Minor fixes for object detection (#5613)

* 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

research/object_detection/meta_architectures/ssd_meta_arch.py

@@ -19,7 +19,6 @@ models.
 """
 from abc import abstractmethod
 
-import re
 import tensorflow as tf
 
 from object_detection.core import box_list
@@ -116,6 +115,25 @@ class SSDFeatureExtractor(object):
     """
     raise NotImplementedError
 
+  def restore_from_classification_checkpoint_fn(self, feature_extractor_scope):
+    """Returns a map of variables to load from a foreign checkpoint.
+
+    Args:
+      feature_extractor_scope: A scope name for the 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():
+      var_name = variable.op.name
+      if var_name.startswith(feature_extractor_scope + '/'):
+        var_name = var_name.replace(feature_extractor_scope + '/', '')
+        variables_to_restore[var_name] = variable
+
+    return variables_to_restore
+
 
 class SSDKerasFeatureExtractor(tf.keras.Model):
   """SSD Feature Extractor definition."""
@@ -218,6 +236,25 @@ class SSDKerasFeatureExtractor(tf.keras.Model):
   def call(self, inputs, **kwargs):
     return self._extract_features(inputs)
 
+  def restore_from_classification_checkpoint_fn(self, feature_extractor_scope):
+    """Returns a map of variables to load from a foreign checkpoint.
+
+    Args:
+      feature_extractor_scope: A scope name for the 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():
+      var_name = variable.op.name
+      if var_name.startswith(feature_extractor_scope + '/'):
+        var_name = var_name.replace(feature_extractor_scope + '/', '')
+        variables_to_restore[var_name] = variable
+
+    return variables_to_restore
+
 
 class SSDMetaArch(model.DetectionModel):
   """SSD Meta-architecture definition."""
@@ -333,13 +370,15 @@ class SSDMetaArch(model.DetectionModel):
       # Slim feature extractors get an explicit naming scope
       self._extract_features_scope = 'FeatureExtractor'
 
-    # TODO(jonathanhuang): handle agnostic mode
-    # weights
-    self._unmatched_class_label = tf.constant([1] + self.num_classes * [0],
-                                              tf.float32)
-    if encode_background_as_zeros:
+    if self._add_background_class and encode_background_as_zeros:
       self._unmatched_class_label = tf.constant((self.num_classes + 1) * [0],
                                                 tf.float32)
+    elif self._add_background_class:
+      self._unmatched_class_label = tf.constant([1] + self.num_classes * [0],
+                                                tf.float32)
+    else:
+      self._unmatched_class_label = tf.constant(self.num_classes * [0],
+                                                tf.float32)
 
     self._target_assigner = target_assigner_instance
 
@@ -606,14 +645,22 @@ class SSDMetaArch(model.DetectionModel):
       detection_boxes = tf.identity(detection_boxes, 'raw_box_locations')
       detection_boxes = tf.expand_dims(detection_boxes, axis=2)
 
-      detection_scores_with_background = self._score_conversion_fn(
-          class_predictions)
-      detection_scores_with_background = tf.identity(
-          detection_scores_with_background, 'raw_box_scores')
-      detection_scores = tf.slice(detection_scores_with_background, [0, 0, 1],
-                                  [-1, -1, -1])
+      detection_scores = self._score_conversion_fn(class_predictions)
+      detection_scores = tf.identity(detection_scores, 'raw_box_scores')
+      if self._add_background_class:
+        detection_scores = tf.slice(detection_scores, [0, 0, 1], [-1, -1, -1])
       additional_fields = None
 
+      batch_size = (
+          shape_utils.combined_static_and_dynamic_shape(preprocessed_images)[0])
+
+      if 'feature_maps' in prediction_dict:
+        feature_map_list = []
+        for feature_map in prediction_dict['feature_maps']:
+          feature_map_list.append(tf.reshape(feature_map, [batch_size, -1]))
+        box_features = tf.concat(feature_map_list, 1)
+        box_features = tf.identity(box_features, 'raw_box_features')
+
       if detection_keypoints is not None:
         additional_fields = {
             fields.BoxListFields.keypoints: detection_keypoints}
@@ -683,17 +730,20 @@ class SSDMetaArch(model.DetectionModel):
             self.groundtruth_lists(fields.BoxListFields.boxes), match_list)
 
       if self._random_example_sampler:
+        batch_cls_per_anchor_weights = tf.reduce_mean(
+            batch_cls_weights, axis=-1)
         batch_sampled_indicator = tf.to_float(
             shape_utils.static_or_dynamic_map_fn(
                 self._minibatch_subsample_fn,
-                [batch_cls_targets, batch_cls_weights],
+                [batch_cls_targets, batch_cls_per_anchor_weights],
                 dtype=tf.bool,
                 parallel_iterations=self._parallel_iterations,
                 back_prop=True))
         batch_reg_weights = tf.multiply(batch_sampled_indicator,
                                         batch_reg_weights)
-        batch_cls_weights = tf.multiply(batch_sampled_indicator,
-                                        batch_cls_weights)
+        batch_cls_weights = tf.multiply(
+            tf.expand_dims(batch_sampled_indicator, -1),
+            batch_cls_weights)
 
       losses_mask = None
       if self.groundtruth_has_field(fields.InputDataFields.is_annotated):
@@ -713,16 +763,32 @@ class SSDMetaArch(model.DetectionModel):
           losses_mask=losses_mask)
 
       if self._expected_classification_loss_under_sampling:
+        # Need to compute losses for assigned targets against the
+        # unmatched_class_label as well as their assigned targets.
+        # simplest thing (but wasteful) is just to calculate all losses
+        # twice
+        batch_size, num_anchors, num_classes = batch_cls_targets.get_shape()
+        unmatched_targets = tf.ones([batch_size, num_anchors, 1
+                                    ]) * self._unmatched_class_label
+
+        unmatched_cls_losses = self._classification_loss(
+            prediction_dict['class_predictions_with_background'],
+            unmatched_targets,
+            weights=batch_cls_weights,
+            losses_mask=losses_mask)
+
         if cls_losses.get_shape().ndims == 3:
           batch_size, num_anchors, num_classes = cls_losses.get_shape()
           cls_losses = tf.reshape(cls_losses, [batch_size, -1])
+          unmatched_cls_losses = tf.reshape(unmatched_cls_losses,
+                                            [batch_size, -1])
           batch_cls_targets = tf.reshape(
               batch_cls_targets, [batch_size, num_anchors * num_classes, -1])
           batch_cls_targets = tf.concat(
               [1 - batch_cls_targets, batch_cls_targets], axis=-1)
 
         cls_losses = self._expected_classification_loss_under_sampling(
-            batch_cls_targets, cls_losses)
+            batch_cls_targets, cls_losses, unmatched_cls_losses)
 
         classification_loss = tf.reduce_sum(cls_losses)
         localization_loss = tf.reduce_sum(location_losses)
@@ -971,6 +1037,26 @@ class SSDMetaArch(model.DetectionModel):
         [combined_shape[0], combined_shape[1], 4]))
     return decoded_boxes, decoded_keypoints
 
+  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.
+    """
+    losses = []
+    slim_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
+    # Copy the slim losses to avoid modifying the collection
+    if slim_losses:
+      losses.extend(slim_losses)
+    if self._box_predictor.is_keras_model:
+      losses.extend(self._box_predictor.losses)
+    if self._feature_extractor.is_keras_model:
+      losses.extend(self._feature_extractor.losses)
+    return losses
+
   def restore_map(self,
                   fine_tune_checkpoint_type='detection',
                   load_all_detection_checkpoint_vars=False):
@@ -997,18 +1083,44 @@ class SSDMetaArch(model.DetectionModel):
     if fine_tune_checkpoint_type not in ['detection', 'classification']:
       raise ValueError('Not supported fine_tune_checkpoint_type: {}'.format(
           fine_tune_checkpoint_type))
-    variables_to_restore = {}
-    for variable in tf.global_variables():
-      var_name = variable.op.name
-      if (fine_tune_checkpoint_type == 'detection' and
-          load_all_detection_checkpoint_vars):
-        variables_to_restore[var_name] = variable
-      else:
-        if var_name.startswith(self._extract_features_scope):
-          if fine_tune_checkpoint_type == 'classification':
-            var_name = (
-                re.split('^' + self._extract_features_scope + '/',
-                         var_name)[-1])
+
+    if fine_tune_checkpoint_type == 'classification':
+      return self._feature_extractor.restore_from_classification_checkpoint_fn(
+          self._extract_features_scope)
+
+    if fine_tune_checkpoint_type == 'detection':
+      variables_to_restore = {}
+      for variable in tf.global_variables():
+        var_name = variable.op.name
+        if load_all_detection_checkpoint_vars:
           variables_to_restore[var_name] = variable
+        else:
+          if var_name.startswith(self._extract_features_scope):
+            variables_to_restore[var_name] = variable
 
     return variables_to_restore
+
+  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 = []
+    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)
+    if self._box_predictor.is_keras_model:
+      update_ops.extend(self._box_predictor.get_updates_for(None))
+      update_ops.extend(self._box_predictor.get_updates_for(
+          self._box_predictor.inputs))
+    if self._feature_extractor.is_keras_model:
+      update_ops.extend(self._feature_extractor.get_updates_for(None))
+      update_ops.extend(self._feature_extractor.get_updates_for(
+          self._feature_extractor.inputs))
+    return update_ops

