revert reverted changes... (#7503)

research/lstm_object_detection/export_tflite_lstd_graph.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 r"""Exports an LSTM detection model to use with tf-lite.
 
 Outputs file:
@@ -86,8 +85,9 @@ python lstm_object_detection/export_tflite_lstd_graph.py \
 """
 
 import tensorflow as tf
-from lstm_object_detection.utils import config_util
+
 from lstm_object_detection import export_tflite_lstd_graph_lib
+from lstm_object_detection.utils import config_util
 
 flags = tf.app.flags
 flags.DEFINE_string('output_directory', None, 'Path to write outputs.')
@@ -122,12 +122,16 @@ def main(argv):
   flags.mark_flag_as_required('trained_checkpoint_prefix')
 
   pipeline_config = config_util.get_configs_from_pipeline_file(
-    FLAGS.pipeline_config_path)
+      FLAGS.pipeline_config_path)
 
   export_tflite_lstd_graph_lib.export_tflite_graph(
-      pipeline_config, FLAGS.trained_checkpoint_prefix, FLAGS.output_directory,
-      FLAGS.add_postprocessing_op, FLAGS.max_detections,
-      FLAGS.max_classes_per_detection, use_regular_nms=FLAGS.use_regular_nms)
+      pipeline_config,
+      FLAGS.trained_checkpoint_prefix,
+      FLAGS.output_directory,
+      FLAGS.add_postprocessing_op,
+      FLAGS.max_detections,
+      FLAGS.max_classes_per_detection,
+      use_regular_nms=FLAGS.use_regular_nms)
 
 
 if __name__ == '__main__':

research/lstm_object_detection/export_tflite_lstd_graph_lib.py

@@ -12,26 +12,26 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 r"""Exports detection models to use with tf-lite.
 
 See export_tflite_lstd_graph.py for usage.
 """
 import os
 import tempfile
+
 import numpy as np
 import tensorflow as tf
+
 from tensorflow.core.framework import attr_value_pb2
 from tensorflow.core.framework import types_pb2
 from tensorflow.core.protobuf import saver_pb2
 from tensorflow.tools.graph_transforms import TransformGraph
+from lstm_object_detection import model_builder
 from object_detection import exporter
 from object_detection.builders import graph_rewriter_builder
 from object_detection.builders import post_processing_builder
 from object_detection.core import box_list
 
-from lstm_object_detection import model_builder
-
 _DEFAULT_NUM_CHANNELS = 3
 _DEFAULT_NUM_COORD_BOX = 4
 
@@ -84,11 +84,11 @@ def append_postprocessing_op(frozen_graph_def,
     num_classes: number of classes in SSD detector
     scale_values: scale values is a dict with following key-value pairs
       {y_scale: 10, x_scale: 10, h_scale: 5, w_scale: 5} that are used in decode
-      centersize boxes
+        centersize boxes
     detections_per_class: In regular NonMaxSuppression, number of anchors used
-    for NonMaxSuppression per class
-    use_regular_nms: Flag to set postprocessing op to use Regular NMS instead
-      of Fast NMS.
+      for NonMaxSuppression per class
+    use_regular_nms: Flag to set postprocessing op to use Regular NMS instead of
+      Fast NMS.
 
   Returns:
     transformed_graph_def: Frozen GraphDef with postprocessing custom op
@@ -165,9 +165,9 @@ def export_tflite_graph(pipeline_config,
   is written to output_dir/tflite_graph.pb.
 
   Args:
-    pipeline_config: Dictionary of configuration objects. Keys are `model`, `train_config`,
-      `train_input_config`, `eval_config`, `eval_input_config`, `lstm_model`.
-      Value are the corresponding config objects.
+    pipeline_config: Dictionary of configuration objects. Keys are `model`,
+      `train_config`, `train_input_config`, `eval_config`, `eval_input_config`,
+      `lstm_model`. Value are the corresponding config objects.
     trained_checkpoint_prefix: a file prefix for the checkpoint containing the
       trained parameters of the SSD model.
     output_dir: A directory to write the tflite graph and anchor file to.
@@ -176,9 +176,9 @@ def export_tflite_graph(pipeline_config,
     max_detections: Maximum number of detections (boxes) to show
     max_classes_per_detection: Number of classes to display per detection
     detections_per_class: In regular NonMaxSuppression, number of anchors used
-    for NonMaxSuppression per class
-    use_regular_nms: Flag to set postprocessing op to use Regular NMS instead
-      of Fast NMS.
+      for NonMaxSuppression per class
+    use_regular_nms: Flag to set postprocessing op to use Regular NMS instead of
+      Fast NMS.
     binary_graph_name: Name of the exported graph file in binary format.
     txt_graph_name: Name of the exported graph file in text format.
 
