sigmoid_focal loss for faster rcnn classifier

PiperOrigin-RevId: 217601740

* Add detection_features to output nodes. The shape of the feature is [batch_size, max_detections, depth].

PiperOrigin-RevId: 217629905

* FPN uses a custom NN resize op for TPU-compatibility. Replace this op with the Tensorflow version at export time for TFLite-compatibility.

PiperOrigin-RevId: 217721184

* Compute `num_groundtruth_boxes` in inputs.tranform_input_data_fn after data augmentation instead of decoders.

PiperOrigin-RevId: 217733432

* 1. Stop gradients from flowing into groundtruth masks with zero paddings.
2. Normalize pixelwise cross entropy loss across the whole batch.

PiperOrigin-RevId: 217735114

* Optimize Input pipeline for Mask R-CNN on TPU with blfoat16: improve the step time from:
1663.6 ms -> 1184.2 ms, about 28.8% improvement.

PiperOrigin-RevId: 217748833

* Fixes to export a TPU compatible model

Adds nodes to each of the output tensor. Also increments the value of class labels by 1.

PiperOrigin-RevId: 217856760

* API changes:
 - change the interface of target assigner to return per-class weights.
 - change the interface of classification loss to take per-class weights.

PiperOrigin-RevId: 217968393

* Add an option to override pipeline config in export_saved_model using command line arg

PiperOrigin-RevId: 218429292

* Include Quantized trained MobileNet V2 SSD and FaceSsd in model zoo.

PiperOrigin-RevId: 218530947

* Write final config to disk in `train` mode only.

PiperOrigin-RevId: 218735512

* Merged commit includes the following changes:
213899768  by Sergio Guadarrama:

    Fixes #3819.

--
213493831  by Sergio Guadarrama:

    Internal change

212057654  by Sergio Guadarrama:

    Internal change

210747685  by Sergio Guadarrama:

    For FPN, when use_depthwise is set to true, use slightly modified mobilenet v1 config.

--
210128931  by Sergio Guadarrama:

    Allow user-defined current_step in NASNet.

--
209092664  by Sergio Guadarrama:

    Add quantized fine-tuning / training / eval and export to slim image classifier binaries.

--
207651347  by Sergio Guadarrama:

    Update mobilenet v1 docs to include revised tflite models.

--
207165245  by Sergio Guadarrama:

    Internal change

207095064  by Sergio Guadarrama:

    Internal change

PiperOrigin-RevId: 213899768

* Update model_lib.py to fix eval_spec name issue.

save people the trouble of training a model and trying to convert only to realize they should have used SSD model

214018767  by Zhichao Lu:

    Add original_image_spatial_shape tensor in input dictionary to store shape of the original input image

--
213914693  by lzc:

    Internal change.

--
213872175  by Zhichao Lu:

    This CL adds a Keras-based mobilenet_v2 feature extractor for object detection models.

    As part of this CL, we use the Keras mobilenet_v2 application's keyword argument layer injection API to allow the generated network to support the object detection hyperparameters.

--
213848499  by Zhichao Lu:

    Replace tf.image.resize_nearest_neighbor with tf.image.resize_images. tf.image.resize_nearest_neighbor only supports 4-D tensors but masks is a 3-D tensor.

--
213758622  by lzc:

    Internal change.

--

PiperOrigin-RevId: 214018767

Release iNaturalist Species-trained models, refactor of evaluation, box predictor for object detection. (#5289)

* Merged commit includes the following changes:
212389173  by Zhichao Lu:

    1. Replace tf.boolean_mask with tf.where

--
212282646  by Zhichao Lu:

    1. Fix a typo in model_builder.py and add a test to cover it.

--
212142989  by Zhichao Lu:

    Only resize masks in meta architecture if it has not already been resized in the input pipeline.

--
212136935  by Zhichao Lu:

    Choose matmul or native crop_and_resize in the model builder instead of faster r-cnn meta architecture.

--
211907984  by Zhichao Lu:

    Make eval input reader repeated field and update config util to handle this field.

--
211858098  by Zhichao Lu:

    Change the implementation of merge_boxes_with_multiple_labels.

--
211843915  by Zhichao Lu:

    Add Mobilenet v2 + FPN support.

--
211655076  by Zhichao Lu:

    Bug fix for generic keys in config overrides

    In generic configuration overrides, we had a duplicate entry for train_input_config and we were missing the eval_input_config and eval_config.

    This change also introduces testing for all config overrides.

--
211157501  by Zhichao Lu:

    Make the locally-modified conv defs a copy.

    So that it doesn't modify MobileNet conv defs globally for other code that
    transitively imports this package.