research/object_detection/meta_architectures/ssd_meta_arch_test.py

@@ -42,7 +42,7 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
                     random_example_sampling=False,
                     weight_regression_loss_by_score=False,
                     use_expected_classification_loss_under_sampling=False,
-                    minimum_negative_sampling=1,
+                    min_num_negative_samples=1,
                     desired_negative_sampling_ratio=3,
                     use_keras=False,
                     predict_mask=False,
@@ -57,7 +57,7 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
         weight_regression_loss_by_score=weight_regression_loss_by_score,
         use_expected_classification_loss_under_sampling=
         use_expected_classification_loss_under_sampling,
-        minimum_negative_sampling=minimum_negative_sampling,
+        min_num_negative_samples=min_num_negative_samples,
         desired_negative_sampling_ratio=desired_negative_sampling_ratio,
         use_keras=use_keras,
         predict_mask=predict_mask,
@@ -344,11 +344,11 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
     preprocessed_input = np.random.rand(batch_size, 2, 2, 3).astype(np.float32)
     groundtruth_boxes1 = np.array([[0, 0, .5, .5]], dtype=np.float32)
     groundtruth_boxes2 = np.array([[0, 0, .5, .5]], dtype=np.float32)
-    groundtruth_classes1 = np.array([[0, 1]], dtype=np.float32)
-    groundtruth_classes2 = np.array([[0, 1]], dtype=np.float32)
+    groundtruth_classes1 = np.array([[1]], dtype=np.float32)
+    groundtruth_classes2 = np.array([[1]], dtype=np.float32)
     expected_localization_loss = 0.0
     expected_classification_loss = (
-        batch_size * num_anchors * (num_classes + 1) * np.log(2.0))
+        batch_size * num_anchors * num_classes * np.log(2.0))
     (localization_loss, classification_loss) = self.execute(
         graph_fn, [
             preprocessed_input, groundtruth_boxes1, groundtruth_boxes2,
@@ -371,7 +371,7 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
           apply_hard_mining=False,
           add_background_class=True,
           use_expected_classification_loss_under_sampling=True,
-          minimum_negative_sampling=1,
+          min_num_negative_samples=1,
           desired_negative_sampling_ratio=desired_negative_sampling_ratio)
       model.provide_groundtruth(groundtruth_boxes_list,
                                 groundtruth_classes_list)
@@ -391,8 +391,7 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
     expected_localization_loss = 0.0
 
     expected_classification_loss = (
-        batch_size * (desired_negative_sampling_ratio * num_anchors +
-                      num_classes * num_anchors) * np.log(2.0))
+        batch_size * (num_anchors + num_classes * num_anchors) * np.log(2.0))
     (localization_loss, classification_loss) = self.execute(
         graph_fn, [
             preprocessed_input, groundtruth_boxes1, groundtruth_boxes2,
@@ -432,11 +431,11 @@ class SsdMetaArchTest(ssd_meta_arch_test_lib.SSDMetaArchTestBase,
     preprocessed_input = np.random.rand(batch_size, 2, 2, 3).astype(np.float32)
     groundtruth_boxes1 = np.array([[0, 0, 1, 1]], dtype=np.float32)
     groundtruth_boxes2 = np.array([[0, 0, 1, 1]], dtype=np.float32)
-    groundtruth_classes1 = np.array([[0, 1]], dtype=np.float32)
-    groundtruth_classes2 = np.array([[1, 0]], dtype=np.float32)
+    groundtruth_classes1 = np.array([[1]], dtype=np.float32)
+    groundtruth_classes2 = np.array([[0]], dtype=np.float32)
     expected_localization_loss = 0.25
     expected_classification_loss = (
-        batch_size * num_anchors * (num_classes + 1) * np.log(2.0))
+        batch_size * num_anchors * num_classes * np.log(2.0))
     (localization_loss, classification_loss) = self.execute(
         graph_fn, [
             preprocessed_input, groundtruth_boxes1, groundtruth_boxes2,

research/object_detection/meta_architectures/ssd_meta_arch_test_lib.py

@@ -119,7 +119,7 @@ class SSDMetaArchTestBase(test_case.TestCase):
                     random_example_sampling=False,
                     weight_regression_loss_by_score=False,
                     use_expected_classification_loss_under_sampling=False,
-                    minimum_negative_sampling=1,
+                    min_num_negative_samples=1,
                     desired_negative_sampling_ratio=3,
                     use_keras=False,
                     predict_mask=False,
@@ -130,10 +130,12 @@ class SSDMetaArchTestBase(test_case.TestCase):
     mock_anchor_generator = MockAnchorGenerator2x2()
     if use_keras:
       mock_box_predictor = test_utils.MockKerasBoxPredictor(
-          is_training, num_classes, predict_mask=predict_mask)
+          is_training, num_classes, add_background_class=add_background_class,
+          predict_mask=predict_mask)
     else:
       mock_box_predictor = test_utils.MockBoxPredictor(
-          is_training, num_classes, predict_mask=predict_mask)
+          is_training, num_classes, add_background_class=add_background_class,
+          predict_mask=predict_mask)
     mock_box_coder = test_utils.MockBoxCoder()
     if use_keras:
       fake_feature_extractor = FakeSSDKerasFeatureExtractor()
@@ -182,7 +184,7 @@ class SSDMetaArchTestBase(test_case.TestCase):
     if use_expected_classification_loss_under_sampling:
       expected_classification_loss_under_sampling = functools.partial(
           ops.expected_classification_loss_under_sampling,
-          minimum_negative_sampling=minimum_negative_sampling,
+          min_num_negative_samples=min_num_negative_samples,
           desired_negative_sampling_ratio=desired_negative_sampling_ratio)
 
     code_size = 4

research/object_detection/metrics/coco_evaluation.py

@@ -248,27 +248,30 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
         detection_boxes_batched,
         detection_scores_batched,
         detection_classes_batched,
-        num_det_boxes_per_image):
+        num_det_boxes_per_image,
+        is_annotated_batched):
       """Update operation for adding batch of images to Coco evaluator."""
 
       for (image_id, gt_box, gt_class, gt_is_crowd, num_gt_box, det_box,
-           det_score, det_class, num_det_box) in zip(
+           det_score, det_class, num_det_box, is_annotated) in zip(
                image_id_batched, groundtruth_boxes_batched,
                groundtruth_classes_batched, groundtruth_is_crowd_batched,
                num_gt_boxes_per_image,
                detection_boxes_batched, detection_scores_batched,
-               detection_classes_batched, num_det_boxes_per_image):
-        self.add_single_ground_truth_image_info(
-            image_id, {
-                'groundtruth_boxes': gt_box[:num_gt_box],
-                'groundtruth_classes': gt_class[:num_gt_box],
-                'groundtruth_is_crowd': gt_is_crowd[:num_gt_box]
-            })
-        self.add_single_detected_image_info(
-            image_id,
-            {'detection_boxes': det_box[:num_det_box],
-             'detection_scores': det_score[:num_det_box],
-             'detection_classes': det_class[:num_det_box]})
+               detection_classes_batched, num_det_boxes_per_image,
+               is_annotated_batched):
+        if is_annotated:
+          self.add_single_ground_truth_image_info(
+              image_id, {
+                  'groundtruth_boxes': gt_box[:num_gt_box],
+                  'groundtruth_classes': gt_class[:num_gt_box],
+                  'groundtruth_is_crowd': gt_is_crowd[:num_gt_box]
+              })
+          self.add_single_detected_image_info(
+              image_id,
+              {'detection_boxes': det_box[:num_det_box],
+               'detection_scores': det_score[:num_det_box],
+               'detection_classes': det_class[:num_det_box]})
 
     # Unpack items from the evaluation dictionary.
     input_data_fields = standard_fields.InputDataFields
@@ -284,6 +287,7 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
     num_gt_boxes_per_image = eval_dict.get(
         'num_groundtruth_boxes_per_image', None)
     num_det_boxes_per_image = eval_dict.get('num_det_boxes_per_image', None)
+    is_annotated = eval_dict.get('is_annotated', None)
 
     if groundtruth_is_crowd is None:
       groundtruth_is_crowd = tf.zeros_like(groundtruth_classes, dtype=tf.bool)
@@ -306,6 +310,11 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
         num_det_boxes_per_image = tf.shape(detection_boxes)[1:2]
       else:
         num_det_boxes_per_image = tf.expand_dims(num_det_boxes_per_image, 0)
+
+      if is_annotated is None:
+        is_annotated = tf.constant([True])
+      else:
+        is_annotated = tf.expand_dims(is_annotated, 0)
     else:
       if num_gt_boxes_per_image is None:
         num_gt_boxes_per_image = tf.tile(
@@ -315,6 +324,8 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
         num_det_boxes_per_image = tf.tile(
             tf.shape(detection_boxes)[1:2],
             multiples=tf.shape(detection_boxes)[0:1])
+      if is_annotated is None:
+        is_annotated = tf.ones_like(image_id, dtype=tf.bool)
 
     update_op = tf.py_func(update_op, [image_id,
                                        groundtruth_boxes,
@@ -324,7 +335,8 @@ class CocoDetectionEvaluator(object_detection_evaluation.DetectionEvaluator):
                                        detection_boxes,
                                        detection_scores,
                                        detection_classes,
-                                       num_det_boxes_per_image], [])
+                                       num_det_boxes_per_image,
+                                       is_annotated], [])
     metric_names = ['DetectionBoxes_Precision/mAP',
                     'DetectionBoxes_Precision/mAP@.50IOU',
                     'DetectionBoxes_Precision/mAP@.75IOU',
@@ -581,8 +593,11 @@ class CocoMaskEvaluator(object_detection_evaluation.DetectionEvaluator):
 
     Args:
       eval_dict: A dictionary that holds tensors for evaluating object detection
-        performance. This dictionary may be produced from
-        eval_util.result_dict_for_single_example().
+        performance. For single-image evaluation, this dictionary may be
+        produced from eval_util.result_dict_for_single_example(). If multi-image
+        evaluation, `eval_dict` should contain the fields
+        'num_groundtruth_boxes_per_image' and 'num_det_boxes_per_image' to
+        properly unpad the tensors from the batch.
 
