Refactor object detection box predictors and fix some issues with model_main. (#4965)

* Merged commit includes the following changes:
206852642  by Zhichao Lu:

    Build the balanced_positive_negative_sampler in the model builder for FasterRCNN. Also adds an option to use the static implementation of the sampler.

--
206803260  by Zhichao Lu:

    Fixes a misplaced argument in resnet fpn feature extractor.

--
206682736  by Zhichao Lu:

    This CL modifies the SSD meta architecture to support both Slim-based and Keras-based box predictors, and begins preparation for Keras box predictor support in the other meta architectures.

    Concretely, this CL adds a new `KerasBoxPredictor` base class and makes the meta architectures appropriately call whichever box predictors they are using.

    We can switch the non-ssd meta architectures to fully support Keras box predictors once the Keras Convolutional Box Predictor CL is submitted.

--
206669634  by Zhichao Lu:

    Adds an alternate method for balanced positive negative sampler using static shapes.

--
206643278  by Zhichao Lu:

    This CL adds a Keras layer hyperparameter configuration object to the hyperparams_builder.

    It automatically converts from Slim layer hyperparameter configs to Keras layer hyperparameters. Namely, it:
    - Builds Keras initializers/regularizers instead of Slim ones
    - sets weights_regularizer/initializer to kernel_regularizer/initializer
    - converts batchnorm decay to momentum
    - converts Slim l2 regularizer weights to the equivalent Keras l2 weights

    This will be used in the conversion of object detection feature extractors & box predictors to newer Tensorflow APIs.

--
206611681  by Zhichao Lu:

    Internal changes.

--
206591619  by Zhichao Lu:

    Clip the to shape when the input tensors are larger than the expected padded static shape

--
206517644  by Zhichao Lu:

    Make MultiscaleGridAnchorGenerator more consistent with MultipleGridAnchorGenerator.

--
206415624  by Zhichao Lu:

    Make the hardcoded feature pyramid network (FPN) levels configurable for both SSD
    Resnet and SSD Mobilenet.

--
206398204  by Zhichao Lu:

    This CL modifies the SSD meta architecture to support both Slim-based and Keras-based feature extractors.

    This allows us to begin the conversion of object detection to newer Tensorflow APIs.

--
206213448  by Zhichao Lu:

    Adding a method to compute the expected classification loss by background/foreground weighting.

--
206204232  by Zhichao Lu:

    Adding the keypoint head to the Mask RCNN pipeline.

--
206200352  by Zhichao Lu:

    - Create Faster R-CNN target assigner in the model builder. This allows configuring matchers in Target assigner to use TPU compatible ops (tf.gather in this case) without any change in meta architecture.
    - As a +ve side effect of the refactoring, we can now re-use a single target assigner for all of second stage heads in Faster R-CNN.

--
206178206  by Zhichao Lu:

    Force ssd feature extractor builder to use keyword arguments so values won't be passed to wrong arguments.

--
206168297  by Zhichao Lu:

    Updating exporter to use freeze_graph.freeze_graph_with_def_protos rather than a homegrown version.

--
206080748  by Zhichao Lu:

    Merge external contributions.

--
206074460  by Zhichao Lu:

    Update to preprocessor to apply temperature and softmax to the multiclass scores on read.

--
205960802  by Zhichao Lu:

    Fixing a bug in hierarchical label expansion script.

--
205944686  by Zhichao Lu:

    Update exporter to support exporting quantized model.

--
205912529  by Zhichao Lu:

    Add a two stage matcher to allow for thresholding by one criteria and then argmaxing on the other.

--
205909017  by Zhichao Lu:

    Add test for grayscale image_resizer

--
205892801  by Zhichao Lu:

    Add flag to decide whether to apply batch norm to conv layers of weight shared box predictor.

--
205824449  by Zhichao Lu:

    make sure that by default mask rcnn box predictor predicts 2 stages.

--
205730139  by Zhichao Lu:

    Updating warning message to be more explicit about variable size mismatch.

--
205696992  by Zhichao Lu:

    Remove utils/ops.py's dependency on core/box_list_ops.py. This will allow re-using TPU compatible ops from utils/ops.py in core/box_list_ops.py.

--
205696867  by Zhichao Lu:

    Refactoring mask rcnn predictor so have each head in a separate file.
    This CL lets us to add new heads more easily in the future to mask rcnn.

--
205492073  by Zhichao Lu:

    Refactor R-FCN box predictor to be TPU compliant.

    - Change utils/ops.py:position_sensitive_crop_regions to operate on single image and set of boxes without `box_ind`
    - Add a batch version that operations on batches of images and batches of boxes.
    - Refactor R-FCN box predictor to use the batched version of position sensitive crop regions.

--
205453567  by Zhichao Lu:

    Fix bug that cannot export inference graph when write_inference_graph flag is True.

--
205316039  by Zhichao Lu:

    Changing input tensor name.

--
205256307  by Zhichao Lu:

    Fix model zoo links for quantized model.

--
205164432  by Zhichao Lu:

    Fixes eval error when label map contains non-ascii characters.

--
205129842  by Zhichao Lu:

    Adds a option to clip the anchors to the window size without filtering the overlapped boxes in Faster-RCNN

--
205094863  by Zhichao Lu:

    Update to label map util to allow the option of adding a background class and fill in gaps in the label map. Useful for using multiclass scores which require a complete label map with explicit background label.

--
204989032  by Zhichao Lu:

    Add tf.prof support to exporter.

--
204825267  by Zhichao Lu:

    Modify mask rcnn box predictor tests for TPU compatibility.

--
204778749  by Zhichao Lu:

    Remove score filtering from postprocessing.py and rely on filtering logic in tf.image.non_max_suppression

--
204775818  by Zhichao Lu:

    Python3 fixes for object_detection.

--
204745920  by Zhichao Lu:

    Object Detection Dataset visualization tool (documentation).

--
204686993  by Zhichao Lu:

    Internal changes.

--
204559667  by Zhichao Lu:

    Refactor box_predictor.py into multiple files.
    The abstract base class remains in the object_detection/core, The other classes have moved to a separate file each in object_detection/predictors

--
204552847  by Zhichao Lu:

    Update blog post link.

--
204508028  by Zhichao Lu:

    Bump down the batch size to 1024 to be a bit more tolerant to OOM and double the number of iterations. This job still converges to 20.5 mAP in 3 hours.

--

PiperOrigin-RevId: 206852642

* Add original post-processing back.

research/object_detection/core/region_similarity_calculator_test.py

@@ -18,6 +18,7 @@ import tensorflow as tf
 
 from object_detection.core import box_list
 from object_detection.core import region_similarity_calculator
+from object_detection.core import standard_fields as fields
 
 
 class RegionSimilarityCalculatorTest(tf.test.TestCase):
@@ -70,6 +71,25 @@ class RegionSimilarityCalculatorTest(tf.test.TestCase):
       self.assertAllClose(iou_output_1, exp_output_1)
       self.assertAllClose(iou_output_2, exp_output_2)
 
+  def test_get_correct_pairwise_similarity_based_on_thresholded_iou(self):
+    corners1 = tf.constant([[4.0, 3.0, 7.0, 5.0], [5.0, 6.0, 10.0, 7.0]])
+    corners2 = tf.constant([[3.0, 4.0, 6.0, 8.0], [14.0, 14.0, 15.0, 15.0],
+                            [0.0, 0.0, 20.0, 20.0]])
+    scores = tf.constant([.3, .6])
+    iou_threshold = .013
+
+    exp_output = tf.constant([[0.3, 0., 0.3], [0.6, 0., 0.]])
+    boxes1 = box_list.BoxList(corners1)
+    boxes1.add_field(fields.BoxListFields.scores, scores)
+    boxes2 = box_list.BoxList(corners2)
+    iou_similarity_calculator = (
+        region_similarity_calculator.ThresholdedIouSimilarity(
+            iou_threshold=iou_threshold))
+    iou_similarity = iou_similarity_calculator.compare(boxes1, boxes2)
+    with self.test_session() as sess:
+      iou_output = sess.run(iou_similarity)
+      self.assertAllClose(iou_output, exp_output)
+
 
 if __name__ == '__main__':
   tf.test.main()

research/object_detection/core/target_assigner.py

@@ -49,7 +49,7 @@ class TargetAssigner(object):
   """Target assigner to compute classification and regression targets."""
 
   def __init__(self, similarity_calc, matcher, box_coder,
-               negative_class_weight=1.0, unmatched_cls_target=None):
+               negative_class_weight=1.0):
     """Construct Object Detection Target Assigner.
 
     Args:
@@ -60,12 +60,6 @@ class TargetAssigner(object):
         groundtruth boxes with respect to anchors.
       negative_class_weight: classification weight to be associated to negative
         anchors (default: 1.0). The weight must be in [0., 1.].
-      unmatched_cls_target: a float32 tensor with shape [d_1, d_2, ..., d_k]
-        which is consistent with the classification target for each
-        anchor (and can be empty for scalar targets).  This shape must thus be
-        compatible with the groundtruth labels that are passed to the "assign"
-        function (which have shape [num_gt_boxes, d_1, d_2, ..., d_k]).
-        If set to None, unmatched_cls_target is set to be [0] for each anchor.
 
     Raises:
       ValueError: if similarity_calc is not a RegionSimilarityCalculator or
@@ -81,17 +75,14 @@ class TargetAssigner(object):
     self._matcher = matcher
     self._box_coder = box_coder
     self._negative_class_weight = negative_class_weight
-    if unmatched_cls_target is None:
-      self._unmatched_cls_target = tf.constant([0], tf.float32)
-    else:
-      self._unmatched_cls_target = unmatched_cls_target
 
   @property
   def box_coder(self):
     return self._box_coder
 
+  # TODO(rathodv): move labels, scores, and weights to groundtruth_boxes fields.
   def assign(self, anchors, groundtruth_boxes, groundtruth_labels=None,
-             groundtruth_weights=None, **params):
+             unmatched_class_label=None, groundtruth_weights=None, **params):
     """Assign classification and regression targets to each anchor.
 
     For a given set of anchors and groundtruth detections, match anchors
@@ -110,6 +101,12 @@ class TargetAssigner(object):
         [d_1, ... d_k] can be empty (corresponding to scalar inputs).  When set
         to None, groundtruth_labels assumes a binary problem where all
         ground_truth boxes get a positive label (of 1).
+      unmatched_class_label: a float32 tensor with shape [d_1, d_2, ..., d_k]
+        which is consistent with the classification target for each
+        anchor (and can be empty for scalar targets).  This shape must thus be
+        compatible with the groundtruth labels that are passed to the "assign"
+        function (which have shape [num_gt_boxes, d_1, d_2, ..., d_k]).
+        If set to None, unmatched_cls_target is set to be [0] for each anchor.
       groundtruth_weights: a float tensor of shape [M] indicating the weight to
         assign to all anchors match to a particular groundtruth box. The weights
         must be in [0., 1.]. If None, all weights are set to 1.
@@ -136,14 +133,17 @@ class TargetAssigner(object):
     if not isinstance(groundtruth_boxes, box_list.BoxList):
       raise ValueError('groundtruth_boxes must be an BoxList')
 
+    if unmatched_class_label is None:
+      unmatched_class_label = tf.constant([0], tf.float32)
+
     if groundtruth_labels is None:
       groundtruth_labels = tf.ones(tf.expand_dims(groundtruth_boxes.num_boxes(),
                                                   0))
       groundtruth_labels = tf.expand_dims(groundtruth_labels, -1)
+
     unmatched_shape_assert = shape_utils.assert_shape_equal(
         shape_utils.combined_static_and_dynamic_shape(groundtruth_labels)[1:],
-        shape_utils.combined_static_and_dynamic_shape(
-            self._unmatched_cls_target))
+        shape_utils.combined_static_and_dynamic_shape(unmatched_class_label))
     labels_and_box_shapes_assert = shape_utils.assert_shape_equal(
         shape_utils.combined_static_and_dynamic_shape(
             groundtruth_labels)[:1],
@@ -155,6 +155,12 @@ class TargetAssigner(object):
       if not num_gt_boxes:
         num_gt_boxes = groundtruth_boxes.num_boxes()
       groundtruth_weights = tf.ones([num_gt_boxes], dtype=tf.float32)
+
+    # set scores on the gt boxes
+    scores = 1 - groundtruth_labels[:, 0]
+
+    groundtruth_boxes.add_field(fields.BoxListFields.scores, scores)
+
     with tf.control_dependencies(
         [unmatched_shape_assert, labels_and_box_shapes_assert]):
       match_quality_matrix = self._similarity_calc.compare(groundtruth_boxes,
@@ -164,6 +170,7 @@ class TargetAssigner(object):
                                                     groundtruth_boxes,
                                                     match)
       cls_targets = self._create_classification_targets(groundtruth_labels,
+                                                        unmatched_class_label,
                                                         match)
       reg_weights = self._create_regression_weights(match, groundtruth_weights)
       cls_weights = self._create_classification_weights(match,
@@ -245,7 +252,8 @@ class TargetAssigner(object):
     """
     return tf.constant([self._box_coder.code_size*[0]], tf.float32)
 
-  def _create_classification_targets(self, groundtruth_labels, match):
+  def _create_classification_targets(self, groundtruth_labels,
+                                     unmatched_class_label, match):
     """Create classification targets for each anchor.
 