--
211112813  by Zhichao Lu:

    Refactoring visualization tools for Estimator's eval_metric_ops. This will make it easier for future models to take advantage of a single interface and mechanics.

--
211109571  by Zhichao Lu:

    A test decorator.

--
210747685  by Zhichao Lu:

    For FPN, when use_depthwise is set to true, use slightly modified mobilenet v1 config.

--
210723882  by Zhichao Lu:

    Integrating the losses mask into the meta architectures. When providing groundtruth, one can optionally specify annotation information (i.e. which images are labeled vs. unlabeled). For any image that is unlabeled, there is no loss accumulation.

--
210673675  by Zhichao Lu:

    Internal change.

--
210546590  by Zhichao Lu:

    Internal change.

--
210529752  by Zhichao Lu:

    Support batched inputs with ops.matmul_crop_and_resize.

    With this change the new inputs are images of shape [batch, heigh, width, depth] and boxes of shape [batch, num_boxes, 4]. The output tensor is of the shape [batch, num_boxes, crop_height, crop_width, depth].

--
210485912  by Zhichao Lu:

    Fix TensorFlow version check in object_detection_tutorial.ipynb

--
210484076  by Zhichao Lu:

    Reduce TPU memory required for single image matmul_crop_and_resize.

    Using tf.einsum eliminates intermediate tensors, tiling and expansion. for an image of size [40, 40, 1024] and boxes of shape [300, 4] HBM memory usage goes down from 3.52G to 1.67G.

--
210468361  by Zhichao Lu:

    Remove PositiveAnchorLossCDF/NegativeAnchorLossCDF to resolve "Main thread is not in main loop error" issue in local training.

--
210100253  by Zhichao Lu:

    Pooling pyramid feature maps: add option to replace max pool with convolution layers.

--
209995842  by Zhichao Lu:

    Fix a bug which prevents variable sharing in Faster RCNN.

--
209965526  by Zhichao Lu:

    Add support for enabling export_to_tpu through the estimator.

--
209946440  by Zhichao Lu:

    Replace deprecated tf.train.Supervisor with tf.train.MonitoredSession. MonitoredSession also takes away the hassle of starting queue runners.

--
209888003  by Zhichao Lu:

    Implement function to handle data where source_id is not set.

    If the field source_id is found to be the empty string for any image during runtime, it will be replaced with a random string. This avoids hash-collisions on dataset where many examples do not have source_id set. Those hash-collisions have unintended site effects and may lead to bugs in the detection pipeline.

--
209842134  by Zhichao Lu:

    Converting loss mask into multiplier, rather than using it as a boolean mask (which changes tensor shape). This is necessary, since other utilities (e.g. hard example miner) require a loss matrix with the same dimensions as the original prediction tensor.

--
209768066  by Zhichao Lu:

    Adding ability to remove loss computation from specific images in a batch, via an optional boolean mask.

--
209722556  by Zhichao Lu:

    Remove dead code.

    (_USE_C_API was flipped to True by default in TensorFlow 1.8)

--
209701861  by Zhichao Lu:

    This CL cleans-up some tf.Example creation snippets, by reusing the convenient tf.train.Feature building functions in dataset_util.

--
209697893  by Zhichao Lu:

    Do not overwrite num_epoch for eval input. This leads to errors in some cases.

--
209694652  by Zhichao Lu:

    Sample boxes by jittering around the currently given boxes.

--
209550300  by Zhichao Lu:

    `create_category_index_from_labelmap()` function now accepts `use_display_name` parameter.
    Also added create_categories_from_labelmap function for convenience

--
209490273  by Zhichao Lu:

    Check result_dict type before accessing image_id via key.

--
209442529  by Zhichao Lu:

    Introducing the capability to sample examples for evaluation. This makes it easy to specify one full epoch of evaluation, or a subset (e.g. sample 1 of every N examples).

--
208941150  by Zhichao Lu:

    Adding the capability of exporting the results in json format.

--
208888798  by Zhichao Lu:

    Fixes wrong dictionary key for num_det_boxes_per_image.

--
208873549  by Zhichao Lu:

    Reduce the number of HLO ops created by matmul_crop_and_resize.

    Do not unroll along the channels dimension. Instead, transpose the input image dimensions, apply tf.matmul and transpose back.

    The number of HLO instructions for 1024 channels reduce from 12368 to 110.

--
208844315  by Zhichao Lu:

    Add an option to use tf.non_maximal_supression_padded in SSD post-process

--
208731380  by Zhichao Lu:

    Add field in box_predictor config to enable mask prediction and update builders accordingly.

--
208699405  by Zhichao Lu:

    This CL creates a keras-based multi-resolution feature map extractor.