     Returns:
       a dictionary of metric names to tuple of value_op and update_op that can
@@ -590,27 +605,41 @@ class CocoMaskEvaluator(object_detection_evaluation.DetectionEvaluator):
       update ops  must be run together and similarly all value ops must be run
       together to guarantee correct behaviour.
     """
-    def update_op(
-        image_id,
-        groundtruth_boxes,
-        groundtruth_classes,
-        groundtruth_instance_masks,
-        groundtruth_is_crowd,
-        detection_scores,
-        detection_classes,
-        detection_masks):
+
+    def update_op(image_id_batched, groundtruth_boxes_batched,
+                  groundtruth_classes_batched,
+                  groundtruth_instance_masks_batched,
+                  groundtruth_is_crowd_batched, num_gt_boxes_per_image,
+                  detection_scores_batched, detection_classes_batched,
+                  detection_masks_batched, num_det_boxes_per_image):
       """Update op for metrics."""
-      self.add_single_ground_truth_image_info(
-          image_id,
-          {'groundtruth_boxes': groundtruth_boxes,
-           'groundtruth_classes': groundtruth_classes,
-           'groundtruth_instance_masks': groundtruth_instance_masks,
-           'groundtruth_is_crowd': groundtruth_is_crowd})
-      self.add_single_detected_image_info(
-          image_id,
-          {'detection_scores': detection_scores,
-           'detection_classes': detection_classes,
-           'detection_masks': detection_masks})
+
+      for (image_id, groundtruth_boxes, groundtruth_classes,
+           groundtruth_instance_masks, groundtruth_is_crowd, num_gt_box,
+           detection_scores, detection_classes,
+           detection_masks, num_det_box) in zip(
+               image_id_batched, groundtruth_boxes_batched,
+               groundtruth_classes_batched, groundtruth_instance_masks_batched,
+               groundtruth_is_crowd_batched, num_gt_boxes_per_image,
+               detection_scores_batched, detection_classes_batched,
+               detection_masks_batched, num_det_boxes_per_image):
+        self.add_single_ground_truth_image_info(
+            image_id, {
+                'groundtruth_boxes':
+                    groundtruth_boxes[:num_gt_box],
+                'groundtruth_classes':
+                    groundtruth_classes[:num_gt_box],
+                'groundtruth_instance_masks':
+                    groundtruth_instance_masks[:num_gt_box],
+                'groundtruth_is_crowd':
+                    groundtruth_is_crowd[:num_gt_box]
+            })
+        self.add_single_detected_image_info(
+            image_id, {
+                'detection_scores': detection_scores[:num_det_box],
+                'detection_classes': detection_classes[:num_det_box],
+                'detection_masks': detection_masks[:num_det_box]
+            })
 
     # Unpack items from the evaluation dictionary.
     input_data_fields = standard_fields.InputDataFields
@@ -622,20 +651,54 @@ class CocoMaskEvaluator(object_detection_evaluation.DetectionEvaluator):
         input_data_fields.groundtruth_instance_masks]
     groundtruth_is_crowd = eval_dict.get(
         input_data_fields.groundtruth_is_crowd, None)
+    num_gt_boxes_per_image = eval_dict.get(
+        input_data_fields.num_groundtruth_boxes, None)
     detection_scores = eval_dict[detection_fields.detection_scores]
     detection_classes = eval_dict[detection_fields.detection_classes]
     detection_masks = eval_dict[detection_fields.detection_masks]
+    num_det_boxes_per_image = eval_dict.get(detection_fields.num_detections,
+                                            None)
 
     if groundtruth_is_crowd is None:
       groundtruth_is_crowd = tf.zeros_like(groundtruth_classes, dtype=tf.bool)
-    update_op = tf.py_func(update_op, [image_id,
-                                       groundtruth_boxes,
-                                       groundtruth_classes,
-                                       groundtruth_instance_masks,
-                                       groundtruth_is_crowd,
-                                       detection_scores,
-                                       detection_classes,
-                                       detection_masks], [])
+
+    if not image_id.shape.as_list():
+      # Apply a batch dimension to all tensors.
+      image_id = tf.expand_dims(image_id, 0)
+      groundtruth_boxes = tf.expand_dims(groundtruth_boxes, 0)
+      groundtruth_classes = tf.expand_dims(groundtruth_classes, 0)
+      groundtruth_instance_masks = tf.expand_dims(groundtruth_instance_masks, 0)
+      groundtruth_is_crowd = tf.expand_dims(groundtruth_is_crowd, 0)
+      detection_scores = tf.expand_dims(detection_scores, 0)
+      detection_classes = tf.expand_dims(detection_classes, 0)
+      detection_masks = tf.expand_dims(detection_masks, 0)
+
+      if num_gt_boxes_per_image is None:
+        num_gt_boxes_per_image = tf.shape(groundtruth_boxes)[1:2]
+      else:
+        num_gt_boxes_per_image = tf.expand_dims(num_gt_boxes_per_image, 0)
+
+      if num_det_boxes_per_image is None:
+        num_det_boxes_per_image = tf.shape(detection_scores)[1:2]
+      else:
+        num_det_boxes_per_image = tf.expand_dims(num_det_boxes_per_image, 0)
+    else:
+      if num_gt_boxes_per_image is None:
+        num_gt_boxes_per_image = tf.tile(
+            tf.shape(groundtruth_boxes)[1:2],
+            multiples=tf.shape(groundtruth_boxes)[0:1])
+      if num_det_boxes_per_image is None:
+        num_det_boxes_per_image = tf.tile(
+            tf.shape(detection_scores)[1:2],
+            multiples=tf.shape(detection_scores)[0:1])
+
+    update_op = tf.py_func(update_op, [
+        image_id, groundtruth_boxes, groundtruth_classes,
+        groundtruth_instance_masks, groundtruth_is_crowd,
+        num_gt_boxes_per_image, detection_scores, detection_classes,
+        detection_masks, num_det_boxes_per_image
+    ], [])
+
     metric_names = ['DetectionMasks_Precision/mAP',
                     'DetectionMasks_Precision/mAP@.50IOU',
                     'DetectionMasks_Precision/mAP@.75IOU',

research/object_detection/metrics/coco_evaluation_test.py

@@ -308,6 +308,99 @@ class CocoEvaluationPyFuncTest(tf.test.TestCase):
     self.assertFalse(coco_evaluator._detection_boxes_list)
     self.assertFalse(coco_evaluator._image_ids)
 
+  def testGetOneMAPWithMatchingGroundtruthAndDetectionsIsAnnotated(self):
+    coco_evaluator = coco_evaluation.CocoDetectionEvaluator(
+        _get_categories_list())
+    image_id = tf.placeholder(tf.string, shape=())
+    groundtruth_boxes = tf.placeholder(tf.float32, shape=(None, 4))
+    groundtruth_classes = tf.placeholder(tf.float32, shape=(None))
+    is_annotated = tf.placeholder(tf.bool, shape=())
+    detection_boxes = tf.placeholder(tf.float32, shape=(None, 4))
+    detection_scores = tf.placeholder(tf.float32, shape=(None))
+    detection_classes = tf.placeholder(tf.float32, shape=(None))
+
+    input_data_fields = standard_fields.InputDataFields
+    detection_fields = standard_fields.DetectionResultFields
+    eval_dict = {
+        input_data_fields.key: image_id,
+        input_data_fields.groundtruth_boxes: groundtruth_boxes,
+        input_data_fields.groundtruth_classes: groundtruth_classes,
+        'is_annotated': is_annotated,
+        detection_fields.detection_boxes: detection_boxes,
+        detection_fields.detection_scores: detection_scores,
+        detection_fields.detection_classes: detection_classes
+    }
+
+    eval_metric_ops = coco_evaluator.get_estimator_eval_metric_ops(eval_dict)
+
+    _, update_op = eval_metric_ops['DetectionBoxes_Precision/mAP']
+
+    with self.test_session() as sess:
+      sess.run(update_op,
+               feed_dict={
+                   image_id: 'image1',
+                   groundtruth_boxes: np.array([[100., 100., 200., 200.]]),
+                   groundtruth_classes: np.array([1]),
+                   is_annotated: True,
+                   detection_boxes: np.array([[100., 100., 200., 200.]]),
+                   detection_scores: np.array([.8]),
+                   detection_classes: np.array([1])
+               })
+      sess.run(update_op,
+               feed_dict={
+                   image_id: 'image2',
+                   groundtruth_boxes: np.array([[50., 50., 100., 100.]]),
+                   groundtruth_classes: np.array([3]),
+                   is_annotated: True,
+                   detection_boxes: np.array([[50., 50., 100., 100.]]),
+                   detection_scores: np.array([.7]),
+                   detection_classes: np.array([3])
+               })
+      sess.run(update_op,
+               feed_dict={
+                   image_id: 'image3',
+                   groundtruth_boxes: np.array([[25., 25., 50., 50.]]),
+                   groundtruth_classes: np.array([2]),
+                   is_annotated: True,
+                   detection_boxes: np.array([[25., 25., 50., 50.]]),
+                   detection_scores: np.array([.9]),
+                   detection_classes: np.array([2])
+               })
+      sess.run(update_op,
+               feed_dict={
+                   image_id: 'image4',
+                   groundtruth_boxes: np.zeros((0, 4)),
+                   groundtruth_classes: np.zeros((0)),
+                   is_annotated: False,  # Note that this image isn't annotated.
+                   detection_boxes: np.array([[25., 25., 50., 50.],
+                                              [25., 25., 70., 50.],
+                                              [25., 25., 80., 50.],
+                                              [25., 25., 90., 50.]]),
+                   detection_scores: np.array([0.6, 0.7, 0.8, 0.9]),
+                   detection_classes: np.array([1, 2, 2, 3])
+               })
+    metrics = {}
+    for key, (value_op, _) in eval_metric_ops.iteritems():
+      metrics[key] = value_op
+    metrics = sess.run(metrics)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP@.50IOU'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP@.75IOU'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP (large)'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP (medium)'],
+                           1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Precision/mAP (small)'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Recall/AR@1'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Recall/AR@10'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Recall/AR@100'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Recall/AR@100 (large)'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Recall/AR@100 (medium)'],
+                           1.0)
+    self.assertAlmostEqual(metrics['DetectionBoxes_Recall/AR@100 (small)'], 1.0)
+    self.assertFalse(coco_evaluator._groundtruth_list)
+    self.assertFalse(coco_evaluator._detection_boxes_list)
+    self.assertFalse(coco_evaluator._image_ids)
+
   def testGetOneMAPWithMatchingGroundtruthAndDetectionsPadded(self):
     coco_evaluator = coco_evaluation.CocoDetectionEvaluator(
         _get_categories_list())
@@ -665,22 +758,40 @@ class CocoMaskEvaluationPyFuncTest(tf.test.TestCase):
     _, update_op = eval_metric_ops['DetectionMasks_Precision/mAP']
 