@@ -197,12 +197,10 @@ def export_tflite_graph(pipeline_config,
 
   num_classes = model_config.ssd.num_classes
   nms_score_threshold = {
-      model_config.ssd.post_processing.batch_non_max_suppression.
-      score_threshold
+      model_config.ssd.post_processing.batch_non_max_suppression.score_threshold
   }
   nms_iou_threshold = {
-      model_config.ssd.post_processing.batch_non_max_suppression.
-      iou_threshold
+      model_config.ssd.post_processing.batch_non_max_suppression.iou_threshold
   }
   scale_values = {}
   scale_values['y_scale'] = {
@@ -226,7 +224,7 @@ def export_tflite_graph(pipeline_config,
     width = image_resizer_config.fixed_shape_resizer.width
     if image_resizer_config.fixed_shape_resizer.convert_to_grayscale:
       num_channels = 1
-    #TODO(richardbrks) figure out how to make with a None defined batch size
+
     shape = [lstm_config.eval_unroll_length, height, width, num_channels]
   else:
     raise ValueError(
@@ -235,14 +233,14 @@ def export_tflite_graph(pipeline_config,
             image_resizer_config.WhichOneof('image_resizer_oneof')))
 
   video_tensor = tf.placeholder(
-    tf.float32, shape=shape, name='input_video_tensor')
+      tf.float32, shape=shape, name='input_video_tensor')
 
-  detection_model = model_builder.build(model_config, lstm_config,
-    is_training=False)
+  detection_model = model_builder.build(
+      model_config, lstm_config, is_training=False)
   preprocessed_video, true_image_shapes = detection_model.preprocess(
-    tf.to_float(video_tensor))
+      tf.to_float(video_tensor))
   predicted_tensors = detection_model.predict(preprocessed_video,
-    true_image_shapes)
+                                              true_image_shapes)
   # predicted_tensors = detection_model.postprocess(predicted_tensors,
   #   true_image_shapes)
   # The score conversion occurs before the post-processing custom op
@@ -311,7 +309,7 @@ def export_tflite_graph(pipeline_config,
       initializer_nodes='')
 
   # Add new operation to do post processing in a custom op (TF Lite only)
-  #(richardbrks) Do use this or detection_model.postprocess?
+
   if add_postprocessing_op:
     transformed_graph_def = append_postprocessing_op(
         frozen_graph_def, max_detections, max_classes_per_detection,

research/lstm_object_detection/export_tflite_lstd_model.py

@@ -13,6 +13,8 @@
 # limitations under the License.
 # ==============================================================================
 
+"""Export a LSTD model in tflite format."""
+
 import os
 from absl import flags
 import tensorflow as tf
@@ -29,34 +31,35 @@ FLAGS = flags.FLAGS
 
 
 def main(_):
-    flags.mark_flag_as_required('export_path')
-    flags.mark_flag_as_required('frozen_graph_path')
-    flags.mark_flag_as_required('pipeline_config_path')
-
-    configs = config_util.get_configs_from_pipeline_file(
-        FLAGS.pipeline_config_path)
-    lstm_config = configs['lstm_model']
-
-    input_arrays = ['input_video_tensor']
-    output_arrays = [
-        'TFLite_Detection_PostProcess',
-        'TFLite_Detection_PostProcess:1',
-        'TFLite_Detection_PostProcess:2',
-        'TFLite_Detection_PostProcess:3',
-    ]
-    input_shapes = {
-        'input_video_tensor': [lstm_config.eval_unroll_length, 320, 320, 3],
-    }
-
-    converter = tf.lite.TFLiteConverter.from_frozen_graph(
-        FLAGS.frozen_graph_path, input_arrays, output_arrays,
-        input_shapes=input_shapes
-    )
-    converter.allow_custom_ops = True
-    tflite_model = converter.convert()
-    ofilename = os.path.join(FLAGS.export_path)
-    open(ofilename, "wb").write(tflite_model)
+  flags.mark_flag_as_required('export_path')
+  flags.mark_flag_as_required('frozen_graph_path')
+  flags.mark_flag_as_required('pipeline_config_path')
+
+  configs = config_util.get_configs_from_pipeline_file(
+      FLAGS.pipeline_config_path)
+  lstm_config = configs['lstm_model']
+
+  input_arrays = ['input_video_tensor']
+  output_arrays = [
+      'TFLite_Detection_PostProcess',
+      'TFLite_Detection_PostProcess:1',
+      'TFLite_Detection_PostProcess:2',
+      'TFLite_Detection_PostProcess:3',
+  ]
+  input_shapes = {
+      'input_video_tensor': [lstm_config.eval_unroll_length, 320, 320, 3],
+  }
+
+  converter = tf.lite.TFLiteConverter.from_frozen_graph(
+      FLAGS.frozen_graph_path,
+      input_arrays,
+      output_arrays,
+      input_shapes=input_shapes)
+  converter.allow_custom_ops = True
+  tflite_model = converter.convert()
+  ofilename = os.path.join(FLAGS.export_path)
+  open(ofilename, 'wb').write(tflite_model)
 