--
208557208  by Zhichao Lu:

    Add TPU tests for Faster R-CNN Meta arch.

    * Tests that two_stage_predict and total_loss tests run successfully on TPU.
    * Small mods to multiclass_non_max_suppression to preserve static shapes.

--
208499278  by Zhichao Lu:

    This CL makes sure the Keras convolutional box predictor & head layers apply activation layers *after* normalization (as opposed to before).

--
208391694  by Zhichao Lu:

    Updating visualization tool to produce multiple evaluation images.

--
208275961  by Zhichao Lu:

    This CL adds a Keras version of the Convolutional Box Predictor, as well as more general infrastructure for making Keras Prediction heads & Keras box predictors.

--
208275585  by Zhichao Lu:

    This CL enables the Keras layer hyperparameter object to build a dedicated activation layer, and to disable activation by default in the op layer construction kwargs.

    This is necessary because in most cases the normalization layer must be applied before the activation layer. So, in Keras models we must set the convolution activation in a dedicated layer after normalization is applied, rather than setting it in the convolution layer construction args.

--
208263792  by Zhichao Lu:

    Add a new SSD mask meta arch that can predict masks for SSD models.
    Changes including:
     - overwrite loss function to add mask loss computation.
     - update ssd_meta_arch to handle masks if predicted in predict and postprocessing.

--
208000218  by Zhichao Lu:

    Make FasterRCNN choose static shape operations only in training mode.

--
207997797  by Zhichao Lu:

    Add static boolean_mask op to box_list_ops.py and use that in faster_rcnn_meta_arch.py to support use_static_shapes option.

--
207993460  by Zhichao Lu:

    Include FGVC detection models in model zoo.

--
207971213  by Zhichao Lu:

    remove the restriction to run tf.nn.top_k op on CPU

--
207961187  by Zhichao Lu:

    Build the first stage NMS function in the model builder and pass it to FasterRCNN meta arch.

--
207960608  by Zhichao Lu:

    Internal Change.

--
207927015  by Zhichao Lu:

    Have an option to use the TPU compatible NMS op cl/206673787, in the batch_multiclass_non_max_suppression function. On setting pad_to_max_output_size to true, the output nmsed boxes are padded to be of length max_size_per_class.

    This can be used in first stage Region Proposal Network in FasterRCNN model by setting the first_stage_nms_pad_to_max_proposals field to true in config proto.

--
207809668  by Zhichao Lu:

    Add option to use depthwise separable conv instead of conv2d in FPN and WeightSharedBoxPredictor. More specifically, there are two related configs:
    - SsdFeatureExtractor.use_depthwise
    - WeightSharedConvolutionalBoxPredictor.use_depthwise

--
207808651  by Zhichao Lu:

    Fix the static balanced positive negative sampler's TPU tests

--
207798658  by Zhichao Lu:

    Fixes a post-refactoring bug where the pre-prediction convolution layers in the convolutional box predictor are ignored.

--
207796470  by Zhichao Lu:

    Make slim endpoints visible in FasterRCNNMetaArch.

--
207787053  by Zhichao Lu:

    Refactor ssd_meta_arch so that the target assigner instance is passed into the SSDMetaArch constructor rather than constructed inside.

--

PiperOrigin-RevId: 212389173

* Fix detection model zoo typo.

* Modify tf example decoder to handle label maps with either `display_name` or `name` fields seamlessly.

Currently, tf example decoder uses only `name` field to look up ids for class text field present in the data. This change uses both `display_name` and `name` fields in the label map to fetch ids for class text.

PiperOrigin-RevId: 212672223

* Modify create_coco_tf_record tool to write out class text instead of class labels.

PiperOrigin-RevId: 212679112

* Fix detection model zoo typo.

PiperOrigin-RevId: 212715692

* Adding the following two optional flags to WeightSharedConvolutionalBoxHead:
1) In the box head, apply clipping to box encodings in the box head.
2) In the class head, apply sigmoid to class predictions at inference time.

PiperOrigin-RevId: 212723242

* Support class confidences in merge boxes with multiple labels.

PiperOrigin-RevId: 212884998

* Creates multiple eval specs for object detection.

PiperOrigin-RevId: 212894556

* Set batch_norm on last layer in Mask Head to None.

PiperOrigin-RevId: 213030087

* Enable bfloat16 training for object detection models.

PiperOrigin-RevId: 213053547

* Skip padding op when unnecessary.

PiperOrigin-RevId: 213065869

* Modify `Matchers` to use groundtruth weights before performing matching.