     with self.test_session() as sess:
-      sess.run(update_op,
-               feed_dict={
-                   image_id: 'image1',
-                   groundtruth_boxes: np.array([[100., 100., 200., 200.]]),
-                   groundtruth_classes: np.array([1]),
-                   groundtruth_masks: np.pad(np.ones([1, 100, 100],
-                                                     dtype=np.uint8),
-                                             ((0, 0), (10, 10), (10, 10)),
-                                             mode='constant'),
-                   detection_scores: np.array([.8]),
-                   detection_classes: np.array([1]),
-                   detection_masks: np.pad(np.ones([1, 100, 100],
-                                                   dtype=np.uint8),
-                                           ((0, 0), (10, 10), (10, 10)),
-                                           mode='constant')
-               })
+      sess.run(
+          update_op,
+          feed_dict={
+              image_id:
+                  'image1',
+              groundtruth_boxes:
+                  np.array([[100., 100., 200., 200.], [50., 50., 100., 100.]]),
+              groundtruth_classes:
+                  np.array([1, 2]),
+              groundtruth_masks:
+                  np.stack([
+                      np.pad(
+                          np.ones([100, 100], dtype=np.uint8), ((10, 10),
+                                                                (10, 10)),
+                          mode='constant'),
+                      np.pad(
+                          np.ones([50, 50], dtype=np.uint8), ((0, 70), (0, 70)),
+                          mode='constant')
+                  ]),
+              detection_scores:
+                  np.array([.9, .8]),
+              detection_classes:
+                  np.array([2, 1]),
+              detection_masks:
+                  np.stack([
+                      np.pad(
+                          np.ones([50, 50], dtype=np.uint8), ((0, 70), (0, 70)),
+                          mode='constant'),
+                      np.pad(
+                          np.ones([100, 100], dtype=np.uint8), ((10, 10),
+                                                                (10, 10)),
+                          mode='constant'),
+                  ])
+          })
       sess.run(update_op,
                feed_dict={
                    image_id: 'image2',
@@ -735,6 +846,106 @@ class CocoMaskEvaluationPyFuncTest(tf.test.TestCase):
     self.assertFalse(coco_evaluator._image_id_to_mask_shape_map)
     self.assertFalse(coco_evaluator._detection_masks_list)
 
+  def testGetOneMAPWithMatchingGroundtruthAndDetectionsBatched(self):
+    coco_evaluator = coco_evaluation.CocoMaskEvaluator(_get_categories_list())
+    batch_size = 3
+    image_id = tf.placeholder(tf.string, shape=(batch_size))
+    groundtruth_boxes = tf.placeholder(tf.float32, shape=(batch_size, None, 4))
+    groundtruth_classes = tf.placeholder(tf.float32, shape=(batch_size, None))
+    groundtruth_masks = tf.placeholder(
+        tf.uint8, shape=(batch_size, None, None, None))
+    detection_scores = tf.placeholder(tf.float32, shape=(batch_size, None))
+    detection_classes = tf.placeholder(tf.float32, shape=(batch_size, None))
+    detection_masks = tf.placeholder(
+        tf.uint8, shape=(batch_size, None, None, None))
+
+    input_data_fields = standard_fields.InputDataFields
+    detection_fields = standard_fields.DetectionResultFields
+    eval_dict = {
+        input_data_fields.key: image_id,
+        input_data_fields.groundtruth_boxes: groundtruth_boxes,
+        input_data_fields.groundtruth_classes: groundtruth_classes,
+        input_data_fields.groundtruth_instance_masks: groundtruth_masks,
+        detection_fields.detection_scores: detection_scores,
+        detection_fields.detection_classes: detection_classes,
+        detection_fields.detection_masks: detection_masks,
+    }
+
+    eval_metric_ops = coco_evaluator.get_estimator_eval_metric_ops(eval_dict)
+
+    _, update_op = eval_metric_ops['DetectionMasks_Precision/mAP']
+
+    with self.test_session() as sess:
+      sess.run(
+          update_op,
+          feed_dict={
+              image_id: ['image1', 'image2', 'image3'],
+              groundtruth_boxes:
+                  np.array([[[100., 100., 200., 200.]],
+                            [[50., 50., 100., 100.]],
+                            [[25., 25., 50., 50.]]]),
+              groundtruth_classes:
+                  np.array([[1], [1], [1]]),
+              groundtruth_masks:
+                  np.stack([
+                      np.pad(
+                          np.ones([1, 100, 100], dtype=np.uint8),
+                          ((0, 0), (0, 0), (0, 0)),
+                          mode='constant'),
+                      np.pad(
+                          np.ones([1, 50, 50], dtype=np.uint8),
+                          ((0, 0), (25, 25), (25, 25)),
+                          mode='constant'),
+                      np.pad(
+                          np.ones([1, 25, 25], dtype=np.uint8),
+                          ((0, 0), (37, 38), (37, 38)),
+                          mode='constant')
+                  ],
+                           axis=0),
+              detection_scores:
+                  np.array([[.8], [.8], [.8]]),
+              detection_classes:
+                  np.array([[1], [1], [1]]),
+              detection_masks:
+                  np.stack([
+                      np.pad(
+                          np.ones([1, 100, 100], dtype=np.uint8),
+                          ((0, 0), (0, 0), (0, 0)),
+                          mode='constant'),
+                      np.pad(
+                          np.ones([1, 50, 50], dtype=np.uint8),
+                          ((0, 0), (25, 25), (25, 25)),
+                          mode='constant'),
+                      np.pad(
+                          np.ones([1, 25, 25], dtype=np.uint8),
+                          ((0, 0), (37, 38), (37, 38)),
+                          mode='constant')
+                  ],
+                           axis=0)
+          })
+    metrics = {}
+    for key, (value_op, _) in eval_metric_ops.iteritems():
+      metrics[key] = value_op
+    metrics = sess.run(metrics)
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP@.50IOU'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP@.75IOU'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP (large)'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP (medium)'],
+                           1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Precision/mAP (small)'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Recall/AR@1'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Recall/AR@10'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Recall/AR@100'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Recall/AR@100 (large)'], 1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Recall/AR@100 (medium)'],
+                           1.0)
+    self.assertAlmostEqual(metrics['DetectionMasks_Recall/AR@100 (small)'], 1.0)
+    self.assertFalse(coco_evaluator._groundtruth_list)
+    self.assertFalse(coco_evaluator._image_ids_with_detections)
+    self.assertFalse(coco_evaluator._image_id_to_mask_shape_map)
+    self.assertFalse(coco_evaluator._detection_masks_list)
+
 
 if __name__ == '__main__':
   tf.test.main()

research/object_detection/model_lib.py

@@ -25,6 +25,7 @@ import os
 import tensorflow as tf
 
 from object_detection import eval_util
+from object_detection import exporter as exporter_lib
 from object_detection import inputs
 from object_detection.builders import graph_rewriter_builder
 from object_detection.builders import model_builder
@@ -306,8 +307,7 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False):
           prediction_dict, features[fields.InputDataFields.true_image_shape])
       losses = [loss_tensor for loss_tensor in losses_dict.values()]
       if train_config.add_regularization_loss:
-        regularization_losses = tf.get_collection(
-            tf.GraphKeys.REGULARIZATION_LOSSES)
+        regularization_losses = detection_model.regularization_losses()
         if regularization_losses:
           regularization_loss = tf.add_n(
               regularization_losses, name='regularization_loss')
@@ -353,20 +353,24 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False):
         for var in optimizer_summary_vars:
           tf.summary.scalar(var.op.name, var)
       summaries = [] if use_tpu else None
+      if train_config.summarize_gradients:
+        summaries = ['gradients', 'gradient_norm', 'global_gradient_norm']
       train_op = tf.contrib.layers.optimize_loss(
           loss=total_loss,
           global_step=global_step,
           learning_rate=None,
           clip_gradients=clip_gradients_value,
           optimizer=training_optimizer,
+          update_ops=detection_model.updates(),
           variables=trainable_variables,
           summaries=summaries,
           name='')  # Preventing scope prefix on all variables.
 
     if mode == tf.estimator.ModeKeys.PREDICT:
+      exported_output = exporter_lib.add_output_tensor_nodes(detections)
       export_outputs = {
           tf.saved_model.signature_constants.PREDICT_METHOD_NAME:
-              tf.estimator.export.PredictOutput(detections)
+              tf.estimator.export.PredictOutput(exported_output)
       }
 
     eval_metric_ops = None
@@ -456,6 +460,7 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False):
 def create_estimator_and_inputs(run_config,
                                 hparams,
                                 pipeline_config_path,
+                                config_override=None,
                                 train_steps=None,
                                 sample_1_of_n_eval_examples=1,
                                 sample_1_of_n_eval_on_train_examples=1,
@@ -465,6 +470,7 @@ def create_estimator_and_inputs(run_config,
                                 num_shards=1,
                                 params=None,
                                 override_eval_num_epochs=True,
+                                save_final_config=False,
                                 **kwargs):
   """Creates `Estimator`, input functions, and steps.
 