 
 if __name__ == '__main__':
-    tf.app.run()
+  tf.app.run()

research/lstm_object_detection/g3doc/exporting_models.md

 # Exporting a tflite model from a checkpoint
 
-Starting from a trained model checkpoint, creating a tflite model requires 2 steps:
-
-* exporting a tflite frozen graph from a checkpoint
-* exporting a tflite model from a frozen graph
+Starting from a trained model checkpoint, creating a tflite model requires 2
+steps:
 
+*   exporting a tflite frozen graph from a checkpoint
+*   exporting a tflite model from a frozen graph
 
 ## Exporting a tflite frozen graph from a checkpoint
 
@@ -20,14 +20,14 @@ python lstm_object_detection/export_tflite_lstd_graph.py \
     --pipeline_config_path ${PIPELINE_CONFIG_PATH} \
     --trained_checkpoint_prefix ${TRAINED_CKPT_PREFIX} \
     --output_directory ${EXPORT_DIR} \
-	--add_preprocessing_op
+    --add_preprocessing_op
 ```
 
-After export, you should see the directory ${EXPORT_DIR} containing the following files:
-
-* `tflite_graph.pb`
-* `tflite_graph.pbtxt` 
+After export, you should see the directory ${EXPORT_DIR} containing the
+following files:
 
+*   `tflite_graph.pb`
+*   `tflite_graph.pbtxt`
 
 ## Exporting a tflite model from a frozen graph
 
@@ -40,10 +40,10 @@ FROZEN_GRAPH_PATH={path to exported tflite_graph.pb}
 EXPORT_PATH={path to filename that will be used for export}
 PIPELINE_CONFIG_PATH={path to pipeline config}
 python lstm_object_detection/export_tflite_lstd_model.py \
-	   --export_path ${EXPORT_PATH} \
-	   --frozen_graph_path ${FROZEN_GRAPH_PATH} \
-	   --pipeline_config_path ${PIPELINE_CONFIG_PATH}
+       --export_path ${EXPORT_PATH} \
+       --frozen_graph_path ${FROZEN_GRAPH_PATH} \
+       --pipeline_config_path ${PIPELINE_CONFIG_PATH}
 ```
 
 After export, you should see the file ${EXPORT_PATH} containing the FlatBuffer
-model to be used by an application.
\ No newline at end of file
+model to be used by an application.

research/lstm_object_detection/test_tflite_model.py

@@ -13,6 +13,9 @@
 # limitations under the License.
 # ==============================================================================
 
+"""Test a tflite model using random input data."""
+
+from __future__ import print_function
 from absl import flags
 import numpy as np
 import tensorflow as tf
@@ -23,28 +26,28 @@ FLAGS = flags.FLAGS
 
 def main(_):
 
-    flags.mark_flag_as_required('model_path')
+  flags.mark_flag_as_required('model_path')
 
-    # Load TFLite model and allocate tensors.
-    interpreter = tf.lite.Interpreter(model_path=FLAGS.model_path)
-    interpreter.allocate_tensors()
+  # Load TFLite model and allocate tensors.
+  interpreter = tf.lite.Interpreter(model_path=FLAGS.model_path)
+  interpreter.allocate_tensors()
 