Groundtruth weights tensor is used to indicate padding in groundtruth box tensor. It is handled in `TargetAssigner` by creating appropriate classification and regression target weights based on the groundtruth box each anchor matches to. However, options such as `force_match_all_rows` in `ArgmaxMatcher` force certain anchors to match to groundtruth boxes that are just paddings thereby reducing the number of anchors that could otherwise match to real groundtruth boxes.

For single stage models like SSD the effect of this is negligible as there are two orders of magnitude more anchors than the number of padded groundtruth boxes. But for Faster R-CNN and Mask R-CNN where there are only 300 anchors in the second stage, a significant number of these match to groundtruth paddings reducing the number of anchors regressing to real groundtruth boxes degrading the performance severely.

Therefore, this change introduces an additional boolean argument `valid_rows` to `Matcher.match` methods and the implementations now ignore such padded groudtruth boxes during matching.

PiperOrigin-RevId: 213345395

* Add release note for iNaturalist Species trained models.

PiperOrigin-RevId: 213347179

* Fix the bug of uninitialized gt_is_crowd_list variable.

PiperOrigin-RevId: 213364858

* ...text exposed to open source public git repo...

PiperOrigin-RevId: 213554260

* Fixed TensorFlow version check in object_detection_tutorial_ipynb

* Changed the minimum version to 1.9.0 for the object detection notebook

The bash script to submit training job for pets detection has runtime-version of 1.8. This will trigger `TypeError: non_max_suppression() got an unexpected keyword argument 'score_threshold'` on the Google Cloud since 1.8 and older does not support this keyword argument. Therefore, update this runtime version to 1.9, which is the most recent runtime version that is published on June 27, 2018. 

See 
https://github.com/tensorflow/models/issues/5056
https://cloud.google.com/ml-engine/docs/tensorflow/runtime-version-list

Updated model_lib to use min_score_threshold and max_num_boxes_to_visualize from the eval config.

* Merged commit includes the following changes:
207771702  by Zhichao Lu:

    Refactoring evaluation utilities so that it is easier to introduce new DetectionEvaluators with eval_metric_ops.

--
207758641  by Zhichao Lu:

    Require tensorflow version 1.9+ for running object detection API.

--
207641470  by Zhichao Lu:

    Clip `num_groundtruth_boxes` in pad_input_data_to_static_shapes() to `max_num_boxes`. This prevents a scenario where tensors are sliced to an invalid range in model_lib.unstack_batch().

--
207621728  by Zhichao Lu:

    This CL adds a FreezableBatchNorm that inherits from the Keras BatchNormalization layer, but supports freezing the `training` parameter at construction time instead of having to do it in the `call` method.

    It also adds a method to the `KerasLayerHyperparams` class that will build an appropriate FreezableBatchNorm layer according to the hyperparameter configuration. If batch_norm is disabled, this method returns and Identity layer.

    These will be used to simplify the conversion to Keras APIs.

--
207610524  by Zhichao Lu:

    Update anchor generators and box predictors for python3 compatibility.

--
207585122  by Zhichao Lu:

    Refactoring convolutional box predictor into separate prediction heads.

--
207549305  by Zhichao Lu:

    Pass all 1s for batch weights if nothing is specified in GT.

--
207336575  by Zhichao Lu:

    Move the new argument 'target_assigner_instance' to the end of the list of arguments to the ssd_meta_arch constructor for backwards compatibility.

--
207327862  by Zhichao Lu:

    Enable support for float output in quantized custom op for postprocessing in SSD Mobilenet model.

--
207323154  by Zhichao Lu:

    Bug fix: change dict.iteritems() to dict.items()

--
207301109  by Zhichao Lu:

    Integrating expected_classification_loss_under_sampling op as an option in the ssd_meta_arch

--
207286221  by Zhichao Lu:

    Adding an option to weight regression loss with foreground scores from the ground truth labels.

--
207231739  by Zhichao Lu:

    Explicitly mentioning the argument names when calling the batch target assigner.

--
207206356  by Zhichao Lu:

    Add include_trainable_variables field to train config to better handle trainable variables.

--
207135930  by Zhichao Lu:

    Internal change.

--
206862541  by Zhichao Lu:

    Do not unpad the outputs from batch_non_max_suppression before sampling.

    Since BalancedPositiveNegativeSampler takes an indicator for valid positions to sample from we can pass the output from NMS directly into Sampler.

--

PiperOrigin-RevId: 207771702

* Remove unused doc.

`iteritems()` was removed from python3. `items()` does the same functionality so changing it will work in both python2 and python3. The only difference as far as I know is `iteritems()` returns a generator where `items` returns a list. But for this this code it will not make any difference where we are just changing the key of the dict to a string.