@@ -472,6 +478,8 @@ def create_estimator_and_inputs(run_config,
     run_config: A `RunConfig`.
     hparams: A `HParams`.
     pipeline_config_path: A path to a pipeline config file.
+    config_override: A pipeline_pb2.TrainEvalPipelineConfig text proto to
+      override the config from `pipeline_config_path`.
     train_steps: Number of training steps. If None, the number of training steps
       is set from the `TrainConfig` proto.
     sample_1_of_n_eval_examples: Integer representing how often an eval example
@@ -499,6 +507,8 @@ def create_estimator_and_inputs(run_config,
       `use_tpu_estimator` is True.
     override_eval_num_epochs: Whether to overwrite the number of epochs to
       1 for eval_input.
+    save_final_config: Whether to save final config (obtained after applying
+      overrides) to `estimator.model_dir`.
     **kwargs: Additional keyword arguments for configuration override.
 
   Returns:
@@ -522,7 +532,8 @@ def create_estimator_and_inputs(run_config,
   create_eval_input_fn = MODEL_BUILD_UTIL_MAP['create_eval_input_fn']
   create_predict_input_fn = MODEL_BUILD_UTIL_MAP['create_predict_input_fn']
 
-  configs = get_configs_from_pipeline_file(pipeline_config_path)
+  configs = get_configs_from_pipeline_file(pipeline_config_path,
+                                           config_override=config_override)
   kwargs.update({
       'train_steps': train_steps,
       'sample_1_of_n_eval_examples': sample_1_of_n_eval_examples
@@ -595,7 +606,7 @@ def create_estimator_and_inputs(run_config,
     estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config)
 
   # Write the as-run pipeline config to disk.
-  if run_config.is_chief:
+  if run_config.is_chief and save_final_config:
     pipeline_config_final = create_pipeline_proto_from_configs(configs)
     config_util.save_pipeline_config(pipeline_config_final, estimator.model_dir)
 
@@ -641,11 +652,17 @@ def create_train_and_eval_specs(train_input_fn,
       input_fn=train_input_fn, max_steps=train_steps)
 
   if eval_spec_names is None:
-    eval_spec_names = [ str(i) for i in range(len(eval_input_fns)) ]
+    eval_spec_names = [str(i) for i in range(len(eval_input_fns))]
 
   eval_specs = []
-  for eval_spec_name, eval_input_fn in zip(eval_spec_names, eval_input_fns):
-    exporter_name = '{}_{}'.format(final_exporter_name, eval_spec_name)
+  for index, (eval_spec_name, eval_input_fn) in enumerate(
+      zip(eval_spec_names, eval_input_fns)):
+    # Uses final_exporter_name as exporter_name for the first eval spec for
+    # backward compatibility.
+    if index == 0:
+      exporter_name = final_exporter_name
+    else:
+      exporter_name = '{}_{}'.format(final_exporter_name, eval_spec_name)
     exporter = tf.estimator.FinalExporter(
         name=exporter_name, serving_input_receiver_fn=predict_input_fn)
     eval_specs.append(
@@ -747,6 +764,7 @@ def populate_experiment(run_config,
       train_steps=train_steps,
       eval_steps=eval_steps,
       model_fn_creator=model_fn_creator,
+      save_final_config=True,
       **kwargs)
   estimator = train_and_eval_dict['estimator']
   train_input_fn = train_and_eval_dict['train_input_fn']

research/object_detection/model_lib_test.py

@@ -310,7 +310,7 @@ class ModelLibTest(tf.test.TestCase):
     self.assertEqual(2, len(eval_specs))
     self.assertEqual(None, eval_specs[0].steps)
     self.assertEqual('holdout', eval_specs[0].name)
-    self.assertEqual('exporter_holdout', eval_specs[0].exporters[0].name)
+    self.assertEqual('exporter', eval_specs[0].exporters[0].name)
     self.assertEqual(None, eval_specs[1].steps)
     self.assertEqual('eval_on_train', eval_specs[1].name)
 

research/object_detection/model_tpu_main.py

@@ -114,6 +114,7 @@ def main(unused_argv):
       use_tpu_estimator=True,
       use_tpu=FLAGS.use_tpu,
       num_shards=FLAGS.num_shards,
+      save_final_config=FLAGS.mode == 'train',
       **kwargs)
   estimator = train_and_eval_dict['estimator']
   train_input_fn = train_and_eval_dict['train_input_fn']

research/object_detection/models/faster_rcnn_resnet_v1_feature_extractor.py

@@ -72,6 +72,8 @@ class FasterRCNNResnetV1FeatureExtractor(
 
     VGG style channel mean subtraction as described here:
     https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md
+    Note that if the number of channels is not equal to 3, the mean subtraction
+    will be skipped and the original resized_inputs will be returned.
 
     Args:
       resized_inputs: A [batch, height_in, width_in, channels] float32 tensor
@@ -82,8 +84,11 @@ class FasterRCNNResnetV1FeatureExtractor(
         tensor representing a batch of images.
 
     """
-    channel_means = [123.68, 116.779, 103.939]
-    return resized_inputs - [[channel_means]]
+    if resized_inputs.shape.as_list()[3] == 3:
+      channel_means = [123.68, 116.779, 103.939]
+      return resized_inputs - [[channel_means]]
+    else:
+      return resized_inputs
 
   def _extract_proposal_features(self, preprocessed_inputs, scope):
     """Extracts first stage RPN features.

research/object_detection/models/feature_map_generators.py

@@ -146,7 +146,6 @@ class KerasMultiResolutionFeatureMaps(tf.keras.Model):
       use_depthwise = feature_map_layout['use_depthwise']
     for index, from_layer in enumerate(feature_map_layout['from_layer']):
       net = []
-      self.convolutions.append(net)
       layer_depth = feature_map_layout['layer_depth'][index]
       conv_kernel_size = 3
       if 'conv_kernel_size' in feature_map_layout:
@@ -231,6 +230,10 @@ class KerasMultiResolutionFeatureMaps(tf.keras.Model):
               conv_hyperparams.build_activation_layer(
                   name=layer_name))
 
+      # Until certain bugs are fixed in checkpointable lists,
+      # this net must be appended only once it's been filled with layers
+      self.convolutions.append(net)
+
   def call(self, image_features):
     """Generate the multi-resolution feature maps.
 
@@ -263,7 +266,8 @@ class KerasMultiResolutionFeatureMaps(tf.keras.Model):
 
 
 def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
-                                  min_depth, insert_1x1_conv, image_features):
+                                  min_depth, insert_1x1_conv, image_features,
+                                  pool_residual=False):
   """Generates multi resolution feature maps from input image features.
 
   Generates multi-scale feature maps for detection as in the SSD papers by
@@ -317,6 +321,13 @@ def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
       should be inserted before shrinking the feature map.
     image_features: A dictionary of handles to activation tensors from the
       base feature extractor.
+    pool_residual: Whether to add an average pooling layer followed by a
+      residual connection between subsequent feature maps when the channel
+      depth match. For example, with option 'layer_depth': [-1, 512, 256, 256],
+      a pooling and residual layer is added between the third and forth feature
+      map. This option is better used with Weight Shared Convolution Box
+      Predictor when all feature maps have the same channel depth to encourage
+      more consistent features across multi-scale feature maps.
 
   Returns:
     feature_maps: an OrderedDict mapping keys (feature map names) to
@@ -350,6 +361,7 @@ def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
       feature_map_keys.append(from_layer)
     else:
       pre_layer = feature_maps[-1]
+      pre_layer_depth = pre_layer.get_shape().as_list()[3]
       intermediate_layer = pre_layer
       if insert_1x1_conv:
         layer_name = '{}_1_Conv2d_{}_1x1_{}'.format(
@@ -383,6 +395,12 @@ def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
             padding='SAME',
             stride=1,
             scope=layer_name)
+        if pool_residual and pre_layer_depth == depth_fn(layer_depth):
+          feature_map += slim.avg_pool2d(
+              pre_layer, [3, 3],
+              padding='SAME',
+              stride=2,
+              scope=layer_name + '_pool')
       else:
         feature_map = slim.conv2d(
             intermediate_layer,
@@ -399,6 +417,7 @@ def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
 def fpn_top_down_feature_maps(image_features,
                               depth,
                               use_depthwise=False,
+                              use_explicit_padding=False,
                               scope=None):
   """Generates `top-down` feature maps for Feature Pyramid Networks.
 
@@ -409,7 +428,9 @@ def fpn_top_down_feature_maps(image_features,
       Spatial resolutions of succesive tensors must reduce exactly by a factor
       of 2.
     depth: depth of output feature maps.
-    use_depthwise: use depthwise separable conv instead of regular conv.
+    use_depthwise: whether to use depthwise separable conv instead of regular
+      conv.
+    use_explicit_padding: whether to use explicit padding.
     scope: A scope name to wrap this op under.
 