     Assign a classification target of for each anchor to the matching
@@ -256,6 +264,11 @@ class TargetAssigner(object):
       groundtruth_labels:  a tensor of shape [num_gt_boxes, d_1, ... d_k]
         with labels for each of the ground_truth boxes. The subshape
         [d_1, ... d_k] can be empty (corresponding to scalar labels).
+      unmatched_class_label: a float32 tensor with shape [d_1, d_2, ..., d_k]
+        which is consistent with the classification target for each
+        anchor (and can be empty for scalar targets).  This shape must thus be
+        compatible with the groundtruth labels that are passed to the "assign"
+        function (which have shape [num_gt_boxes, d_1, d_2, ..., d_k]).
       match: a matcher.Match object that provides a matching between anchors
         and groundtruth boxes.
 
@@ -266,8 +279,8 @@ class TargetAssigner(object):
     """
     return match.gather_based_on_match(
         groundtruth_labels,
-        unmatched_value=self._unmatched_cls_target,
-        ignored_value=self._unmatched_cls_target)
+        unmatched_value=unmatched_class_label,
+        ignored_value=unmatched_class_label)
 
   def _create_regression_weights(self, match, groundtruth_weights):
     """Set regression weight for each anchor.
@@ -327,8 +340,7 @@ class TargetAssigner(object):
 # TODO(rathodv): This method pulls in all the implementation dependencies into
 # core. Therefore its best to have this factory method outside of core.
 def create_target_assigner(reference, stage=None,
-                           negative_class_weight=1.0,
-                           unmatched_cls_target=None):
+                           negative_class_weight=1.0, use_matmul_gather=False):
   """Factory function for creating standard target assigners.
 
   Args:
@@ -336,12 +348,8 @@ def create_target_assigner(reference, stage=None,
     stage: string denoting stage: {proposal, detection}.
     negative_class_weight: classification weight to be associated to negative
       anchors (default: 1.0)
-    unmatched_cls_target: a float32 tensor with shape [d_1, d_2, ..., d_k]
-      which is consistent with the classification target for each
-      anchor (and can be empty for scalar targets).  This shape must thus be
-      compatible with the groundtruth labels that are passed to the Assign
-      function (which have shape [num_gt_boxes, d_1, d_2, ..., d_k]).
-      If set to None, unmatched_cls_target is set to be 0 for each anchor.
+    use_matmul_gather: whether to use matrix multiplication based gather which
+      are better suited for TPUs.
 
   Returns:
     TargetAssigner: desired target assigner.
@@ -358,7 +366,8 @@ def create_target_assigner(reference, stage=None,
     similarity_calc = sim_calc.IouSimilarity()
     matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.7,
                                            unmatched_threshold=0.3,
-                                           force_match_for_each_row=True)
+                                           force_match_for_each_row=True,
+                                           use_matmul_gather=use_matmul_gather)
     box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
         scale_factors=[10.0, 10.0, 5.0, 5.0])
 
@@ -366,7 +375,8 @@ def create_target_assigner(reference, stage=None,
     similarity_calc = sim_calc.IouSimilarity()
     # Uses all proposals with IOU < 0.5 as candidate negatives.
     matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
-                                           negatives_lower_than_unmatched=True)
+                                           negatives_lower_than_unmatched=True,
+                                           use_matmul_gather=use_matmul_gather)
     box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
         scale_factors=[10.0, 10.0, 5.0, 5.0])
 
@@ -375,21 +385,22 @@ def create_target_assigner(reference, stage=None,
     matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
                                            unmatched_threshold=0.1,
                                            force_match_for_each_row=False,
-                                           negatives_lower_than_unmatched=False)
+                                           negatives_lower_than_unmatched=False,
+                                           use_matmul_gather=use_matmul_gather)
     box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder()
 
   else:
     raise ValueError('No valid combination of reference and stage.')
 
   return TargetAssigner(similarity_calc, matcher, box_coder,
-                        negative_class_weight=negative_class_weight,
-                        unmatched_cls_target=unmatched_cls_target)
+                        negative_class_weight=negative_class_weight)
 
 
 def batch_assign_targets(target_assigner,
                          anchors_batch,
                          gt_box_batch,
                          gt_class_targets_batch,
+                         unmatched_class_label=None,
                          gt_weights_batch=None):
   """Batched assignment of classification and regression targets.
 
@@ -403,6 +414,11 @@ def batch_assign_targets(target_assigner,
       each tensor has shape [num_gt_boxes_i, classification_target_size] and
       num_gt_boxes_i is the number of boxes in the ith boxlist of
       gt_box_batch.
+    unmatched_class_label: a float32 tensor with shape [d_1, d_2, ..., d_k]
+      which is consistent with the classification target for each
+      anchor (and can be empty for scalar targets).  This shape must thus be
+      compatible with the groundtruth labels that are passed to the "assign"
+      function (which have shape [num_gt_boxes, d_1, d_2, ..., d_k]).
     gt_weights_batch: A list of 1-D tf.float32 tensors of shape
       [num_boxes] containing weights for groundtruth boxes.
 
@@ -442,9 +458,9 @@ def batch_assign_targets(target_assigner,
     gt_weights_batch = [None] * len(gt_class_targets_batch)
   for anchors, gt_boxes, gt_class_targets, gt_weights in zip(
       anchors_batch, gt_box_batch, gt_class_targets_batch, gt_weights_batch):
-    (cls_targets, cls_weights, reg_targets,
-     reg_weights, match) = target_assigner.assign(
-         anchors, gt_boxes, gt_class_targets, gt_weights)
+    (cls_targets, cls_weights, reg_targets, reg_weights,
+     match) = target_assigner.assign(anchors, gt_boxes, gt_class_targets,
+                                     unmatched_class_label, gt_weights)
     cls_targets_list.append(cls_targets)
     cls_weights_list.append(cls_weights)
     reg_targets_list.append(reg_targets)

research/object_detection/core/target_assigner_test.py

@@ -37,10 +37,11 @@ class TargetAssignerTest(test_case.TestCase):
                                              unmatched_threshold=0.5)
       box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder, unmatched_cls_target=None)
+          similarity_calc, matcher, box_coder)
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist)
+      result = target_assigner.assign(
+          anchors_boxlist, groundtruth_boxlist, unmatched_class_label=None)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -81,10 +82,11 @@ class TargetAssignerTest(test_case.TestCase):
                                              unmatched_threshold=0.3)
       box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder, unmatched_cls_target=None)
+          similarity_calc, matcher, box_coder)
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist)
+      result = target_assigner.assign(
+          anchors_boxlist, groundtruth_boxlist, unmatched_class_label=None)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -120,12 +122,13 @@ class TargetAssignerTest(test_case.TestCase):
       box_coder = keypoint_box_coder.KeypointBoxCoder(
           num_keypoints=6, scale_factors=[10.0, 10.0, 5.0, 5.0])
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder, unmatched_cls_target=None)
+          similarity_calc, matcher, box_coder)
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
       groundtruth_boxlist.add_field(fields.BoxListFields.keypoints,
                                     groundtruth_keypoints)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist)
+      result = target_assigner.assign(
+          anchors_boxlist, groundtruth_boxlist, unmatched_class_label=None)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -174,12 +177,13 @@ class TargetAssignerTest(test_case.TestCase):
       box_coder = keypoint_box_coder.KeypointBoxCoder(
           num_keypoints=6, scale_factors=[10.0, 10.0, 5.0, 5.0])
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder, unmatched_cls_target=None)
+          similarity_calc, matcher, box_coder)
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
       groundtruth_boxlist.add_field(fields.BoxListFields.keypoints,
                                     groundtruth_keypoints)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist)
+      result = target_assigner.assign(
+          anchors_boxlist, groundtruth_boxlist, unmatched_class_label=None)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -221,15 +225,17 @@ class TargetAssignerTest(test_case.TestCase):
       matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
                                              unmatched_threshold=0.5)
       box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
-      unmatched_cls_target = tf.constant([1, 0, 0, 0, 0, 0, 0], tf.float32)
+      unmatched_class_label = tf.constant([1, 0, 0, 0, 0, 0, 0], tf.float32)
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder,
-          unmatched_cls_target=unmatched_cls_target)
+          similarity_calc, matcher, box_coder)
 
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist,
-                                      groundtruth_labels)
+      result = target_assigner.assign(
+          anchors_boxlist,
+          groundtruth_boxlist,
+          groundtruth_labels,
+          unmatched_class_label=unmatched_class_label)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -275,16 +281,18 @@ class TargetAssignerTest(test_case.TestCase):
       matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
                                              unmatched_threshold=0.5)
       box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
-      unmatched_cls_target = tf.constant([1, 0, 0, 0, 0, 0, 0], tf.float32)
+      unmatched_class_label = tf.constant([1, 0, 0, 0, 0, 0, 0], tf.float32)
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder,
-          unmatched_cls_target=unmatched_cls_target)
+          similarity_calc, matcher, box_coder)
 
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist,
+      result = target_assigner.assign(
+          anchors_boxlist,
+          groundtruth_boxlist,
           groundtruth_labels,
-                                      groundtruth_weights)
+          unmatched_class_label=unmatched_class_label,
+          groundtruth_weights=groundtruth_weights)
       (_, cls_weights, _, reg_weights, _) = result
       return (cls_weights, reg_weights)
 
@@ -318,15 +326,17 @@ class TargetAssignerTest(test_case.TestCase):
                                              unmatched_threshold=0.5)
       box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
 
-      unmatched_cls_target = tf.constant([[0, 0], [0, 0]], tf.float32)
+      unmatched_class_label = tf.constant([[0, 0], [0, 0]], tf.float32)
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder,
-          unmatched_cls_target=unmatched_cls_target)
+          similarity_calc, matcher, box_coder)
 
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist,
-                                      groundtruth_labels)
+      result = target_assigner.assign(
+          anchors_boxlist,
+          groundtruth_boxlist,
+          groundtruth_labels,
+          unmatched_class_label=unmatched_class_label)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -371,14 +381,16 @@ class TargetAssignerTest(test_case.TestCase):
       matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
                                              unmatched_threshold=0.5)
       box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
-      unmatched_cls_target = tf.constant([0, 0, 0], tf.float32)
+      unmatched_class_label = tf.constant([0, 0, 0], tf.float32)
       anchors_boxlist = box_list.BoxList(anchor_means)
       groundtruth_boxlist = box_list.BoxList(groundtruth_box_corners)
       target_assigner = targetassigner.TargetAssigner(
-          similarity_calc, matcher, box_coder,
-          unmatched_cls_target=unmatched_cls_target)
-      result = target_assigner.assign(anchors_boxlist, groundtruth_boxlist,
-                                      groundtruth_labels)
+          similarity_calc, matcher, box_coder)
+      result = target_assigner.assign(
+          anchors_boxlist,
+          groundtruth_boxlist,
+          groundtruth_labels,
+          unmatched_class_label=unmatched_class_label)
       (cls_targets, cls_weights, reg_targets, reg_weights, _) = result
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
@@ -415,10 +427,9 @@ class TargetAssignerTest(test_case.TestCase):
     similarity_calc = region_similarity_calculator.NegSqDistSimilarity()
     matcher = bipartite_matcher.GreedyBipartiteMatcher()
     box_coder = mean_stddev_box_coder.MeanStddevBoxCoder()
-    unmatched_cls_target = tf.constant([1, 0, 0, 0, 0, 0, 0], tf.float32)
+    unmatched_class_label = tf.constant([1, 0, 0, 0, 0, 0, 0], tf.float32)
     target_assigner = targetassigner.TargetAssigner(
-        similarity_calc, matcher, box_coder,
-        unmatched_cls_target=unmatched_cls_target)
+        similarity_calc, matcher, box_coder)
 
     prior_means = tf.constant([[0.0, 0.0, 0.5, 0.5],
                                [0.5, 0.5, 1.0, 0.8],
@@ -436,17 +447,20 @@ class TargetAssignerTest(test_case.TestCase):
                                       [0, 0, 0, 0, 0, 1, 0],
                                       [0, 0, 0, 1, 0, 0, 0]], tf.float32)
     with self.assertRaisesRegexp(ValueError, 'Unequal shapes'):
-      target_assigner.assign(priors, boxes, groundtruth_labels,
+      target_assigner.assign(
+          priors,
+          boxes,
+          groundtruth_labels,
+          unmatched_class_label=unmatched_class_label,
           num_valid_rows=3)
 
   def test_raises_error_on_invalid_groundtruth_labels(self):
     similarity_calc = region_similarity_calculator.NegSqDistSimilarity()
     matcher = bipartite_matcher.GreedyBipartiteMatcher()
     box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=1.0)
-    unmatched_cls_target = tf.constant([[0, 0], [0, 0], [0, 0]], tf.float32)
+    unmatched_class_label = tf.constant([[0, 0], [0, 0], [0, 0]], tf.float32)
     target_assigner = targetassigner.TargetAssigner(
-        similarity_calc, matcher, box_coder,
-        unmatched_cls_target=unmatched_cls_target)
+        similarity_calc, matcher, box_coder)
 
     prior_means = tf.constant([[0.0, 0.0, 0.5, 0.5]])
     priors = box_list.BoxList(prior_means)
@@ -458,41 +472,22 @@ class TargetAssignerTest(test_case.TestCase):
     groundtruth_labels = tf.constant([[[0, 1], [1, 0]]], tf.float32)
 
     with self.assertRaises(ValueError):
-      target_assigner.assign(priors, boxes, groundtruth_labels,
+      target_assigner.assign(
+          priors,
+          boxes,
+          groundtruth_labels,
+          unmatched_class_label=unmatched_class_label,
           num_valid_rows=3)
 
 
 class BatchTargetAssignerTest(test_case.TestCase):
 