-    # Get input and output tensors.
-    input_details = interpreter.get_input_details()
-    print 'input_details:', input_details
-    output_details = interpreter.get_output_details()
-    print 'output_details:', output_details
+  # Get input and output tensors.
+  input_details = interpreter.get_input_details()
+  print('input_details:', input_details)
+  output_details = interpreter.get_output_details()
+  print('output_details:', output_details)
 
-    # Test model on random input data.
-    input_shape = input_details[0]['shape']
-    # change the following line to feed into your own data.
-    input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
-    interpreter.set_tensor(input_details[0]['index'], input_data)
+  # Test model on random input data.
+  input_shape = input_details[0]['shape']
+  # change the following line to feed into your own data.
+  input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
+  interpreter.set_tensor(input_details[0]['index'], input_data)
 
-    interpreter.invoke()
-    output_data = interpreter.get_tensor(output_details[0]['index'])
-    print output_data
+  interpreter.invoke()
+  output_data = interpreter.get_tensor(output_details[0]['index'])
+  print(output_data)
 
 
 if __name__ == '__main__':
-    tf.app.run()
+  tf.app.run()

research/lstm_object_detection/tflite/BUILD

@@ -59,12 +59,19 @@ cc_library(
     name = "mobile_lstd_tflite_client",
     srcs = ["mobile_lstd_tflite_client.cc"],
     hdrs = ["mobile_lstd_tflite_client.h"],
+    defines = select({
+        "//conditions:default": [],
+        "enable_edgetpu": ["ENABLE_EDGETPU"],
+    }),
     deps = [
         ":mobile_ssd_client",
         ":mobile_ssd_tflite_client",
-        "@com_google_absl//absl/base:core_headers",
         "@com_google_glog//:glog",
+        "@com_google_absl//absl/base:core_headers",
         "@org_tensorflow//tensorflow/lite/kernels:builtin_ops",
-    ],
+    ] + select({
+        "//conditions:default": [],
+        "enable_edgetpu": ["@libedgetpu//libedgetpu:header"],
+    }),
     alwayslink = 1,
 )

research/lstm_object_detection/tflite/WORKSPACE

@@ -90,12 +90,6 @@ http_archive(
     sha256 = "79d102c61e2a479a0b7e5fc167bcfaa4832a0c6aad4a75fa7da0480564931bcc",
 )
 
-#
-# http_archive(
-#     name = "com_google_protobuf",
-#     strip_prefix = "protobuf-master",
-#     urls = ["https://github.com/protocolbuffers/protobuf/archive/master.zip"],
-# )
 
 # Needed by TensorFlow
 http_archive(

research/lstm_object_detection/tflite/mobile_lstd_tflite_client.cc

@@ -66,6 +66,11 @@ bool MobileLSTDTfLiteClient::InitializeInterpreter(
     interpreter_->UseNNAPI(false);
   }
 
+#ifdef ENABLE_EDGETPU
+  interpreter_->SetExternalContext(kTfLiteEdgeTpuContext,
+                                   edge_tpu_context_.get());
+#endif
+
   // Inputs are: normalized_input_image_tensor, raw_inputs/init_lstm_c,
   // raw_inputs/init_lstm_h
   if (interpreter_->inputs().size() != 3) {

research/lstm_object_detection/tflite/mobile_ssd_tflite_client.h

@@ -76,6 +76,10 @@ class MobileSSDTfLiteClient : public MobileSSDClient {
   std::unique_ptr<::tflite::MutableOpResolver> resolver_;
   std::unique_ptr<::tflite::Interpreter> interpreter_;
 
+#ifdef ENABLE_EDGETPU
+  std::unique_ptr<edgetpu::EdgeTpuContext> edge_tpu_context_;
+#endif
+
  private:
   // MobileSSDTfLiteClient is neither copyable nor movable.
   MobileSSDTfLiteClient(const MobileSSDTfLiteClient&) = delete;
@@ -103,10 +107,6 @@ class MobileSSDTfLiteClient : public MobileSSDClient {
   bool FloatInference(const uint8_t* input_data);
   bool QuantizedInference(const uint8_t* input_data);
   void GetOutputBoxesAndScoreTensorsFromUInt8();
-
-#ifdef ENABLE_EDGETPU
-  std::unique_ptr<edgetpu::EdgeTpuContext> edge_tpu_context_;
-#endif
 };
 
 }  // namespace tflite