   Returns:
@@ -420,8 +441,10 @@ def fpn_top_down_feature_maps(image_features,
     num_levels = len(image_features)
     output_feature_maps_list = []
     output_feature_map_keys = []
+    padding = 'VALID' if use_explicit_padding else 'SAME'
+    kernel_size = 3
     with slim.arg_scope(
-        [slim.conv2d, slim.separable_conv2d], padding='SAME', stride=1):
+        [slim.conv2d, slim.separable_conv2d], padding=padding, stride=1):
       top_down = slim.conv2d(
           image_features[-1][1],
           depth, [1, 1], activation_fn=None, normalizer_fn=None,
@@ -436,14 +459,20 @@ def fpn_top_down_feature_maps(image_features,
             image_features[level][1], depth, [1, 1],
             activation_fn=None, normalizer_fn=None,
             scope='projection_%d' % (level + 1))
+        if use_explicit_padding:
+          # slice top_down to the same shape as residual
+          residual_shape = tf.shape(residual)
+          top_down = top_down[:, :residual_shape[1], :residual_shape[2], :]
         top_down += residual
         if use_depthwise:
           conv_op = functools.partial(slim.separable_conv2d, depth_multiplier=1)
         else:
           conv_op = slim.conv2d
+        if use_explicit_padding:
+          top_down = ops.fixed_padding(top_down, kernel_size)
         output_feature_maps_list.append(conv_op(
             top_down,
-            depth, [3, 3],
+            depth, [kernel_size, kernel_size],
             scope='smoothing_%d' % (level + 1)))
         output_feature_map_keys.append('top_down_%s' % image_features[level][0])
       return collections.OrderedDict(reversed(

research/object_detection/models/feature_map_generators_test.py

@@ -45,6 +45,11 @@ EMBEDDED_SSD_MOBILENET_V1_LAYOUT = {
     'conv_kernel_size': [-1, -1, 3, 3, 2],
 }
 
+SSD_MOBILENET_V1_WEIGHT_SHARED_LAYOUT = {
+    'from_layer': ['Conv2d_13_pointwise', '', '', ''],
+    'layer_depth': [-1, 256, 256, 256],
+}
+
 
 @parameterized.parameters(
     {'use_keras': False},
@@ -67,7 +72,8 @@ class MultiResolutionFeatureMapGeneratorTest(tf.test.TestCase):
     text_format.Merge(conv_hyperparams_text_proto, conv_hyperparams)
     return hyperparams_builder.KerasLayerHyperparams(conv_hyperparams)
 
-  def _build_feature_map_generator(self, feature_map_layout, use_keras):
+  def _build_feature_map_generator(self, feature_map_layout, use_keras,
+                                   pool_residual=False):
     if use_keras:
       return feature_map_generators.KerasMultiResolutionFeatureMaps(
           feature_map_layout=feature_map_layout,
@@ -86,7 +92,8 @@ class MultiResolutionFeatureMapGeneratorTest(tf.test.TestCase):
             depth_multiplier=1,
             min_depth=32,
             insert_1x1_conv=True,
-            image_features=image_features)
+            image_features=image_features,
+            pool_residual=pool_residual)
       return feature_map_generator
 
   def test_get_expected_feature_map_shapes_with_inception_v2(self, use_keras):
@@ -209,6 +216,34 @@ class MultiResolutionFeatureMapGeneratorTest(tf.test.TestCase):
           (key, value.shape) for key, value in out_feature_maps.items())
       self.assertDictEqual(expected_feature_map_shapes, out_feature_map_shapes)
 
+  def test_feature_map_shapes_with_pool_residual_ssd_mobilenet_v1(
+      self, use_keras):
+    image_features = {
+        'Conv2d_13_pointwise': tf.random_uniform([4, 8, 8, 1024],
+                                                 dtype=tf.float32),
+    }
+
+    feature_map_generator = self._build_feature_map_generator(
+        feature_map_layout=SSD_MOBILENET_V1_WEIGHT_SHARED_LAYOUT,
+        use_keras=use_keras,
+        pool_residual=True
+    )
+    feature_maps = feature_map_generator(image_features)
+
+    expected_feature_map_shapes = {
+        'Conv2d_13_pointwise': (4, 8, 8, 1024),
+        'Conv2d_13_pointwise_2_Conv2d_1_3x3_s2_256': (4, 4, 4, 256),
+        'Conv2d_13_pointwise_2_Conv2d_2_3x3_s2_256': (4, 2, 2, 256),
+        'Conv2d_13_pointwise_2_Conv2d_3_3x3_s2_256': (4, 1, 1, 256)}
+
+    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 = dict(
+          (key, value.shape) for key, value in out_feature_maps.items())
+      self.assertDictEqual(expected_feature_map_shapes, out_feature_map_shapes)
+
   def test_get_expected_variable_names_with_inception_v2(self, use_keras):
     image_features = {
         'Mixed_3c': tf.random_uniform([4, 28, 28, 256], dtype=tf.float32),

research/object_detection/models/keras_applications/mobilenet_v2.py

@@ -82,6 +82,8 @@ class _LayersOverride(object):
     self._conv_hyperparams = conv_hyperparams
     self._use_explicit_padding = use_explicit_padding
     self._min_depth = min_depth
+    self.regularizer = tf.keras.regularizers.l2(0.00004 * 0.5)
+    self.initializer = tf.truncated_normal_initializer(stddev=0.09)
 
   def _FixedPaddingLayer(self, kernel_size):
     return tf.keras.layers.Lambda(lambda x: ops.fixed_padding(x, kernel_size))
@@ -114,6 +116,9 @@ class _LayersOverride(object):
 
     if self._conv_hyperparams:
       kwargs = self._conv_hyperparams.params(**kwargs)
+    else:
+      kwargs['kernel_regularizer'] = self.regularizer
+      kwargs['kernel_initializer'] = self.initializer
 
     kwargs['padding'] = 'same'
     kernel_size = kwargs.get('kernel_size')
@@ -144,6 +149,8 @@ class _LayersOverride(object):
     """
     if self._conv_hyperparams:
       kwargs = self._conv_hyperparams.params(**kwargs)
+    else:
+      kwargs['depthwise_initializer'] = self.initializer
 
     kwargs['padding'] = 'same'
     kernel_size = kwargs.get('kernel_size')

research/object_detection/models/ssd_mobilenet_v1_fpn_feature_extractor.py

@@ -31,11 +31,10 @@ slim = tf.contrib.slim
 
 # A modified config of mobilenet v1 that makes it more detection friendly,
 def _create_modified_mobilenet_config():
-  conv_defs = copy.copy(mobilenet_v1.MOBILENETV1_CONV_DEFS)
+  conv_defs = copy.deepcopy(mobilenet_v1.MOBILENETV1_CONV_DEFS)
   conv_defs[-2] = mobilenet_v1.DepthSepConv(kernel=[3, 3], stride=2, depth=512)
   conv_defs[-1] = mobilenet_v1.DepthSepConv(kernel=[3, 3], stride=1, depth=256)
   return conv_defs
-_CONV_DEFS = _create_modified_mobilenet_config()
 
 
 class SSDMobileNetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
@@ -98,6 +97,9 @@ class SSDMobileNetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     self._fpn_min_level = fpn_min_level
     self._fpn_max_level = fpn_max_level
     self._additional_layer_depth = additional_layer_depth
+    self._conv_defs = None
+    if self._use_depthwise:
+      self._conv_defs = _create_modified_mobilenet_config()
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -141,7 +143,7 @@ class SSDMobileNetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
               final_endpoint='Conv2d_13_pointwise',
               min_depth=self._min_depth,
               depth_multiplier=self._depth_multiplier,
-              conv_defs=_CONV_DEFS if self._use_depthwise else None,
+              conv_defs=self._conv_defs,
               use_explicit_padding=self._use_explicit_padding,
               scope=scope)
 
@@ -159,7 +161,8 @@ class SSDMobileNetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
           fpn_features = feature_map_generators.fpn_top_down_feature_maps(
               [(key, image_features[key]) for key in feature_block_list],
               depth=depth_fn(self._additional_layer_depth),
-              use_depthwise=self._use_depthwise)
+              use_depthwise=self._use_depthwise,
+              use_explicit_padding=self._use_explicit_padding)
           feature_maps = []
           for level in range(self._fpn_min_level, base_fpn_max_level + 1):
             feature_maps.append(fpn_features['top_down_{}'.format(
@@ -167,18 +170,23 @@ class SSDMobileNetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
           last_feature_map = fpn_features['top_down_{}'.format(
               feature_blocks[base_fpn_max_level - 2])]
           # Construct coarse features
+          padding = 'VALID' if self._use_explicit_padding else 'SAME'
+          kernel_size = 3
           for i in range(base_fpn_max_level + 1, self._fpn_max_level + 1):
             if self._use_depthwise:
               conv_op = functools.partial(
                   slim.separable_conv2d, depth_multiplier=1)
             else:
               conv_op = slim.conv2d
+            if self._use_explicit_padding:
+              last_feature_map = ops.fixed_padding(
+                  last_feature_map, kernel_size)
             last_feature_map = conv_op(
                 last_feature_map,
                 num_outputs=depth_fn(self._additional_layer_depth),
-                kernel_size=[3, 3],
+                kernel_size=[kernel_size, kernel_size],
                 stride=2,
-                padding='SAME',
+                padding=padding,
                 scope='bottom_up_Conv2d_{}'.format(i - base_fpn_max_level + 13))
             feature_maps.append(last_feature_map)
     return feature_maps

research/object_detection/models/ssd_mobilenet_v2_fpn_feature_extractor.py

@@ -30,17 +30,14 @@ from nets.mobilenet import mobilenet_v2
 slim = tf.contrib.slim
 
 
-# A modified config of mobilenet v2 that makes it more detection friendly,
+# A modified config of mobilenet v2 that makes it more detection friendly.
 def _create_modified_mobilenet_config():
-  conv_defs = copy.copy(mobilenet_v2.V2_DEF)
+  conv_defs = copy.deepcopy(mobilenet_v2.V2_DEF)
   conv_defs['spec'][-1] = mobilenet.op(
       slim.conv2d, stride=1, kernel_size=[1, 1], num_outputs=256)
   return conv_defs
 
 
-_CONV_DEFS = _create_modified_mobilenet_config()
-
-
 class SSDMobileNetV2FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
   """SSD Feature Extractor using MobilenetV2 FPN features."""
 