-  def _get_agnostic_target_assigner(self):
-    similarity_calc = region_similarity_calculator.IouSimilarity()
-    matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
-                                           unmatched_threshold=0.5)
-    box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
-    return targetassigner.TargetAssigner(
-        similarity_calc, matcher, box_coder,
-        unmatched_cls_target=None)
-
-  def _get_multi_class_target_assigner(self, num_classes):
+  def _get_target_assigner(self):
     similarity_calc = region_similarity_calculator.IouSimilarity()
     matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
                                            unmatched_threshold=0.5)
     box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
-    unmatched_cls_target = tf.constant([1] + num_classes * [0], tf.float32)
-    return targetassigner.TargetAssigner(
-        similarity_calc, matcher, box_coder,
-        unmatched_cls_target=unmatched_cls_target)
-
-  def _get_multi_dimensional_target_assigner(self, target_dimensions):
-    similarity_calc = region_similarity_calculator.IouSimilarity()
-    matcher = argmax_matcher.ArgMaxMatcher(matched_threshold=0.5,
-                                           unmatched_threshold=0.5)
-    box_coder = mean_stddev_box_coder.MeanStddevBoxCoder(stddev=0.1)
-    unmatched_cls_target = tf.constant(np.zeros(target_dimensions),
-                                       tf.float32)
-    return targetassigner.TargetAssigner(
-        similarity_calc, matcher, box_coder,
-        unmatched_cls_target=unmatched_cls_target)
+    return targetassigner.TargetAssigner(similarity_calc, matcher, box_coder)
 
   def test_batch_assign_targets(self):
 
@@ -502,7 +497,7 @@ class BatchTargetAssignerTest(test_case.TestCase):
       gt_box_batch = [box_list1, box_list2]
       gt_class_targets = [None, None]
       anchors_boxlist = box_list.BoxList(anchor_means)
-      agnostic_target_assigner = self._get_agnostic_target_assigner()
+      agnostic_target_assigner = self._get_target_assigner()
       (cls_targets, cls_weights, reg_targets, reg_weights,
        _) = targetassigner.batch_assign_targets(
            agnostic_target_assigner, anchors_boxlist, gt_box_batch,
@@ -550,12 +545,13 @@ class BatchTargetAssignerTest(test_case.TestCase):
       gt_box_batch = [box_list1, box_list2]
       gt_class_targets = [class_targets1, class_targets2]
       anchors_boxlist = box_list.BoxList(anchor_means)
-      multiclass_target_assigner = self._get_multi_class_target_assigner(
-          num_classes=3)
+      multiclass_target_assigner = self._get_target_assigner()
+      num_classes = 3
+      unmatched_class_label = tf.constant([1] + num_classes * [0], tf.float32)
       (cls_targets, cls_weights, reg_targets, reg_weights,
        _) = targetassigner.batch_assign_targets(
            multiclass_target_assigner, anchors_boxlist, gt_box_batch,
-           gt_class_targets)
+           gt_class_targets, unmatched_class_label)
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
     groundtruth_boxlist1 = np.array([[0., 0., 0.2, 0.2]], dtype=np.float32)
@@ -613,12 +609,13 @@ class BatchTargetAssignerTest(test_case.TestCase):
       gt_class_targets = [class_targets1, class_targets2]
       gt_weights = [groundtruth_weights1, groundtruth_weights2]
       anchors_boxlist = box_list.BoxList(anchor_means)
-      multiclass_target_assigner = self._get_multi_class_target_assigner(
-          num_classes=3)
+      multiclass_target_assigner = self._get_target_assigner()
+      num_classes = 3
+      unmatched_class_label = tf.constant([1] + num_classes * [0], tf.float32)
       (cls_targets, cls_weights, reg_targets, reg_weights,
        _) = targetassigner.batch_assign_targets(
            multiclass_target_assigner, anchors_boxlist, gt_box_batch,
-           gt_class_targets, gt_weights)
+           gt_class_targets, unmatched_class_label, gt_weights)
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
     groundtruth_boxlist1 = np.array([[0., 0., 0.2, 0.2],
@@ -680,12 +677,14 @@ class BatchTargetAssignerTest(test_case.TestCase):
       gt_box_batch = [box_list1, box_list2]
       gt_class_targets = [class_targets1, class_targets2]
       anchors_boxlist = box_list.BoxList(anchor_means)
-      multiclass_target_assigner = self._get_multi_dimensional_target_assigner(
-          target_dimensions=(2, 3))
+      multiclass_target_assigner = self._get_target_assigner()
+      target_dimensions = (2, 3)
+      unmatched_class_label = tf.constant(np.zeros(target_dimensions),
+                                          tf.float32)
       (cls_targets, cls_weights, reg_targets, reg_weights,
        _) = targetassigner.batch_assign_targets(
            multiclass_target_assigner, anchors_boxlist, gt_box_batch,
-           gt_class_targets)
+           gt_class_targets, unmatched_class_label)
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
     groundtruth_boxlist1 = np.array([[0., 0., 0.2, 0.2]], dtype=np.float32)
@@ -754,13 +753,13 @@ class BatchTargetAssignerTest(test_case.TestCase):
       gt_class_targets_batch = [gt_class_targets]
       anchors_boxlist = box_list.BoxList(anchor_means)
 
-      multiclass_target_assigner = self._get_multi_class_target_assigner(
-          num_classes=3)
-
+      multiclass_target_assigner = self._get_target_assigner()
+      num_classes = 3
+      unmatched_class_label = tf.constant([1] + num_classes * [0], tf.float32)
       (cls_targets, cls_weights, reg_targets, reg_weights,
        _) = targetassigner.batch_assign_targets(
            multiclass_target_assigner, anchors_boxlist,
-           gt_box_batch, gt_class_targets_batch)
+           gt_box_batch, gt_class_targets_batch, unmatched_class_label)
       return (cls_targets, cls_weights, reg_targets, reg_weights)
 
     groundtruth_box_corners = np.zeros((0, 4), dtype=np.float32)

research/object_detection/dataset_tools/oid_hierarchical_labels_expansion.py

@@ -162,12 +162,12 @@ def main(parsed_args):
       for line in source:
         if not header:
           header = line
+          target.writelines(header)
           continue
         if labels_file:
           expanded_lines = expansion_generator.expand_labels_from_csv(line)
         else:
           expanded_lines = expansion_generator.expand_boxes_from_csv(line)
-        expanded_lines = [header] + expanded_lines
         target.writelines(expanded_lines)
 
 

research/object_detection/export_inference_graph.py

@@ -140,10 +140,10 @@ def main(_):
     ]
   else:
     input_shape = None
-  exporter.export_inference_graph(FLAGS.input_type, pipeline_config,
-                                  FLAGS.trained_checkpoint_prefix,
-                                  FLAGS.output_directory, input_shape,
-                                  FLAGS.write_inference_graph)
+  exporter.export_inference_graph(
+      FLAGS.input_type, pipeline_config, FLAGS.trained_checkpoint_prefix,
+      FLAGS.output_directory, input_shape=input_shape,
+      write_inference_graph=FLAGS.write_inference_graph)
 
 
 if __name__ == '__main__':

research/object_detection/export_tflite_ssd_graph_lib.py

@@ -258,6 +258,7 @@ def export_tflite_graph(pipeline_config, trained_checkpoint_prefix, output_dir,
       restore_op_name='save/restore_all',
       filename_tensor_name='save/Const:0',
       clear_devices=True,
+      output_graph='',
       initializer_nodes='')
 
   # Add new operation to do post processing in a custom op (TF Lite only)

research/object_detection/exporter.py

@@ -14,17 +14,16 @@
 # ==============================================================================
 
 """Functions to export object detection inference graph."""
-import logging
 import os
 import tempfile
 import tensorflow as tf
 from tensorflow.core.protobuf import saver_pb2
-from tensorflow.python import pywrap_tensorflow
 from tensorflow.python.client import session
-from tensorflow.python.framework import graph_util
 from tensorflow.python.platform import gfile
 from tensorflow.python.saved_model import signature_constants
+from tensorflow.python.tools import freeze_graph
 from tensorflow.python.training import saver as saver_lib
+from object_detection.builders import graph_rewriter_builder
 from object_detection.builders import model_builder
 from object_detection.core import standard_fields as fields
 from object_detection.data_decoders import tf_example_decoder
@@ -32,70 +31,7 @@ from object_detection.utils import config_util
 
 slim = tf.contrib.slim
 
-
-# TODO(derekjchow): Replace with freeze_graph.freeze_graph_with_def_protos when
-# newer version of Tensorflow becomes more common.
-def freeze_graph_with_def_protos(
-    input_graph_def,
-    input_saver_def,
-    input_checkpoint,
-    output_node_names,
-    restore_op_name,
-    filename_tensor_name,
-    clear_devices,
-    initializer_nodes,
-    variable_names_blacklist=''):
-  """Converts all variables in a graph and checkpoint into constants."""
-  del restore_op_name, filename_tensor_name  # Unused by updated loading code.
-
-  # 'input_checkpoint' may be a prefix if we're using Saver V2 format
-  if not saver_lib.checkpoint_exists(input_checkpoint):
-    raise ValueError(
-        'Input checkpoint "' + input_checkpoint + '" does not exist!')
-
-  if not output_node_names:
-    raise ValueError(
-        'You must supply the name of a node to --output_node_names.')
-
-  # Remove all the explicit device specifications for this node. This helps to
-  # make the graph more portable.
-  if clear_devices:
-    for node in input_graph_def.node:
-      node.device = ''
-
-  with tf.Graph().as_default():
-    tf.import_graph_def(input_graph_def, name='')
-    config = tf.ConfigProto(graph_options=tf.GraphOptions())
-    with session.Session(config=config) as sess:
-      if input_saver_def:
-        saver = saver_lib.Saver(saver_def=input_saver_def)
-        saver.restore(sess, input_checkpoint)
-      else:
-        var_list = {}
-        reader = pywrap_tensorflow.NewCheckpointReader(input_checkpoint)
-        var_to_shape_map = reader.get_variable_to_shape_map()
-        for key in var_to_shape_map:
-          try:
-            tensor = sess.graph.get_tensor_by_name(key + ':0')
-          except KeyError:
-            # This tensor doesn't exist in the graph (for example it's
-            # 'global_step' or a similar housekeeping element) so skip it.
-            continue
-          var_list[key] = tensor
-        saver = saver_lib.Saver(var_list=var_list)
-        saver.restore(sess, input_checkpoint)
-        if initializer_nodes:
-          sess.run(initializer_nodes)
-
-      variable_names_blacklist = (variable_names_blacklist.split(',') if
-                                  variable_names_blacklist else None)
-      output_graph_def = graph_util.convert_variables_to_constants(
-          sess,
-          input_graph_def,
-          output_node_names.split(','),
-          variable_names_blacklist=variable_names_blacklist)
-
-  return output_graph_def
+freeze_graph_with_def_protos = freeze_graph.freeze_graph_with_def_protos
 
 
 def replace_variable_values_with_moving_averages(graph,
@@ -247,18 +183,6 @@ def _add_output_tensor_nodes(postprocessed_tensors,
   return outputs
 
 
-def write_frozen_graph(frozen_graph_path, frozen_graph_def):
-  """Writes frozen graph to disk.
-
-  Args:
-    frozen_graph_path: Path to write inference graph.
-    frozen_graph_def: tf.GraphDef holding frozen graph.
-  """
-  with gfile.GFile(frozen_graph_path, 'wb') as f:
-    f.write(frozen_graph_def.SerializeToString())
-  logging.info('%d ops in the final graph.', len(frozen_graph_def.node))
-
-
 def write_saved_model(saved_model_path,
                       frozen_graph_def,
                       inputs,
@@ -384,6 +308,7 @@ def _export_inference_graph(input_type,
       output_collection_name=output_collection_name,
       graph_hook_fn=graph_hook_fn)
 
+  profile_inference_graph(tf.get_default_graph())
   saver_kwargs = {}
   if use_moving_averages:
     # This check is to be compatible with both version of SaverDef.
@@ -421,16 +346,17 @@ def _export_inference_graph(input_type,
   else:
     output_node_names = ','.join(outputs.keys())
 
-  frozen_graph_def = freeze_graph_with_def_protos(
+  frozen_graph_def = freeze_graph.freeze_graph_with_def_protos(
       input_graph_def=tf.get_default_graph().as_graph_def(),
       input_saver_def=input_saver_def,
       input_checkpoint=checkpoint_to_use,
       output_node_names=output_node_names,
       restore_op_name='save/restore_all',
       filename_tensor_name='save/Const:0',
+      output_graph=frozen_graph_path,
       clear_devices=True,
       initializer_nodes='')
-  write_frozen_graph(frozen_graph_path, frozen_graph_def)
+
   write_saved_model(saved_model_path, frozen_graph_def,
                     placeholder_tensor, outputs)
 
@@ -461,6 +387,11 @@ def export_inference_graph(input_type,
   """
   detection_model = model_builder.build(pipeline_config.model,
                                         is_training=False)
+  graph_rewriter_fn = None
+  if pipeline_config.HasField('graph_rewriter'):
+    graph_rewriter_config = pipeline_config.graph_rewriter
+    graph_rewriter_fn = graph_rewriter_builder.build(graph_rewriter_config,
+                                                     is_training=False)
   _export_inference_graph(
       input_type,
       detection_model,
@@ -470,7 +401,39 @@ def export_inference_graph(input_type,
       additional_output_tensor_names,
       input_shape,
       output_collection_name,
-      graph_hook_fn=None,
+      graph_hook_fn=graph_rewriter_fn,
       write_inference_graph=write_inference_graph)
   pipeline_config.eval_config.use_moving_averages = False
   config_util.save_pipeline_config(pipeline_config, output_directory)
+
+
+def profile_inference_graph(graph):
+  """Profiles the inference graph.
+
+  Prints model parameters and computation FLOPs given an inference graph.
+  BatchNorms are excluded from the parameter count due to the fact that
+  BatchNorms are usually folded. BatchNorm, Initializer, Regularizer
+  and BiasAdd are not considered in FLOP count.
+
+  Args:
+    graph: the inference graph.
+  """
+  tfprof_vars_option = (
+      tf.contrib.tfprof.model_analyzer.TRAINABLE_VARS_PARAMS_STAT_OPTIONS)
+  tfprof_flops_option = tf.contrib.tfprof.model_analyzer.FLOAT_OPS_OPTIONS
+
+  # Batchnorm is usually folded during inference.
+  tfprof_vars_option['trim_name_regexes'] = ['.*BatchNorm.*']
+  # Initializer and Regularizer are only used in training.
+  tfprof_flops_option['trim_name_regexes'] = [
+      '.*BatchNorm.*', '.*Initializer.*', '.*Regularizer.*', '.*BiasAdd.*'
+  ]
+
+  tf.contrib.tfprof.model_analyzer.print_model_analysis(
+      graph,
+      tfprof_options=tfprof_vars_option)
+
+  tf.contrib.tfprof.model_analyzer.print_model_analysis(
+      graph,
+      tfprof_options=tfprof_flops_option)
+

research/object_detection/exporter_test.py

@@ -20,8 +20,10 @@ import six
 import tensorflow as tf
 from google.protobuf import text_format
 from object_detection import exporter
+from object_detection.builders import graph_rewriter_builder
 from object_detection.builders import model_builder
 from object_detection.core import model
+from object_detection.protos import graph_rewriter_pb2
 from object_detection.protos import pipeline_pb2
 
 if six.PY2:
@@ -75,8 +77,10 @@ class FakeModel(model.DetectionModel):
 
 class ExportInferenceGraphTest(tf.test.TestCase):
 
-  def _save_checkpoint_from_mock_model(self, checkpoint_path,
-                                       use_moving_averages):
+  def _save_checkpoint_from_mock_model(self,
+                                       checkpoint_path,
+                                       use_moving_averages,
+                                       enable_quantization=False):
     g = tf.Graph()
     with g.as_default():
       mock_model = FakeModel()
@@ -86,20 +90,28 @@ class ExportInferenceGraphTest(tf.test.TestCase):
       mock_model.postprocess(predictions, true_image_shapes)
       if use_moving_averages:
         tf.train.ExponentialMovingAverage(0.0).apply()
-      slim.get_or_create_global_step()
+      tf.train.get_or_create_global_step()
+      if enable_quantization:
+        graph_rewriter_config = graph_rewriter_pb2.GraphRewriter()
+        graph_rewriter_config.quantization.delay = 500000
+        graph_rewriter_fn = graph_rewriter_builder.build(
+            graph_rewriter_config, is_training=False)
+        graph_rewriter_fn()
       saver = tf.train.Saver()
       init = tf.global_variables_initializer()
       with self.test_session() as sess:
         sess.run(init)
         saver.save(sess, checkpoint_path)
 
-  def _load_inference_graph(self, inference_graph_path):
+  def _load_inference_graph(self, inference_graph_path, is_binary=True):
     od_graph = tf.Graph()
     with od_graph.as_default():
       od_graph_def = tf.GraphDef()
       with tf.gfile.GFile(inference_graph_path) as fid:
-        serialized_graph = fid.read()
-        od_graph_def.ParseFromString(serialized_graph)
+        if is_binary:
+          od_graph_def.ParseFromString(fid.read())
+        else:
+          text_format.Parse(fid.read(), od_graph_def)
         tf.import_graph_def(od_graph_def, name='')
     return od_graph
 
@@ -284,6 +296,42 @@ class ExportInferenceGraphTest(tf.test.TestCase):
         [var_name for var_name, _ in tf.train.list_variables(output_directory)])
     self.assertTrue(expected_variables.issubset(actual_variables))
 
+  def test_export_model_with_quantization_nodes(self):
+    tmp_dir = self.get_temp_dir()
+    trained_checkpoint_prefix = os.path.join(tmp_dir, 'model.ckpt')
+    self._save_checkpoint_from_mock_model(
+        trained_checkpoint_prefix,
+        use_moving_averages=False,
+        enable_quantization=True)
+    output_directory = os.path.join(tmp_dir, 'output')
+    inference_graph_path = os.path.join(output_directory,
+                                        'inference_graph.pbtxt')
+    with mock.patch.object(
+        model_builder, 'build', autospec=True) as mock_builder:
+      mock_builder.return_value = FakeModel()
+      pipeline_config = pipeline_pb2.TrainEvalPipelineConfig()
+      text_format.Merge(
+          """graph_rewriter {
+               quantization {
+                 delay: 50000
+                 activation_bits: 8
+                 weight_bits: 8
+               }
+             }""", pipeline_config)
+      exporter.export_inference_graph(
+          input_type='image_tensor',
+          pipeline_config=pipeline_config,
+          trained_checkpoint_prefix=trained_checkpoint_prefix,
+          output_directory=output_directory,
+          write_inference_graph=True)
+    self._load_inference_graph(inference_graph_path, is_binary=False)
+    has_quant_nodes = False
+    for v in tf.global_variables():
+      if v.op.name.endswith('act_quant/min'):
+        has_quant_nodes = True
+        break
+    self.assertTrue(has_quant_nodes)
+
   def test_export_model_with_all_output_nodes(self):
     tmp_dir = self.get_temp_dir()
     trained_checkpoint_prefix = os.path.join(tmp_dir, 'model.ckpt')
@@ -564,16 +612,16 @@ class ExportInferenceGraphTest(tf.test.TestCase):
       output_node_names = ','.join(outputs.keys())
       saver = tf.train.Saver()
       input_saver_def = saver.as_saver_def()
-      frozen_graph_def = exporter.freeze_graph_with_def_protos(
+      exporter.freeze_graph_with_def_protos(
           input_graph_def=tf.get_default_graph().as_graph_def(),
           input_saver_def=input_saver_def,
           input_checkpoint=trained_checkpoint_prefix,
           output_node_names=output_node_names,
           restore_op_name='save/restore_all',
           filename_tensor_name='save/Const:0',
+          output_graph=inference_graph_path,
           clear_devices=True,
           initializer_nodes='')
-      exporter.write_frozen_graph(inference_graph_path, frozen_graph_def)
 
     inference_graph = self._load_inference_graph(inference_graph_path)
     tf_example_np = np.expand_dims(self._create_tf_example(
@@ -719,6 +767,7 @@ class ExportInferenceGraphTest(tf.test.TestCase):
           output_node_names=output_node_names,
           restore_op_name='save/restore_all',
           filename_tensor_name='save/Const:0',
+          output_graph='',
           clear_devices=True,
           initializer_nodes='')
       exporter.write_saved_model(

research/object_detection/g3doc/installation.md

@@ -48,6 +48,7 @@ pip install --user jupyter
 pip install --user matplotlib
 ```
 
+<!-- common_typos_disable -->
 **Note**: sometimes "sudo apt-get install protobuf-compiler" will install
 Protobuf 3+ versions for you and some users have issues when using 3.5.
 If that is your case, try the [manual](#Manual-protobuf-compiler-installation-and-usage) installation.
@@ -88,6 +89,7 @@ protoc object_detection/protos/*.proto --python_out=.
 
 ## Manual protobuf-compiler installation and usage
 Download and install the 3.0 release of protoc, then unzip the file.
+
 ```bash
 # From tensorflow/models/research/
 wget -O protobuf.zip https://github.com/google/protobuf/releases/download/v3.0.0/protoc-3.0.0-linux-x86_64.zip

research/object_detection/inputs.py

@@ -219,20 +219,8 @@ def pad_input_data_to_static_shapes(tensor_dict, max_num_boxes, num_classes,
 
   padded_tensor_dict = {}
   for tensor_name in tensor_dict:
-    expected_shape = padding_shapes[tensor_name]
-    current_shape = shape_utils.combined_static_and_dynamic_shape(
-        tensor_dict[tensor_name])
-    trailing_paddings = [
-        expected_shape_dim - current_shape_dim if expected_shape_dim else 0
-        for expected_shape_dim, current_shape_dim in zip(
-            expected_shape, current_shape)
-    ]
-    paddings = tf.stack([tf.zeros(len(trailing_paddings), dtype=tf.int32),
-                         trailing_paddings],
-                        axis=1)
-    padded_tensor_dict[tensor_name] = tf.pad(
-        tensor_dict[tensor_name], paddings=paddings)
-    padded_tensor_dict[tensor_name].set_shape(expected_shape)
+    padded_tensor_dict[tensor_name] = shape_utils.pad_or_clip_nd(
+        tensor_dict[tensor_name], padding_shapes[tensor_name])
   return padded_tensor_dict
 
 
@@ -529,7 +517,7 @@ def create_predict_input_fn(model_config, predict_input_config):
       `ServingInputReceiver`.
     """
     del params
-    example = tf.placeholder(dtype=tf.string, shape=[], name='input_feature')
+    example = tf.placeholder(dtype=tf.string, shape=[], name='tf_example')
 
     num_classes = config_util.get_number_of_classes(model_config)
     model = model_builder.build(model_config, is_training=False)

research/object_detection/inputs_test.py

@@ -657,6 +657,42 @@ class PadInputDataToStaticShapesFnTest(tf.test.TestCase):
         padded_tensor_dict[fields.InputDataFields.groundtruth_classes]
         .shape.as_list(), [3, 3])
 
+  def test_clip_boxes_and_classes(self):
+    input_tensor_dict = {
+        fields.InputDataFields.groundtruth_boxes:
+            tf.placeholder(tf.float32, [None, 4]),
+        fields.InputDataFields.groundtruth_classes:
+            tf.placeholder(tf.int32, [None, 3]),
+    }
+    padded_tensor_dict = inputs.pad_input_data_to_static_shapes(
+        tensor_dict=input_tensor_dict,
+        max_num_boxes=3,
+        num_classes=3,
+        spatial_image_shape=[5, 6])
+
+    self.assertAllEqual(
+        padded_tensor_dict[fields.InputDataFields.groundtruth_boxes]
+        .shape.as_list(), [3, 4])
+    self.assertAllEqual(
+        padded_tensor_dict[fields.InputDataFields.groundtruth_classes]
+        .shape.as_list(), [3, 3])
+
+    with self.test_session() as sess:
+      out_tensor_dict = sess.run(
+          padded_tensor_dict,
+          feed_dict={
+              input_tensor_dict[fields.InputDataFields.groundtruth_boxes]:
+                  np.random.rand(5, 4),
+              input_tensor_dict[fields.InputDataFields.groundtruth_classes]:
+                  np.random.rand(2, 3),
+          })
+
+    self.assertAllEqual(
+        out_tensor_dict[fields.InputDataFields.groundtruth_boxes].shape, [3, 4])
+    self.assertAllEqual(
+        out_tensor_dict[fields.InputDataFields.groundtruth_classes].shape,
+        [3, 3])
+
   def test_do_not_pad_dynamic_images(self):
     input_tensor_dict = {
         fields.InputDataFields.image:

research/object_detection/meta_architectures/faster_rcnn_meta_arch.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 """Faster R-CNN meta-architecture definition.
 
 General tensorflow implementation of Faster R-CNN detection models.
@@ -98,7 +97,6 @@ from functools import partial
 import tensorflow as tf
 
 from object_detection.anchor_generators import grid_anchor_generator
-from object_detection.core import balanced_positive_negative_sampler as sampler
 from object_detection.core import box_list
 from object_detection.core import box_list_ops
 from object_detection.core import box_predictor
@@ -107,6 +105,7 @@ from object_detection.core import model
 from object_detection.core import post_processing
 from object_detection.core import standard_fields as fields
 from object_detection.core import target_assigner
+from object_detection.predictors import convolutional_box_predictor
 from object_detection.utils import ops
 from object_detection.utils import shape_utils
 
@@ -228,12 +227,13 @@ class FasterRCNNMetaArch(model.DetectionModel):
                feature_extractor,
                number_of_stages,
                first_stage_anchor_generator,
+               first_stage_target_assigner,
                first_stage_atrous_rate,
                first_stage_box_predictor_arg_scope_fn,
                first_stage_box_predictor_kernel_size,
                first_stage_box_predictor_depth,
                first_stage_minibatch_size,
-               first_stage_positive_balance_fraction,
+               first_stage_sampler,
                first_stage_nms_score_threshold,
                first_stage_nms_iou_threshold,
                first_stage_max_proposals,
@@ -242,9 +242,10 @@ class FasterRCNNMetaArch(model.DetectionModel):
                initial_crop_size,
                maxpool_kernel_size,
                maxpool_stride,
+               second_stage_target_assigner,
                second_stage_mask_rcnn_box_predictor,
                second_stage_batch_size,
-               second_stage_balance_fraction,
+               second_stage_sampler,
                second_stage_non_max_suppression_fn,
                second_stage_score_conversion_fn,
                second_stage_localization_loss_weight,
@@ -254,7 +255,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
                hard_example_miner=None,
                parallel_iterations=16,
                add_summaries=True,
-               use_matmul_crop_and_resize=False):
+               use_matmul_crop_and_resize=False,
+               clip_anchors_to_image=False):
     """FasterRCNNMetaArch Constructor.
 