@@ -100,6 +97,9 @@ class SSDMobileNetV2FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
     self._fpn_min_level = fpn_min_level
     self._fpn_max_level = fpn_max_level
     self._additional_layer_depth = additional_layer_depth
+    self._conv_defs = None
+    if self._use_depthwise:
+      self._conv_defs = _create_modified_mobilenet_config()
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.
@@ -142,7 +142,7 @@ class SSDMobileNetV2FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
               ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
               final_endpoint='layer_19',
               depth_multiplier=self._depth_multiplier,
-              conv_defs=_CONV_DEFS if self._use_depthwise else None,
+              conv_defs=self._conv_defs,
               use_explicit_padding=self._use_explicit_padding,
               scope=scope)
       depth_fn = lambda d: max(int(d * self._depth_multiplier), self._min_depth)
@@ -158,7 +158,8 @@ class SSDMobileNetV2FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
           fpn_features = feature_map_generators.fpn_top_down_feature_maps(
               [(key, image_features[key]) for key in feature_block_list],
               depth=depth_fn(self._additional_layer_depth),
-              use_depthwise=self._use_depthwise)
+              use_depthwise=self._use_depthwise,
+              use_explicit_padding=self._use_explicit_padding)
           feature_maps = []
           for level in range(self._fpn_min_level, base_fpn_max_level + 1):
             feature_maps.append(fpn_features['top_down_{}'.format(
@@ -166,18 +167,23 @@ class SSDMobileNetV2FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
           last_feature_map = fpn_features['top_down_{}'.format(
               feature_blocks[base_fpn_max_level - 2])]
           # Construct coarse features
+          padding = 'VALID' if self._use_explicit_padding else 'SAME'
+          kernel_size = 3
           for i in range(base_fpn_max_level + 1, self._fpn_max_level + 1):
             if self._use_depthwise:
               conv_op = functools.partial(
                   slim.separable_conv2d, depth_multiplier=1)
             else:
               conv_op = slim.conv2d
+            if self._use_explicit_padding:
+              last_feature_map = ops.fixed_padding(
+                  last_feature_map, kernel_size)
             last_feature_map = conv_op(
                 last_feature_map,
                 num_outputs=depth_fn(self._additional_layer_depth),
-                kernel_size=[3, 3],
+                kernel_size=[kernel_size, kernel_size],
                 stride=2,
-                padding='SAME',
+                padding=padding,
                 scope='bottom_up_Conv2d_{}'.format(i - base_fpn_max_level + 19))
             feature_maps.append(last_feature_map)
     return feature_maps

research/object_detection/models/ssd_mobilenet_v2_keras_feature_extractor.py

@@ -85,41 +85,44 @@ class SSDMobileNetV2KerasFeatureExtractor(
         override_base_feature_extractor_hyperparams=
         override_base_feature_extractor_hyperparams,
         name=name)
-    feature_map_layout = {
+    self._feature_map_layout = {
         'from_layer': ['layer_15/expansion_output', 'layer_19', '', '', '', ''],
         'layer_depth': [-1, -1, 512, 256, 256, 128],
         'use_depthwise': self._use_depthwise,
         'use_explicit_padding': self._use_explicit_padding,
     }
 
-    with tf.name_scope('MobilenetV2'):
-      full_mobilenet_v2 = mobilenet_v2.mobilenet_v2(
-          batchnorm_training=(is_training and not freeze_batchnorm),
-          conv_hyperparams=(conv_hyperparams
-                            if self._override_base_feature_extractor_hyperparams
-                            else None),
-          weights=None,
-          use_explicit_padding=use_explicit_padding,
-          alpha=self._depth_multiplier,
-          min_depth=self._min_depth,
-          include_top=False)
-      conv2d_11_pointwise = full_mobilenet_v2.get_layer(
-          name='block_13_expand_relu').output
-      conv2d_13_pointwise = full_mobilenet_v2.get_layer(name='out_relu').output
-      self.mobilenet_v2 = tf.keras.Model(
-          inputs=full_mobilenet_v2.inputs,
-          outputs=[conv2d_11_pointwise, conv2d_13_pointwise])
-
-      self.feature_map_generator = (
-          feature_map_generators.KerasMultiResolutionFeatureMaps(
-              feature_map_layout=feature_map_layout,
-              depth_multiplier=self._depth_multiplier,
-              min_depth=self._min_depth,
-              insert_1x1_conv=True,
-              is_training=is_training,
-              conv_hyperparams=conv_hyperparams,
-              freeze_batchnorm=freeze_batchnorm,
-              name='FeatureMaps'))
+    self.mobilenet_v2 = None
+    self.feature_map_generator = None
+
+  def build(self, input_shape):
+    full_mobilenet_v2 = mobilenet_v2.mobilenet_v2(
+        batchnorm_training=(self._is_training and not self._freeze_batchnorm),
+        conv_hyperparams=(self._conv_hyperparams
+                          if self._override_base_feature_extractor_hyperparams
+                          else None),
+        weights=None,
+        use_explicit_padding=self._use_explicit_padding,
+        alpha=self._depth_multiplier,
+        min_depth=self._min_depth,
+        include_top=False)
+    conv2d_11_pointwise = full_mobilenet_v2.get_layer(
+        name='block_13_expand_relu').output
+    conv2d_13_pointwise = full_mobilenet_v2.get_layer(name='out_relu').output
+    self.mobilenet_v2 = tf.keras.Model(
+        inputs=full_mobilenet_v2.inputs,
+        outputs=[conv2d_11_pointwise, conv2d_13_pointwise])
+    self.feature_map_generator = (
+        feature_map_generators.KerasMultiResolutionFeatureMaps(
+            feature_map_layout=self._feature_map_layout,
+            depth_multiplier=self._depth_multiplier,
+            min_depth=self._min_depth,
+            insert_1x1_conv=True,
+            is_training=self._is_training,
+            conv_hyperparams=self._conv_hyperparams,
+            freeze_batchnorm=self._freeze_batchnorm,
+            name='FeatureMaps'))
+    self.built = True
 
   def preprocess(self, resized_inputs):
     """SSD preprocessing.