     Args:
@@ -285,6 +287,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
       first_stage_anchor_generator: An anchor_generator.AnchorGenerator object
         (note that currently we only support
         grid_anchor_generator.GridAnchorGenerator objects)
+      first_stage_target_assigner: Target assigner to use for first stage of
+        Faster R-CNN (RPN).
       first_stage_atrous_rate: A single integer indicating the atrous rate for
         the single convolution op which is applied to the `rpn_features_to_crop`
         tensor to obtain a tensor to be used for box prediction. Some feature
@@ -304,8 +308,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
         "batch size" refers to the number of anchors selected as contributing
         to the loss function for any given image within the image batch and is
         only called "batch_size" due to terminology from the Faster R-CNN paper.
-      first_stage_positive_balance_fraction: Fraction of positive examples
-        per image for the RPN. The recommended value for Faster RCNN is 0.5.
+      first_stage_sampler: Sampler to use for first stage loss (RPN loss).
       first_stage_nms_score_threshold: Score threshold for non max suppression
         for the Region Proposal Network (RPN).  This value is expected to be in
         [0, 1] as it is applied directly after a softmax transformation.  The
@@ -325,6 +328,10 @@ class FasterRCNNMetaArch(model.DetectionModel):
         max pool op on the cropped feature map during ROI pooling.
       maxpool_stride: A single integer indicating the stride of the max pool
         op on the cropped feature map during ROI pooling.
+      second_stage_target_assigner: Target assigner to use for second stage of
+        Faster R-CNN. If the model is configured with multiple prediction heads,
+        this target assigner is used to generate targets for all heads (with the
+        correct `unmatched_class_label`).
       second_stage_mask_rcnn_box_predictor: Mask R-CNN box predictor to use for
         the second stage.
       second_stage_batch_size: The batch size used for computing the
@@ -332,9 +339,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
         "batch size" refers to the number of proposals selected as contributing
         to the loss function for any given image within the image batch and is
         only called "batch_size" due to terminology from the Faster R-CNN paper.
-      second_stage_balance_fraction: Fraction of positive examples to use
-        per image for the box classifier. The recommended value for Faster RCNN
-        is 0.25.
+      second_stage_sampler:  Sampler to use for second stage loss (box
+        classifier loss).
       second_stage_non_max_suppression_fn: batch_multiclass_non_max_suppression
         callable that takes `boxes`, `scores`, optional `clip_window` and
         optional (kwarg) `mask` inputs (with all other inputs already set)
@@ -364,6 +370,9 @@ class FasterRCNNMetaArch(model.DetectionModel):
       use_matmul_crop_and_resize: Force the use of matrix multiplication based
         crop and resize instead of standard tf.image.crop_and_resize while
         computing second stage input feature maps.
+      clip_anchors_to_image: Normally, anchors generated for a given image size
+      are pruned during training if they lie outside the image window. This
+      option clips the anchors to be within the image instead of pruning.
 
     Raises:
       ValueError: If `second_stage_batch_size` > `first_stage_max_proposals` at
@@ -388,13 +397,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
     self._feature_extractor = feature_extractor
     self._number_of_stages = number_of_stages
 
-    # The first class is reserved as background.
-    unmatched_cls_target = tf.constant(
-        [1] + self._num_classes * [0], dtype=tf.float32)
-    self._proposal_target_assigner = target_assigner.create_target_assigner(
-        'FasterRCNN', 'proposal')
-    self._detector_target_assigner = target_assigner.create_target_assigner(
-        'FasterRCNN', 'detection', unmatched_cls_target=unmatched_cls_target)
+    self._proposal_target_assigner = first_stage_target_assigner
+    self._detector_target_assigner = second_stage_target_assigner
     # Both proposal and detector target assigners use the same box coder
     self._box_coder = self._proposal_target_assigner.box_coder
 
@@ -407,14 +411,19 @@ class FasterRCNNMetaArch(model.DetectionModel):
         first_stage_box_predictor_kernel_size)
     self._first_stage_box_predictor_depth = first_stage_box_predictor_depth
     self._first_stage_minibatch_size = first_stage_minibatch_size
-    self._first_stage_sampler = sampler.BalancedPositiveNegativeSampler(
-        positive_fraction=first_stage_positive_balance_fraction)
-    self._first_stage_box_predictor = box_predictor.ConvolutionalBoxPredictor(
-        self._is_training, num_classes=1,
+    self._first_stage_sampler = first_stage_sampler
+    self._first_stage_box_predictor = (
+        convolutional_box_predictor.ConvolutionalBoxPredictor(
+            self._is_training,
+            num_classes=1,
             conv_hyperparams_fn=self._first_stage_box_predictor_arg_scope_fn,
-        min_depth=0, max_depth=0, num_layers_before_predictor=0,
-        use_dropout=False, dropout_keep_prob=1.0, kernel_size=1,
-        box_code_size=self._box_coder.code_size)
+            min_depth=0,
+            max_depth=0,
+            num_layers_before_predictor=0,
+            use_dropout=False,
+            dropout_keep_prob=1.0,
+            kernel_size=1,
+            box_code_size=self._box_coder.code_size))
 
     self._first_stage_nms_score_threshold = first_stage_nms_score_threshold
     self._first_stage_nms_iou_threshold = first_stage_nms_iou_threshold
@@ -435,8 +444,7 @@ class FasterRCNNMetaArch(model.DetectionModel):
     self._mask_rcnn_box_predictor = second_stage_mask_rcnn_box_predictor
 
     self._second_stage_batch_size = second_stage_batch_size
-    self._second_stage_sampler = sampler.BalancedPositiveNegativeSampler(
-        positive_fraction=second_stage_balance_fraction)
+    self._second_stage_sampler = second_stage_sampler
 
     self._second_stage_nms_fn = second_stage_non_max_suppression_fn
     self._second_stage_score_conversion_fn = second_stage_score_conversion_fn
@@ -454,6 +462,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
     self._hard_example_miner = hard_example_miner
     self._parallel_iterations = parallel_iterations
 
+    self.clip_anchors_to_image = clip_anchors_to_image
+
     if self._number_of_stages <= 0 or self._number_of_stages > 3:
       raise ValueError('Number of stages should be a value in {1, 2, 3}.')
 
@@ -639,6 +649,10 @@ class FasterRCNNMetaArch(model.DetectionModel):
     # the image window at training time and clipping at inference time.
     clip_window = tf.to_float(tf.stack([0, 0, image_shape[1], image_shape[2]]))
     if self._is_training:
+      if self.clip_anchors_to_image:
+        anchors_boxlist = box_list_ops.clip_to_window(
+            anchors_boxlist, clip_window, filter_nonoverlapping=False)
+      else:
         (rpn_box_encodings, rpn_objectness_predictions_with_background,
          anchors_boxlist) = self._remove_invalid_anchors_and_predictions(
              rpn_box_encodings, rpn_objectness_predictions_with_background,
@@ -761,11 +775,16 @@ class FasterRCNNMetaArch(model.DetectionModel):
             flattened_proposal_feature_maps,
             scope=self.second_stage_feature_extractor_scope))
 
+    if self._mask_rcnn_box_predictor.is_keras_model:
+      box_predictions = self._mask_rcnn_box_predictor(
+          [box_classifier_features],
+          prediction_stage=2)
+    else:
       box_predictions = self._mask_rcnn_box_predictor.predict(
           [box_classifier_features],
           num_predictions_per_location=[1],
           scope=self.second_stage_box_predictor_scope,
-        predict_boxes_and_classes=True)
+          prediction_stage=2)
 
     refined_box_encodings = tf.squeeze(
         box_predictions[box_predictor.BOX_ENCODINGS],
@@ -834,12 +853,16 @@ class FasterRCNNMetaArch(model.DetectionModel):
     if self._is_training:
       curr_box_classifier_features = prediction_dict['box_classifier_features']
       detection_classes = prediction_dict['class_predictions_with_background']
+      if self._mask_rcnn_box_predictor.is_keras_model:
+        mask_predictions = self._mask_rcnn_box_predictor(
+            [curr_box_classifier_features],
+            prediction_stage=3)
+      else:
         mask_predictions = self._mask_rcnn_box_predictor.predict(
             [curr_box_classifier_features],
             num_predictions_per_location=[1],
             scope=self.second_stage_box_predictor_scope,
-          predict_boxes_and_classes=False,
-          predict_auxiliary_outputs=True)
+            prediction_stage=3)
       prediction_dict['mask_predictions'] = tf.squeeze(mask_predictions[
           box_predictor.MASK_PREDICTIONS], axis=1)
     else:
@@ -865,12 +888,16 @@ class FasterRCNNMetaArch(model.DetectionModel):
               flattened_detected_feature_maps,
               scope=self.second_stage_feature_extractor_scope))
 
+      if self._mask_rcnn_box_predictor.is_keras_model:
+        mask_predictions = self._mask_rcnn_box_predictor(
+            [curr_box_classifier_features],
+            prediction_stage=3)
+      else:
         mask_predictions = self._mask_rcnn_box_predictor.predict(
             [curr_box_classifier_features],
             num_predictions_per_location=[1],
             scope=self.second_stage_box_predictor_scope,
-          predict_boxes_and_classes=False,
-          predict_auxiliary_outputs=True)
+            prediction_stage=3)
 
       detection_masks = tf.squeeze(mask_predictions[
           box_predictor.MASK_PREDICTIONS], axis=1)
@@ -976,6 +1003,10 @@ class FasterRCNNMetaArch(model.DetectionModel):
     if len(num_anchors_per_location) != 1:
       raise RuntimeError('anchor_generator is expected to generate anchors '
                          'corresponding to a single feature map.')
+    if self._first_stage_box_predictor.is_keras_model:
+      box_predictions = self._first_stage_box_predictor(
+          [rpn_box_predictor_features])
+    else:
       box_predictions = self._first_stage_box_predictor.predict(
           [rpn_box_predictor_features],
           num_anchors_per_location,
@@ -1393,8 +1424,11 @@ class FasterRCNNMetaArch(model.DetectionModel):
       a BoxList contained sampled proposals.
     """
     (cls_targets, cls_weights, _, _, _) = self._detector_target_assigner.assign(
-        proposal_boxlist, groundtruth_boxlist,
-        groundtruth_classes_with_background)
+        proposal_boxlist,
+        groundtruth_boxlist,
+        groundtruth_classes_with_background,
+        unmatched_class_label=tf.constant(
+            [1] + self._num_classes * [0], dtype=tf.float32))
     # Selects all boxes as candidates if none of them is selected according
     # to cls_weights. This could happen as boxes within certain IOU ranges
     # are ignored. If triggered, the selected boxes will still be ignored
@@ -1672,7 +1706,8 @@ class FasterRCNNMetaArch(model.DetectionModel):
        batch_reg_weights, _) = target_assigner.batch_assign_targets(
            self._proposal_target_assigner, box_list.BoxList(anchors),
            groundtruth_boxlists,
-           len(groundtruth_boxlists) * [None], groundtruth_weights_list)
+           len(groundtruth_boxlists) * [None],
+           gt_weights_batch=groundtruth_weights_list)
       batch_cls_targets = tf.squeeze(batch_cls_targets, axis=2)
 
       def _minibatch_subsample_fn(inputs):
@@ -1792,9 +1827,13 @@ class FasterRCNNMetaArch(model.DetectionModel):
 
       (batch_cls_targets_with_background, batch_cls_weights, batch_reg_targets,
        batch_reg_weights, _) = target_assigner.batch_assign_targets(
-           self._detector_target_assigner, proposal_boxlists,
-           groundtruth_boxlists, groundtruth_classes_with_background_list,
-           groundtruth_weights_list)
+           self._detector_target_assigner,
+           proposal_boxlists,
+           groundtruth_boxlists,
+           groundtruth_classes_with_background_list,
+           unmatched_class_label=tf.constant(
+               [1] + self._num_classes * [0], dtype=tf.float32),
+           gt_weights_batch=groundtruth_weights_list)
 
       class_predictions_with_background = tf.reshape(
           class_predictions_with_background,
@@ -1866,18 +1905,12 @@ class FasterRCNNMetaArch(model.DetectionModel):
           raise ValueError('Groundtruth instance masks not provided. '
                            'Please configure input reader.')
 
-        # Create a new target assigner that matches the proposals to groundtruth
-        # and returns the mask targets.
-        # TODO(rathodv): Move `unmatched_cls_target` from constructor to assign
-        # function. This will enable reuse of a single target assigner for both
-        # class targets and mask targets.
-        mask_target_assigner = target_assigner.create_target_assigner(
-            'FasterRCNN', 'detection',
-            unmatched_cls_target=tf.zeros(image_shape[1:3], dtype=tf.float32))
-        (batch_mask_targets, _, _,
-         batch_mask_target_weights, _) = target_assigner.batch_assign_targets(
-             mask_target_assigner, proposal_boxlists, groundtruth_boxlists,
-             groundtruth_masks_list, groundtruth_weights_list)
+        unmatched_mask_label = tf.zeros(image_shape[1:3], dtype=tf.float32)
+        (batch_mask_targets, _, _, batch_mask_target_weights,
+         _) = target_assigner.batch_assign_targets(
+             self._detector_target_assigner, proposal_boxlists,
+             groundtruth_boxlists, groundtruth_masks_list, unmatched_mask_label,
+             groundtruth_weights_list)
 
         # Pad the prediction_masks with to add zeros for background class to be
         # consistent with class predictions.

research/object_detection/meta_architectures/faster_rcnn_meta_arch_test_lib.py

@@ -21,7 +21,9 @@ from object_detection.anchor_generators import grid_anchor_generator
 from object_detection.builders import box_predictor_builder
 from object_detection.builders import hyperparams_builder
 from object_detection.builders import post_processing_builder
+from object_detection.core import balanced_positive_negative_sampler as sampler
 from object_detection.core import losses
+from object_detection.core import target_assigner
 from object_detection.meta_architectures import faster_rcnn_meta_arch
 from object_detection.protos import box_predictor_pb2
 from object_detection.protos import hyperparams_pb2
@@ -153,7 +155,9 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
                    predict_masks=False,
                    pad_to_max_dimension=None,
                    masks_are_class_agnostic=False,
-                   use_matmul_crop_and_resize=False):
+                   use_matmul_crop_and_resize=False,
+                   clip_anchors_to_image=False,
+                   use_matmul_gather_in_matcher=False):
 
     def image_resizer_fn(image, masks=None):
       """Fake image resizer function."""
@@ -186,6 +190,10 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         first_stage_anchor_scales,
         first_stage_anchor_aspect_ratios,
         anchor_stride=first_stage_anchor_strides)
+    first_stage_target_assigner = target_assigner.create_target_assigner(
+        'FasterRCNN',
+        'proposal',
+        use_matmul_gather=use_matmul_gather_in_matcher)
 
     fake_feature_extractor = FakeFasterRCNNFeatureExtractor()
 
@@ -211,7 +219,8 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
     first_stage_atrous_rate = 1
     first_stage_box_predictor_depth = 512
     first_stage_minibatch_size = 3
-    first_stage_positive_balance_fraction = .5
+    first_stage_sampler = sampler.BalancedPositiveNegativeSampler(
+        positive_fraction=0.5, is_static=False)
 
     first_stage_nms_score_threshold = -1.0
     first_stage_nms_iou_threshold = 1.0
@@ -230,9 +239,14 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
     """
     post_processing_config = post_processing_pb2.PostProcessing()
     text_format.Merge(post_processing_text_proto, post_processing_config)
+
+    second_stage_target_assigner = target_assigner.create_target_assigner(
+        'FasterRCNN', 'detection',
+        use_matmul_gather=use_matmul_gather_in_matcher)
     second_stage_non_max_suppression_fn, _ = post_processing_builder.build(
         post_processing_config)
-    second_stage_balance_fraction = 1.0
+    second_stage_sampler = sampler.BalancedPositiveNegativeSampler(
+        positive_fraction=1.0, is_static=False)
 
     second_stage_score_conversion_fn = tf.identity
     second_stage_localization_loss_weight = 1.0
@@ -261,6 +275,7 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         'feature_extractor': fake_feature_extractor,
         'number_of_stages': number_of_stages,
         'first_stage_anchor_generator': first_stage_anchor_generator,
+        'first_stage_target_assigner': first_stage_target_assigner,
         'first_stage_atrous_rate': first_stage_atrous_rate,
         'first_stage_box_predictor_arg_scope_fn':
         first_stage_box_predictor_arg_scope_fn,
@@ -268,8 +283,7 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         first_stage_box_predictor_kernel_size,
         'first_stage_box_predictor_depth': first_stage_box_predictor_depth,
         'first_stage_minibatch_size': first_stage_minibatch_size,
-        'first_stage_positive_balance_fraction':
-        first_stage_positive_balance_fraction,
+        'first_stage_sampler': first_stage_sampler,
         'first_stage_nms_score_threshold': first_stage_nms_score_threshold,
         'first_stage_nms_iou_threshold': first_stage_nms_iou_threshold,
         'first_stage_max_proposals': first_stage_max_proposals,
@@ -277,8 +291,9 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         first_stage_localization_loss_weight,
         'first_stage_objectness_loss_weight':
         first_stage_objectness_loss_weight,
+        'second_stage_target_assigner': second_stage_target_assigner,
         'second_stage_batch_size': second_stage_batch_size,
-        'second_stage_balance_fraction': second_stage_balance_fraction,
+        'second_stage_sampler': second_stage_sampler,
         'second_stage_non_max_suppression_fn':
         second_stage_non_max_suppression_fn,
         'second_stage_score_conversion_fn': second_stage_score_conversion_fn,
@@ -289,7 +304,8 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         'second_stage_classification_loss':
         second_stage_classification_loss,
         'hard_example_miner': hard_example_miner,
-        'use_matmul_crop_and_resize': use_matmul_crop_and_resize
+        'use_matmul_crop_and_resize': use_matmul_crop_and_resize,
+        'clip_anchors_to_image': clip_anchors_to_image
     }
 
     return self._get_model(
@@ -469,7 +485,8 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
         self.assertAllEqual(tensor_dict_out[key].shape, expected_shapes[key])
 
   def _test_predict_gives_correct_shapes_in_train_mode_both_stages(
-      self, use_matmul_crop_and_resize=False):
+      self, use_matmul_crop_and_resize=False,
+      clip_anchors_to_image=False):
     test_graph = tf.Graph()
     with test_graph.as_default():
       model = self._build_model(
@@ -477,7 +494,8 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
           number_of_stages=2,
           second_stage_batch_size=7,
           predict_masks=False,
-          use_matmul_crop_and_resize=use_matmul_crop_and_resize)
+          use_matmul_crop_and_resize=use_matmul_crop_and_resize,
+          clip_anchors_to_image=clip_anchors_to_image)
 
       batch_size = 2
       image_size = 10
@@ -547,6 +565,10 @@ class FasterRCNNMetaArchTestBase(tf.test.TestCase):
     self._test_predict_gives_correct_shapes_in_train_mode_both_stages(
         use_matmul_crop_and_resize=True)
 
+  def test_predict_gives_correct_shapes_in_train_mode_clip_anchors(self):
+    self._test_predict_gives_correct_shapes_in_train_mode_both_stages(
+        clip_anchors_to_image=True)
+
   def _test_postprocess_first_stage_only_inference_mode(
       self, pad_to_max_dimension=None):
     model = self._build_model(

research/object_detection/meta_architectures/rfcn_meta_arch.py

@@ -55,20 +55,22 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
                feature_extractor,
                number_of_stages,
                first_stage_anchor_generator,
+               first_stage_target_assigner,
                first_stage_atrous_rate,
                first_stage_box_predictor_arg_scope_fn,
                first_stage_box_predictor_kernel_size,
                first_stage_box_predictor_depth,
                first_stage_minibatch_size,
-               first_stage_positive_balance_fraction,
+               first_stage_sampler,
                first_stage_nms_score_threshold,
                first_stage_nms_iou_threshold,
                first_stage_max_proposals,
                first_stage_localization_loss_weight,
                first_stage_objectness_loss_weight,
+               second_stage_target_assigner,
                second_stage_rfcn_box_predictor,
                second_stage_batch_size,
-               second_stage_balance_fraction,
+               second_stage_sampler,
                second_stage_non_max_suppression_fn,
                second_stage_score_conversion_fn,
                second_stage_localization_loss_weight,
@@ -77,7 +79,8 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
                hard_example_miner,
                parallel_iterations=16,
                add_summaries=True,
-               use_matmul_crop_and_resize=False):
+               use_matmul_crop_and_resize=False,
+               clip_anchors_to_image=False):
     """RFCNMetaArch Constructor.
 
     Args:
@@ -97,6 +100,8 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
       first_stage_anchor_generator: An anchor_generator.AnchorGenerator object
         (note that currently we only support
         grid_anchor_generator.GridAnchorGenerator objects)
+      first_stage_target_assigner: Target assigner to use for first stage of
+        R-FCN (RPN).
       first_stage_atrous_rate: A single integer indicating the atrous rate for
         the single convolution op which is applied to the `rpn_features_to_crop`
         tensor to obtain a tensor to be used for box prediction. Some feature
@@ -116,8 +121,8 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         "batch size" refers to the number of anchors selected as contributing
         to the loss function for any given image within the image batch and is
         only called "batch_size" due to terminology from the Faster R-CNN paper.
-      first_stage_positive_balance_fraction: Fraction of positive examples
-        per image for the RPN. The recommended value for Faster RCNN is 0.5.
+      first_stage_sampler: The sampler for the boxes used to calculate the RPN
+        loss after the first stage.
       first_stage_nms_score_threshold: Score threshold for non max suppression
         for the Region Proposal Network (RPN).  This value is expected to be in
         [0, 1] as it is applied directly after a softmax transformation.  The
@@ -130,6 +135,10 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         Region Proposal Network (RPN).
       first_stage_localization_loss_weight: A float
       first_stage_objectness_loss_weight: A float
+      second_stage_target_assigner: Target assigner to use for second stage of
+        R-FCN. If the model is configured with multiple prediction heads, this
+        target assigner is used to generate targets for all heads (with the
+        correct `unmatched_class_label`).
       second_stage_rfcn_box_predictor: RFCN box predictor to use for
         second stage.
       second_stage_batch_size: The batch size used for computing the
@@ -137,9 +146,8 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         "batch size" refers to the number of proposals selected as contributing
         to the loss function for any given image within the image batch and is
         only called "batch_size" due to terminology from the Faster R-CNN paper.
-      second_stage_balance_fraction: Fraction of positive examples to use
-        per image for the box classifier. The recommended value for Faster RCNN
-        is 0.25.
+      second_stage_sampler: The sampler for the boxes used for second stage
+        box classifier.
       second_stage_non_max_suppression_fn: batch_multiclass_non_max_suppression
         callable that takes `boxes`, `scores`, optional `clip_window` and
         optional (kwarg) `mask` inputs (with all other inputs already set)
@@ -163,6 +171,9 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
       use_matmul_crop_and_resize: Force the use of matrix multiplication based
         crop and resize instead of standard tf.image.crop_and_resize while
         computing second stage input feature maps.
+      clip_anchors_to_image: The anchors generated are clip to the
+        window size without filtering the nonoverlapping anchors. This generates
+        a static number of anchors. This argument is unused.
 
     Raises:
       ValueError: If `second_stage_batch_size` > `first_stage_max_proposals`
@@ -178,12 +189,13 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         feature_extractor,
         number_of_stages,
         first_stage_anchor_generator,
+        first_stage_target_assigner,
         first_stage_atrous_rate,
         first_stage_box_predictor_arg_scope_fn,
         first_stage_box_predictor_kernel_size,
         first_stage_box_predictor_depth,
         first_stage_minibatch_size,
-        first_stage_positive_balance_fraction,
+        first_stage_sampler,
         first_stage_nms_score_threshold,
         first_stage_nms_iou_threshold,
         first_stage_max_proposals,
@@ -192,9 +204,10 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
         None,  # initial_crop_size is not used in R-FCN
         None,  # maxpool_kernel_size is not use in R-FCN
         None,  # maxpool_stride is not use in R-FCN
+        second_stage_target_assigner,
         None,  # fully_connected_box_predictor is not used in R-FCN.
         second_stage_batch_size,
-        second_stage_balance_fraction,
+        second_stage_sampler,
         second_stage_non_max_suppression_fn,
         second_stage_score_conversion_fn,
         second_stage_localization_loss_weight,
@@ -274,6 +287,11 @@ class RFCNMetaArch(faster_rcnn_meta_arch.FasterRCNNMetaArch):
             rpn_features,
             scope=self.second_stage_feature_extractor_scope))
 
+    if self._rfcn_box_predictor.is_keras_model:
+      box_predictions = self._rfcn_box_predictor(
+          [box_classifier_features],
+          proposal_boxes=proposal_boxes_normalized)
+    else:
       box_predictions = self._rfcn_box_predictor.predict(
           [box_classifier_features],
           num_predictions_per_location=[1],

research/object_detection/meta_architectures/ssd_meta_arch.py

@@ -35,7 +35,7 @@ slim = tf.contrib.slim
 
 
 class SSDFeatureExtractor(object):
-  """SSD Feature Extractor definition."""
+  """SSD Slim Feature Extractor definition."""
 
   def __init__(self,
                is_training,
@@ -77,6 +77,10 @@ class SSDFeatureExtractor(object):
     self._override_base_feature_extractor_hyperparams = (
         override_base_feature_extractor_hyperparams)
 