research/object_detection/models/ssd_pnasnet_feature_extractor.py

+# 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.
+# ==============================================================================
+
+"""SSDFeatureExtractor for PNASNet features.
+
+Based on PNASNet ImageNet model: https://arxiv.org/abs/1712.00559
+"""
+
+import tensorflow as tf
+
+from object_detection.meta_architectures import ssd_meta_arch
+from object_detection.models import feature_map_generators
+from object_detection.utils import context_manager
+from object_detection.utils import ops
+from nets.nasnet import pnasnet
+
+slim = tf.contrib.slim
+
+
+def pnasnet_large_arg_scope_for_detection(is_batch_norm_training=False):
+  """Defines the default arg scope for the PNASNet Large for object detection.
+
+  This provides a small edit to switch batch norm training on and off.
+
+  Args:
+    is_batch_norm_training: Boolean indicating whether to train with batch norm.
+    Default is False.
+
+  Returns:
+    An `arg_scope` to use for the PNASNet Large Model.
+  """
+  imagenet_scope = pnasnet.pnasnet_large_arg_scope()
+  with slim.arg_scope(imagenet_scope):
+    with slim.arg_scope([slim.batch_norm],
+                        is_training=is_batch_norm_training) as sc:
+      return sc
+
+
+class SSDPNASNetFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
+  """SSD Feature Extractor using PNASNet features."""
+
+  def __init__(self,
+               is_training,
+               depth_multiplier,
+               min_depth,
+               pad_to_multiple,
+               conv_hyperparams_fn,
+               reuse_weights=None,
+               use_explicit_padding=False,
+               use_depthwise=False,
+               override_base_feature_extractor_hyperparams=False):
+    """PNASNet Feature Extractor for SSD Models.
+
+    Args:
+      is_training: whether the network is in training mode.
+      depth_multiplier: float depth multiplier for feature extractor.
+      min_depth: minimum feature extractor depth.
+      pad_to_multiple: the nearest multiple to zero pad the input height and
+        width dimensions to.
+      conv_hyperparams_fn: A function to construct tf slim arg_scope for conv2d
+        and separable_conv2d ops in the layers that are added on top of the
+        base feature extractor.
+      reuse_weights: Whether to reuse variables. Default is None.
+      use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
+        inputs so that the output dimensions are the same as if 'SAME' padding
+        were used.
+      use_depthwise: Whether to use depthwise convolutions.
+      override_base_feature_extractor_hyperparams: Whether to override
+        hyperparameters of the base feature extractor with the one from
+        `conv_hyperparams_fn`.
+    """
+    super(SSDPNASNetFeatureExtractor, self).__init__(
+        is_training=is_training,
+        depth_multiplier=depth_multiplier,
+        min_depth=min_depth,
+        pad_to_multiple=pad_to_multiple,
+        conv_hyperparams_fn=conv_hyperparams_fn,
+        reuse_weights=reuse_weights,
+        use_explicit_padding=use_explicit_padding,
+        use_depthwise=use_depthwise,
+        override_base_feature_extractor_hyperparams=
+        override_base_feature_extractor_hyperparams)
+
+  def preprocess(self, resized_inputs):
+    """SSD preprocessing.
+
+    Maps pixel values to the range [-1, 1].
+
+    Args:
+      resized_inputs: a [batch, height, width, channels] float tensor
+        representing a batch of images.
+
+    Returns:
+      preprocessed_inputs: a [batch, height, width, channels] float tensor
+        representing a batch of images.
+    """
+    return (2.0 / 255.0) * resized_inputs - 1.0
+
+  def extract_features(self, preprocessed_inputs):
+    """Extract features from preprocessed inputs.
+
+    Args:
+      preprocessed_inputs: a [batch, height, width, channels] float tensor
+        representing a batch of images.
+
+    Returns:
+      feature_maps: a list of tensors where the ith tensor has shape
+        [batch, height_i, width_i, depth_i]
+    """
+
+    feature_map_layout = {
+        'from_layer': ['Cell_7', 'Cell_11', '', '', '', ''],
+        'layer_depth': [-1, -1, 512, 256, 256, 128],
+        'use_explicit_padding': self._use_explicit_padding,
+        'use_depthwise': self._use_depthwise,
+    }
+
+    with slim.arg_scope(
+        pnasnet_large_arg_scope_for_detection(
+            is_batch_norm_training=self._is_training)):
+      with slim.arg_scope([slim.conv2d, slim.batch_norm, slim.separable_conv2d],
+                          reuse=self._reuse_weights):
+        with (slim.arg_scope(self._conv_hyperparams_fn())
+              if self._override_base_feature_extractor_hyperparams else
+              context_manager.IdentityContextManager()):
+          _, image_features = pnasnet.build_pnasnet_large(
+              ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
+              num_classes=None,
+              is_training=self._is_training,
+              final_endpoint='Cell_11')
+    with tf.variable_scope('SSD_feature_maps', reuse=self._reuse_weights):
+      with slim.arg_scope(self._conv_hyperparams_fn()):
+        feature_maps = feature_map_generators.multi_resolution_feature_maps(
+            feature_map_layout=feature_map_layout,
+            depth_multiplier=self._depth_multiplier,
+            min_depth=self._min_depth,
+            insert_1x1_conv=True,
+            image_features=image_features)
+
+    return feature_maps.values()
+
+  def restore_from_classification_checkpoint_fn(self, feature_extractor_scope):
+    """Returns a map of variables to load from a foreign checkpoint.
+
+    Note that this overrides the default implementation in
+    ssd_meta_arch.SSDFeatureExtractor which does not work for PNASNet
+    checkpoints.
+
+    Args:
+      feature_extractor_scope: A scope name for the first 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():
+      if variable.op.name.startswith(feature_extractor_scope):
+        var_name = variable.op.name.replace(feature_extractor_scope + '/', '')
+        var_name += '/ExponentialMovingAverage'
+        variables_to_restore[var_name] = variable
+    return variables_to_restore

research/object_detection/models/ssd_pnasnet_feature_extractor_test.py

+# 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 ssd_pnas_feature_extractor."""
+import numpy as np
+import tensorflow as tf
+
+from object_detection.models import ssd_feature_extractor_test
+from object_detection.models import ssd_pnasnet_feature_extractor
+
+slim = tf.contrib.slim
+
+
+class SsdPnasNetFeatureExtractorTest(
+    ssd_feature_extractor_test.SsdFeatureExtractorTestBase):
+
+  def _create_feature_extractor(self, depth_multiplier, pad_to_multiple,
+                                is_training=True, use_explicit_padding=False):
+    """Constructs a new feature extractor.
+
+    Args:
+      depth_multiplier: float depth multiplier for feature extractor
+      pad_to_multiple: the nearest multiple to zero pad the input height and
+        width dimensions to.
+      is_training: whether the network is in training mode.
+      use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
+        inputs so that the output dimensions are the same as if 'SAME' padding
+        were used.
+    Returns:
+      an ssd_meta_arch.SSDFeatureExtractor object.
+    """
+    min_depth = 32
+    return ssd_pnasnet_feature_extractor.SSDPNASNetFeatureExtractor(
+        is_training, depth_multiplier, min_depth, pad_to_multiple,
+        self.conv_hyperparams_fn,
+        use_explicit_padding=use_explicit_padding)
+
+  def test_extract_features_returns_correct_shapes_128(self):
+    image_height = 128
+    image_width = 128
+    depth_multiplier = 1.0
+    pad_to_multiple = 1
+    expected_feature_map_shape = [(2, 8, 8, 2160), (2, 4, 4, 4320),
+                                  (2, 2, 2, 512), (2, 1, 1, 256),
+                                  (2, 1, 1, 256), (2, 1, 1, 128)]
+    self.check_extract_features_returns_correct_shape(
+        2, image_height, image_width, depth_multiplier, pad_to_multiple,
+        expected_feature_map_shape)
+
+  def test_extract_features_returns_correct_shapes_299(self):
+    image_height = 299
+    image_width = 299
+    depth_multiplier = 1.0
+    pad_to_multiple = 1
+    expected_feature_map_shape = [(2, 19, 19, 2160), (2, 10, 10, 4320),
+                                  (2, 5, 5, 512), (2, 3, 3, 256),
+                                  (2, 2, 2, 256), (2, 1, 1, 128)]
+    self.check_extract_features_returns_correct_shape(
+        2, image_height, image_width, depth_multiplier, pad_to_multiple,
+        expected_feature_map_shape)
+
+  def test_preprocess_returns_correct_value_range(self):
+    image_height = 128
+    image_width = 128
+    depth_multiplier = 1
+    pad_to_multiple = 1
+    test_image = np.random.rand(2, image_height, image_width, 3)
+    feature_extractor = self._create_feature_extractor(depth_multiplier,
+                                                       pad_to_multiple)
+    preprocessed_image = feature_extractor.preprocess(test_image)
+    self.assertTrue(np.all(np.less_equal(np.abs(preprocessed_image), 1.0)))
+
+
+if __name__ == '__main__':
+  tf.test.main()

research/object_detection/models/ssd_resnet_v1_fpn_feature_extractor.py

@@ -113,6 +113,8 @@ class _SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
 
     VGG style channel mean subtraction as described here:
     https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-mdnge.
+    Note that if the number of channels is not equal to 3, the mean subtraction
+    will be skipped and the original resized_inputs will be returned.
 
     Args:
       resized_inputs: a [batch, height, width, channels] float tensor
@@ -122,8 +124,11 @@ class _SSDResnetV1FpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       preprocessed_inputs: a [batch, height, width, channels] float tensor
         representing a batch of images.
     """
-    channel_means = [123.68, 116.779, 103.939]
-    return resized_inputs - [[channel_means]]
+    if resized_inputs.shape.as_list()[3] == 3:
+      channel_means = [123.68, 116.779, 103.939]
+      return resized_inputs - [[channel_means]]
+    else:
+      return resized_inputs
 
   def _filter_features(self, image_features):
     # TODO(rathodv): Change resnet endpoint to strip scope prefixes instead

research/object_detection/models/ssd_resnet_v1_fpn_feature_extractor_testbase.py

@@ -82,12 +82,15 @@ class SSDResnetFPNFeatureExtractorTestBase(
     image_width = 128
     depth_multiplier = 1
     pad_to_multiple = 1
-    test_image = np.random.rand(4, image_height, image_width, 3)
+    test_image = tf.constant(np.random.rand(4, image_height, image_width, 3))
     feature_extractor = self._create_feature_extractor(depth_multiplier,
                                                        pad_to_multiple)
     preprocessed_image = feature_extractor.preprocess(test_image)
-    self.assertAllClose(preprocessed_image,
-                        test_image - [[123.68, 116.779, 103.939]])
+    with self.test_session() as sess:
+      test_image_out, preprocessed_image_out = sess.run(
+          [test_image, preprocessed_image])
+      self.assertAllClose(preprocessed_image_out,
+                          test_image_out - [[123.68, 116.779, 103.939]])
 
   def test_variables_only_created_in_scope(self):
     depth_multiplier = 1
@@ -103,5 +106,3 @@ class SSDResnetFPNFeatureExtractorTestBase(
         self.assertTrue(
             variable.name.startswith(self._resnet_scope_name())
             or variable.name.startswith(self._fpn_scope_name()))
-
-

research/object_detection/models/ssd_resnet_v1_ppn_feature_extractor.py

@@ -98,6 +98,8 @@ class _SSDResnetPpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
 
     VGG style channel mean subtraction as described here:
     https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-mdnge.
+    Note that if the number of channels is not equal to 3, the mean subtraction
+    will be skipped and the original resized_inputs will be returned.
 
     Args:
       resized_inputs: a [batch, height, width, channels] float tensor
@@ -107,8 +109,11 @@ class _SSDResnetPpnFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       preprocessed_inputs: a [batch, height, width, channels] float tensor
         representing a batch of images.
     """
-    channel_means = [123.68, 116.779, 103.939]
-    return resized_inputs - [[channel_means]]
+    if resized_inputs.shape.as_list()[3] == 3:
+      channel_means = [123.68, 116.779, 103.939]
+      return resized_inputs - [[channel_means]]
+    else:
+      return resized_inputs
 
   def extract_features(self, preprocessed_inputs):
     """Extract features from preprocessed inputs.