+  @property
+  def is_keras_model(self):
+    return False
+
   @abstractmethod
   def preprocess(self, resized_inputs):
     """Preprocesses images for feature extraction (minus image resizing).
@@ -113,6 +117,105 @@ class SSDFeatureExtractor(object):
     raise NotImplementedError
 
 
+class SSDKerasFeatureExtractor(tf.keras.Model):
+  """SSD Feature Extractor definition."""
+
+  def __init__(self,
+               is_training,
+               depth_multiplier,
+               min_depth,
+               pad_to_multiple,
+               conv_hyperparams_config,
+               freeze_batchnorm,
+               inplace_batchnorm_update,
+               use_explicit_padding=False,
+               use_depthwise=False,
+               override_base_feature_extractor_hyperparams=False):
+    """Constructor.
+
+    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_config: A hyperparams.proto object containing
+        convolution hyperparameters for the layers added on top of the
+        base feature extractor.
+      freeze_batchnorm: Whether to freeze batch norm parameters during
+        training or not. When training with a small batch size (e.g. 1), it is
+        desirable to freeze batch norm update and use pretrained batch norm
+        params.
+      inplace_batchnorm_update: Whether to update batch norm moving average
+        values inplace. When this is false train op must add a control
+        dependency on tf.graphkeys.UPDATE_OPS collection in order to update
+        batch norm statistics.
+      use_explicit_padding: Whether to use explicit padding when extracting
+        features. Default is False.
+      use_depthwise: Whether to use depthwise convolutions. Default is False.
+      override_base_feature_extractor_hyperparams: Whether to override
+        hyperparameters of the base feature extractor with the one from
+        `conv_hyperparams_config`.
+    """
+    super(SSDKerasFeatureExtractor, self).__init__()
+
+    self._is_training = is_training
+    self._depth_multiplier = depth_multiplier
+    self._min_depth = min_depth
+    self._pad_to_multiple = pad_to_multiple
+    self._conv_hyperparams_config = conv_hyperparams_config
+    self._freeze_batchnorm = freeze_batchnorm
+    self._inplace_batchnorm_update = inplace_batchnorm_update
+    self._use_explicit_padding = use_explicit_padding
+    self._use_depthwise = use_depthwise
+    self._override_base_feature_extractor_hyperparams = (
+        override_base_feature_extractor_hyperparams)
+
+  @property
+  def is_keras_model(self):
+    return True
+
+  @abstractmethod
+  def preprocess(self, resized_inputs):
+    """Preprocesses images for feature extraction (minus image resizing).
+
+    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.
+      true_image_shapes: int32 tensor of shape [batch, 3] where each row is
+        of the form [height, width, channels] indicating the shapes
+        of true images in the resized images, as resized images can be padded
+        with zeros.
+    """
+    raise NotImplementedError
+
+  @abstractmethod
+  def _extract_features(self, preprocessed_inputs):
+    """Extracts features from preprocessed inputs.
+
+    This function is responsible for extracting feature maps from preprocessed
+    images.
+
+    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]
+    """
+    raise NotImplementedError
+
+  # This overrides the keras.Model `call` method with the _extract_features
+  # method.
+  def call(self, inputs, **kwargs):
+    return self._extract_features(inputs)
+
+
 class SSDMetaArch(model.DetectionModel):
   """SSD Meta-architecture definition."""
 
@@ -211,10 +314,6 @@ class SSDMetaArch(model.DetectionModel):
     self._freeze_batchnorm = freeze_batchnorm
     self._inplace_batchnorm_update = inplace_batchnorm_update
 
-    # Needed for fine-tuning from classification checkpoints whose
-    # variables do not have the feature extractor scope.
-    self._extract_features_scope = 'FeatureExtractor'
-
     self._anchor_generator = anchor_generator
     self._box_predictor = box_predictor
 
@@ -224,21 +323,30 @@ class SSDMetaArch(model.DetectionModel):
     self._region_similarity_calculator = region_similarity_calculator
     self._add_background_class = add_background_class
 
+    # Needed for fine-tuning from classification checkpoints whose
+    # variables do not have the feature extractor scope.
+    if self._feature_extractor.is_keras_model:
+      # Keras feature extractors will have a name they implicitly use to scope.
+      # So, all contained variables are prefixed by this name.
+      # To load from classification checkpoints, need to filter out this name.
+      self._extract_features_scope = feature_extractor.name
+    else:
+      # Slim feature extractors get an explicit naming scope
+      self._extract_features_scope = 'FeatureExtractor'
+
     # TODO(jonathanhuang): handle agnostic mode
     # weights
-    unmatched_cls_target = None
-    unmatched_cls_target = tf.constant([1] + self.num_classes * [0],
+    self._unmatched_class_label = tf.constant([1] + self.num_classes * [0],
                                               tf.float32)
     if encode_background_as_zeros:
-      unmatched_cls_target = tf.constant((self.num_classes + 1) * [0],
+      self._unmatched_class_label = tf.constant((self.num_classes + 1) * [0],
                                                 tf.float32)
 
     self._target_assigner = target_assigner.TargetAssigner(
         self._region_similarity_calculator,
         self._matcher,
         self._box_coder,
-        negative_class_weight=negative_class_weight,
-        unmatched_cls_target=unmatched_cls_target)
+        negative_class_weight=negative_class_weight)
 
     self._classification_loss = classification_loss
     self._localization_loss = localization_loss
@@ -383,6 +491,9 @@ class SSDMetaArch(model.DetectionModel):
     """
     batchnorm_updates_collections = (None if self._inplace_batchnorm_update
                                      else tf.GraphKeys.UPDATE_OPS)
+    if self._feature_extractor.is_keras_model:
+      feature_maps = self._feature_extractor(preprocessed_inputs)
+    else:
       with slim.arg_scope([slim.batch_norm],
                           is_training=(self._is_training and
                                        not self._freeze_batchnorm),
@@ -391,6 +502,7 @@ class SSDMetaArch(model.DetectionModel):
                                [preprocessed_inputs]):
           feature_maps = self._feature_extractor.extract_features(
               preprocessed_inputs)
+
     feature_map_spatial_dims = self._get_feature_map_spatial_dims(
         feature_maps)
     image_shape = shape_utils.combined_static_and_dynamic_shape(
@@ -400,8 +512,16 @@ class SSDMetaArch(model.DetectionModel):
             feature_map_spatial_dims,
             im_height=image_shape[1],
             im_width=image_shape[2]))
+    if self._box_predictor.is_keras_model:
+      prediction_dict = self._box_predictor(feature_maps)
+    else:
+      with slim.arg_scope([slim.batch_norm],
+                          is_training=(self._is_training and
+                                       not self._freeze_batchnorm),
+                          updates_collections=batchnorm_updates_collections):
         prediction_dict = self._box_predictor.predict(
             feature_maps, self._anchor_generator.num_anchors_per_location())
+
     box_encodings = tf.concat(prediction_dict['box_encodings'], axis=1)
     if box_encodings.shape.ndims == 4 and box_encodings.shape[2] == 1:
       box_encodings = tf.squeeze(box_encodings, axis=2)
@@ -710,7 +830,8 @@ class SSDMetaArch(model.DetectionModel):
         boxlist.add_field(fields.BoxListFields.keypoints, keypoints)
     return target_assigner.batch_assign_targets(
         self._target_assigner, self.anchors, groundtruth_boxlists,
-        groundtruth_classes_with_background_list, groundtruth_weights_list)
+        groundtruth_classes_with_background_list, self._unmatched_class_label,
+        groundtruth_weights_list)
 
   def _summarize_target_assignment(self, groundtruth_boxes_list, match_list):
     """Creates tensorflow summaries for the input boxes and anchors.
@@ -872,3 +993,4 @@ class SSDMetaArch(model.DetectionModel):
           variables_to_restore[var_name] = variable
 
     return variables_to_restore
+

research/object_detection/meta_architectures/ssd_meta_arch_test.py

@@ -15,6 +15,8 @@
 
 """Tests for object_detection.meta_architectures.ssd_meta_arch."""
 import functools
+from absl.testing import parameterized
+
 import numpy as np
 import tensorflow as tf
 
@@ -29,6 +31,7 @@ from object_detection.utils import test_case
 from object_detection.utils import test_utils
 
 slim = tf.contrib.slim
+keras = tf.keras.layers
 
 
 class FakeSSDFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
@@ -51,6 +54,30 @@ class FakeSSDFeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       return [features]
 
 
+class FakeSSDKerasFeatureExtractor(ssd_meta_arch.SSDKerasFeatureExtractor):
+
+  def __init__(self):
+    with tf.name_scope('mock_model'):
+      super(FakeSSDKerasFeatureExtractor, self).__init__(
+          is_training=True,
+          depth_multiplier=0,
+          min_depth=0,
+          pad_to_multiple=1,
+          conv_hyperparams_config=None,
+          freeze_batchnorm=False,
+          inplace_batchnorm_update=False,
+      )
+
+      self._conv = keras.Conv2D(filters=32, kernel_size=1, name='layer1')
+
+  def preprocess(self, resized_inputs):
+    return tf.identity(resized_inputs)
+
+  def _extract_features(self, preprocessed_inputs, **kwargs):
+    with tf.name_scope('mock_model'):
+      return [self._conv(preprocessed_inputs)]
+
+
 class MockAnchorGenerator2x2(anchor_generator.AnchorGenerator):
   """Sets up a simple 2x2 anchor grid on the unit square."""
 
@@ -79,19 +106,31 @@ def _get_value_for_matching_key(dictionary, suffix):
   raise ValueError('key not found {}'.format(suffix))
 
 
-class SsdMetaArchTest(test_case.TestCase):
+@parameterized.parameters(
+    {'use_keras': False},
+    {'use_keras': True},
+)
+class SsdMetaArchTest(test_case.TestCase, parameterized.TestCase):
 
   def _create_model(self,
                     apply_hard_mining=True,
                     normalize_loc_loss_by_codesize=False,
                     add_background_class=True,
-                    random_example_sampling=False):
+                    random_example_sampling=False,
+                    use_keras=False):
     is_training = False
     num_classes = 1
     mock_anchor_generator = MockAnchorGenerator2x2()
+    if use_keras:
+      mock_box_predictor = test_utils.MockKerasBoxPredictor(
+          is_training, num_classes)
+    else:
       mock_box_predictor = test_utils.MockBoxPredictor(
           is_training, num_classes)
     mock_box_coder = test_utils.MockBoxCoder()
+    if use_keras:
+      fake_feature_extractor = FakeSSDKerasFeatureExtractor()
+    else:
       fake_feature_extractor = FakeSSDFeatureExtractor()
     mock_matcher = test_utils.MockMatcher()
     region_similarity_calculator = sim_calc.IouSimilarity()
@@ -152,25 +191,26 @@ class SsdMetaArchTest(test_case.TestCase):
         random_example_sampler=random_example_sampler)
     return model, num_classes, mock_anchor_generator.num_anchors(), code_size
 
-  def test_preprocess_preserves_shapes_with_dynamic_input_image(self):
+  def test_preprocess_preserves_shapes_with_dynamic_input_image(
+      self, use_keras):
     image_shapes = [(3, None, None, 3),
                     (None, 10, 10, 3),
                     (None, None, None, 3)]
-    model, _, _, _ = self._create_model()
+    model, _, _, _ = self._create_model(use_keras=use_keras)
     for image_shape in image_shapes:
       image_placeholder = tf.placeholder(tf.float32, shape=image_shape)
       preprocessed_inputs, _ = model.preprocess(image_placeholder)
       self.assertAllEqual(preprocessed_inputs.shape.as_list(), image_shape)
 
-  def test_preprocess_preserves_shape_with_static_input_image(self):
+  def test_preprocess_preserves_shape_with_static_input_image(self, use_keras):
     def graph_fn(input_image):
-      model, _, _, _ = self._create_model()
+      model, _, _, _ = self._create_model(use_keras=use_keras)
       return model.preprocess(input_image)
     input_image = np.random.rand(2, 3, 3, 3).astype(np.float32)
     preprocessed_inputs, _ = self.execute(graph_fn, [input_image])
     self.assertAllEqual(preprocessed_inputs.shape, [2, 3, 3, 3])
 
-  def test_predict_result_shapes_on_image_with_dynamic_shape(self):
+  def test_predict_result_shapes_on_image_with_dynamic_shape(self, use_keras):
     batch_size = 3
     image_size = 2
     input_shapes = [(None, image_size, image_size, 3),
@@ -180,16 +220,17 @@ class SsdMetaArchTest(test_case.TestCase):
     for input_shape in input_shapes:
       tf_graph = tf.Graph()
       with tf_graph.as_default():
-        model, num_classes, num_anchors, code_size = self._create_model()
+        model, num_classes, num_anchors, code_size = self._create_model(
+            use_keras=use_keras)
         preprocessed_input_placeholder = tf.placeholder(tf.float32,
                                                         shape=input_shape)
         prediction_dict = model.predict(
             preprocessed_input_placeholder, true_image_shapes=None)
 
-        self.assertTrue('box_encodings' in prediction_dict)
-        self.assertTrue('class_predictions_with_background' in prediction_dict)
-        self.assertTrue('feature_maps' in prediction_dict)
-        self.assertTrue('anchors' in prediction_dict)
+        self.assertIn('box_encodings', prediction_dict)
+        self.assertIn('class_predictions_with_background', prediction_dict)
+        self.assertIn('feature_maps', prediction_dict)
+        self.assertIn('anchors', prediction_dict)
 
         init_op = tf.global_variables_initializer()
       with self.test_session(graph=tf_graph) as sess:
@@ -210,10 +251,11 @@ class SsdMetaArchTest(test_case.TestCase):
           prediction_out['class_predictions_with_background'].shape,
           expected_class_predictions_with_background_shape_out)
 
-  def test_predict_result_shapes_on_image_with_static_shape(self):
+  def test_predict_result_shapes_on_image_with_static_shape(self, use_keras):
 
     with tf.Graph().as_default():
-      _, num_classes, num_anchors, code_size = self._create_model()
+      _, num_classes, num_anchors, code_size = self._create_model(
+          use_keras=use_keras)
 
     def graph_fn(input_image):
       model, _, _, _ = self._create_model()
@@ -235,7 +277,7 @@ class SsdMetaArchTest(test_case.TestCase):
     self.assertAllEqual(class_predictions.shape,
                         expected_class_predictions_shape)
 
-  def test_postprocess_results_are_correct(self):
+  def test_postprocess_results_are_correct(self, use_keras):
     batch_size = 2
     image_size = 2
     input_shapes = [(batch_size, image_size, image_size, 3),
@@ -266,17 +308,17 @@ class SsdMetaArchTest(test_case.TestCase):
     for input_shape in input_shapes:
       tf_graph = tf.Graph()
       with tf_graph.as_default():
-        model, _, _, _ = self._create_model()
+        model, _, _, _ = self._create_model(use_keras=use_keras)
         input_placeholder = tf.placeholder(tf.float32, shape=input_shape)
         preprocessed_inputs, true_image_shapes = model.preprocess(
             input_placeholder)
         prediction_dict = model.predict(preprocessed_inputs,
                                         true_image_shapes)
         detections = model.postprocess(prediction_dict, true_image_shapes)
-        self.assertTrue('detection_boxes' in detections)
-        self.assertTrue('detection_scores' in detections)
-        self.assertTrue('detection_classes' in detections)
-        self.assertTrue('num_detections' in detections)
+        self.assertIn('detection_boxes', detections)
+        self.assertIn('detection_scores', detections)
+        self.assertIn('detection_classes', detections)
+        self.assertIn('num_detections', detections)
         init_op = tf.global_variables_initializer()
       with self.test_session(graph=tf_graph) as sess:
         sess.run(init_op)
@@ -295,10 +337,10 @@ class SsdMetaArchTest(test_case.TestCase):
       self.assertAllClose(detections_out['num_detections'],
                           expected_num_detections)
 
-  def test_loss_results_are_correct(self):
+  def test_loss_results_are_correct(self, use_keras):
 
     with tf.Graph().as_default():
-      _, num_classes, num_anchors, _ = self._create_model()
+      _, num_classes, num_anchors, _ = self._create_model(use_keras=use_keras)
     def graph_fn(preprocessed_tensor, groundtruth_boxes1, groundtruth_boxes2,
                  groundtruth_classes1, groundtruth_classes2):
       groundtruth_boxes_list = [groundtruth_boxes1, groundtruth_boxes2]
@@ -331,16 +373,18 @@ class SsdMetaArchTest(test_case.TestCase):
     self.assertAllClose(localization_loss, expected_localization_loss)
     self.assertAllClose(classification_loss, expected_classification_loss)
 
-  def test_loss_results_are_correct_with_normalize_by_codesize_true(self):
+  def test_loss_results_are_correct_with_normalize_by_codesize_true(
+      self, use_keras):
 
     with tf.Graph().as_default():
-      _, _, _, _ = self._create_model()
+      _, _, _, _ = self._create_model(use_keras=use_keras)
     def graph_fn(preprocessed_tensor, groundtruth_boxes1, groundtruth_boxes2,
                  groundtruth_classes1, groundtruth_classes2):
       groundtruth_boxes_list = [groundtruth_boxes1, groundtruth_boxes2]
       groundtruth_classes_list = [groundtruth_classes1, groundtruth_classes2]
       model, _, _, _ = self._create_model(apply_hard_mining=False,
-                                          normalize_loc_loss_by_codesize=True)
+                                          normalize_loc_loss_by_codesize=True,
+                                          use_keras=use_keras)
       model.provide_groundtruth(groundtruth_boxes_list,
                                 groundtruth_classes_list)
       prediction_dict = model.predict(preprocessed_tensor,
@@ -362,10 +406,10 @@ class SsdMetaArchTest(test_case.TestCase):
                                                 groundtruth_classes2])
     self.assertAllClose(localization_loss, expected_localization_loss)
 
-  def test_loss_results_are_correct_with_hard_example_mining(self):
+  def test_loss_results_are_correct_with_hard_example_mining(self, use_keras):
 
     with tf.Graph().as_default():
-      _, num_classes, num_anchors, _ = self._create_model()
+      _, num_classes, num_anchors, _ = self._create_model(use_keras=use_keras)
     def graph_fn(preprocessed_tensor, groundtruth_boxes1, groundtruth_boxes2,
                  groundtruth_classes1, groundtruth_classes2):
       groundtruth_boxes_list = [groundtruth_boxes1, groundtruth_boxes2]
@@ -397,18 +441,20 @@ class SsdMetaArchTest(test_case.TestCase):
     self.assertAllClose(localization_loss, expected_localization_loss)
     self.assertAllClose(classification_loss, expected_classification_loss)
 
-  def test_loss_results_are_correct_without_add_background_class(self):
+  def test_loss_results_are_correct_without_add_background_class(
+      self, use_keras):
 
     with tf.Graph().as_default():
       _, num_classes, num_anchors, _ = self._create_model(
-          add_background_class=False)
+          add_background_class=False, use_keras=use_keras)
 
     def graph_fn(preprocessed_tensor, groundtruth_boxes1, groundtruth_boxes2,
                  groundtruth_classes1, groundtruth_classes2):
       groundtruth_boxes_list = [groundtruth_boxes1, groundtruth_boxes2]
       groundtruth_classes_list = [groundtruth_classes1, groundtruth_classes2]
       model, _, _, _ = self._create_model(
-          apply_hard_mining=False, add_background_class=False)
+          apply_hard_mining=False, add_background_class=False,
+          use_keras=use_keras)
       model.provide_groundtruth(groundtruth_boxes_list,
                                 groundtruth_classes_list)
       prediction_dict = model.predict(
@@ -434,8 +480,8 @@ class SsdMetaArchTest(test_case.TestCase):
     self.assertAllClose(localization_loss, expected_localization_loss)
     self.assertAllClose(classification_loss, expected_classification_loss)
 
-  def test_restore_map_for_detection_ckpt(self):
-    model, _, _, _ = self._create_model()
+  def test_restore_map_for_detection_ckpt(self, use_keras):
+    model, _, _, _ = self._create_model(use_keras=use_keras)
     model.predict(tf.constant(np.array([[[[0, 0], [1, 1]], [[1, 0], [0, 1]]]],
                                        dtype=np.float32)),
                   true_image_shapes=None)
@@ -454,13 +500,21 @@ class SsdMetaArchTest(test_case.TestCase):
       for var in sess.run(tf.report_uninitialized_variables()):
         self.assertNotIn('FeatureExtractor', var)
 
-  def test_restore_map_for_classification_ckpt(self):
+  def test_restore_map_for_classification_ckpt(self, use_keras):
     # Define mock tensorflow classification graph and save variables.
     test_graph_classification = tf.Graph()
     with test_graph_classification.as_default():
       image = tf.placeholder(dtype=tf.float32, shape=[1, 20, 20, 3])
+      if use_keras:
+        with tf.name_scope('mock_model'):
+          layer_one = keras.Conv2D(32, kernel_size=1, name='layer1')
+          net = layer_one(image)
+          layer_two = keras.Conv2D(3, kernel_size=1, name='layer2')
+          layer_two(net)
+      else:
         with tf.variable_scope('mock_model'):
-        net = slim.conv2d(image, num_outputs=32, kernel_size=1, scope='layer1')
+          net = slim.conv2d(image, num_outputs=32, kernel_size=1,
+                            scope='layer1')
           slim.conv2d(net, num_outputs=3, kernel_size=1, scope='layer2')
 
       init_op = tf.global_variables_initializer()
@@ -474,7 +528,7 @@ class SsdMetaArchTest(test_case.TestCase):
     # classification checkpoint.
     test_graph_detection = tf.Graph()
     with test_graph_detection.as_default():
-      model, _, _, _ = self._create_model()
+      model, _, _, _ = self._create_model(use_keras=use_keras)
       inputs_shape = [2, 2, 2, 3]
       inputs = tf.to_float(tf.random_uniform(
           inputs_shape, minval=0, maxval=255, dtype=tf.int32))
@@ -491,10 +545,10 @@ class SsdMetaArchTest(test_case.TestCase):
         for var in sess.run(tf.report_uninitialized_variables()):
           self.assertNotIn('FeatureExtractor', var)
 
-  def test_load_all_det_checkpoint_vars(self):
+  def test_load_all_det_checkpoint_vars(self, use_keras):
     test_graph_detection = tf.Graph()
     with test_graph_detection.as_default():
-      model, _, _, _ = self._create_model()
+      model, _, _, _ = self._create_model(use_keras=use_keras)
       inputs_shape = [2, 2, 2, 3]
       inputs = tf.to_float(
           tf.random_uniform(inputs_shape, minval=0, maxval=255, dtype=tf.int32))
@@ -508,18 +562,22 @@ class SsdMetaArchTest(test_case.TestCase):
       self.assertIsInstance(var_map, dict)
       self.assertIn('another_variable', var_map)
 
-  def test_loss_results_are_correct_with_random_example_sampling(self):
+  def test_loss_results_are_correct_with_random_example_sampling(
+      self,
+      use_keras):
 
     with tf.Graph().as_default():
       _, num_classes, num_anchors, _ = self._create_model(
-          random_example_sampling=True)
+          random_example_sampling=True,
+          use_keras=use_keras)
     print num_classes, num_anchors
 
     def graph_fn(preprocessed_tensor, groundtruth_boxes1, groundtruth_boxes2,
                  groundtruth_classes1, groundtruth_classes2):
       groundtruth_boxes_list = [groundtruth_boxes1, groundtruth_boxes2]
       groundtruth_classes_list = [groundtruth_classes1, groundtruth_classes2]
-      model, _, _, _ = self._create_model(random_example_sampling=True)
+      model, _, _, _ = self._create_model(random_example_sampling=True,
+                                          use_keras=use_keras)
       model.provide_groundtruth(groundtruth_boxes_list,
                                 groundtruth_classes_list)
       prediction_dict = model.predict(

research/object_detection/model_lib.py

@@ -202,6 +202,10 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False):
     params = params or {}
     total_loss, train_op, detections, export_outputs = None, None, None, None
     is_training = mode == tf.estimator.ModeKeys.TRAIN
+
+    # Make sure to set the Keras learning phase. True during training,
+    # False for inference.
+    tf.keras.backend.set_learning_phase(is_training)
     detection_model = detection_model_fn(is_training=is_training,
                                          add_summaries=(not use_tpu))
     scaffold_fn = None
@@ -279,7 +283,7 @@ def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False):
     if mode in (tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL):
       losses_dict = detection_model.loss(
           prediction_dict, features[fields.InputDataFields.true_image_shape])
-      losses = [loss_tensor for loss_tensor in losses_dict.itervalues()]
+      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)

research/object_detection/models/feature_map_generators.py

@@ -221,8 +221,8 @@ def fpn_top_down_feature_maps(image_features, depth, scope=None):
             depth, [3, 3],
             scope='smoothing_%d' % (level + 1)))
         output_feature_map_keys.append('top_down_%s' % image_features[level][0])
-      return collections.OrderedDict(
-          reversed(zip(output_feature_map_keys, output_feature_maps_list)))
+      return collections.OrderedDict(reversed(
+          list(zip(output_feature_map_keys, output_feature_maps_list))))
 
 
 def pooling_pyramid_feature_maps(base_feature_map_depth, num_layers,
@@ -288,4 +288,3 @@ def pooling_pyramid_feature_maps(base_feature_map_depth, num_layers,
       feature_maps.append(feature_map)
   return collections.OrderedDict(
       [(x, y) for (x, y) in zip(feature_map_keys, feature_maps)])
-

research/object_detection/models/ssd_inception_v2_feature_extractor.py

@@ -61,8 +61,15 @@ class SSDInceptionV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       ValueError: If `override_base_feature_extractor_hyperparams` is False.
     """
     super(SSDInceptionV2FeatureExtractor, self).__init__(
-        is_training, depth_multiplier, min_depth, pad_to_multiple,
-        conv_hyperparams_fn, reuse_weights, use_explicit_padding, use_depthwise,
+        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)
     if not self._override_base_feature_extractor_hyperparams:
       raise ValueError('SSD Inception V2 feature extractor always uses'

research/object_detection/models/ssd_inception_v3_feature_extractor.py

@@ -61,8 +61,15 @@ class SSDInceptionV3FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
       ValueError: If `override_base_feature_extractor_hyperparams` is False.
     """
     super(SSDInceptionV3FeatureExtractor, self).__init__(
-        is_training, depth_multiplier, min_depth, pad_to_multiple,
-        conv_hyperparams_fn, reuse_weights, use_explicit_padding, use_depthwise,
+        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)
 
     if not self._override_base_feature_extractor_hyperparams: