GitLab

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

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

    1. Replace tf.boolean_mask with tf.where

--
212282646  by Zhichao Lu:

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

--
212142989  by Zhichao Lu:

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

--
212136935  by Zhichao Lu:

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

--
211907984  by Zhichao Lu:

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

--
211858098  by Zhichao Lu:

    Change the implementation of merge_boxes_with_multiple_labels.

--
211843915  by Zhichao Lu:

    Add Mobilenet v2 + FPN support.

--
211655076  by Zhichao Lu:

    Bug fix for generic keys in config overrides

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

    This change also introduces testing for all config overrides.

--
211157501  by Zhichao Lu:

    Make the locally-modified conv defs a copy.

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

--
211112813  by Zhichao Lu:

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

--
211109571  by Zhichao Lu:

    A test decorator.

--
210747685  by Zhichao Lu:

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

--
210723882  by Zhichao Lu:

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

--
210673675  by Zhichao Lu:

    Internal change.

--
210546590  by Zhichao Lu:

    Internal change.

--
210529752  by Zhichao Lu:

    Support batched inputs with ops.matmul_crop_and_resize.

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

--
210485912  by Zhichao Lu:

    Fix TensorFlow version check in object_detection_tutorial.ipynb

--
210484076  by Zhichao Lu:

    Reduce TPU memory required for single image matmul_crop_and_resize.

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

--
210468361  by Zhichao Lu:

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

--
210100253  by Zhichao Lu:

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

--
209995842  by Zhichao Lu:

    Fix a bug which prevents variable sharing in Faster RCNN.

--
209965526  by Zhichao Lu:

    Add support for enabling export_to_tpu through the estimator.

--
209946440  by Zhichao Lu:

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

--
209888003  by Zhichao Lu:

    Implement function to handle data where source_id is not set.

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

--
209842134  by Zhichao Lu:

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

--
209768066  by Zhichao Lu:

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

--
209722556  by Zhichao Lu:

    Remove dead code.

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

--
209701861  by Zhichao Lu:

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

--
209697893  by Zhichao Lu:

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

--
209694652  by Zhichao Lu:

    Sample boxes by jittering around the currently given boxes.

--
209550300  by Zhichao Lu:

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

--
209490273  by Zhichao Lu:

    Check result_dict type before accessing image_id via key.

--
209442529  by Zhichao Lu:

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

--
208941150  by Zhichao Lu:

    Adding the capability of exporting the results in json format.

--
208888798  by Zhichao Lu:

    Fixes wrong dictionary key for num_det_boxes_per_image.

--
208873549  by Zhichao Lu:

    Reduce the number of HLO ops created by matmul_crop_and_resize.

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

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

--
208844315  by Zhichao Lu:

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

--
208731380  by Zhichao Lu:

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

--
208699405  by Zhichao Lu:

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

--
208557208  by Zhichao Lu:

    Add TPU tests for Faster R-CNN Meta arch.

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

--
208499278  by Zhichao Lu:

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

--
208391694  by Zhichao Lu:

    Updating visualization tool to produce multiple evaluation images.

--
208275961  by Zhichao Lu:

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

--
208275585  by Zhichao Lu:

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

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

--
208263792  by Zhichao Lu:

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

--
208000218  by Zhichao Lu:

    Make FasterRCNN choose static shape operations only in training mode.

--
207997797  by Zhichao Lu:

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

--
207993460  by Zhichao Lu:

    Include FGVC detection models in model zoo.

--
207971213  by Zhichao Lu:

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

--
207961187  by Zhichao Lu:

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

--
207960608  by Zhichao Lu:

    Internal Change.

--
207927015  by Zhichao Lu:

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

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

--
207809668  by Zhichao Lu:

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

--
207808651  by Zhichao Lu:

    Fix the static balanced positive negative sampler's TPU tests

--
207798658  by Zhichao Lu:

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

--
207796470  by Zhichao Lu:

    Make slim endpoints visible in FasterRCNNMetaArch.

--
207787053  by Zhichao Lu:

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

--

PiperOrigin-RevId: 212389173

* Fix detection model zoo typo.

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

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

PiperOrigin-RevId: 212672223

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

PiperOrigin-RevId: 212679112

* Fix detection model zoo typo.

PiperOrigin-RevId: 212715692

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

PiperOrigin-RevId: 212723242

* Support class confidences in merge boxes with multiple labels.

PiperOrigin-RevId: 212884998

* Creates multiple eval specs for object detection.

PiperOrigin-RevId: 212894556

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

PiperOrigin-RevId: 213030087

* Enable bfloat16 training for object detection models.

PiperOrigin-RevId: 213053547

* Skip padding op when unnecessary.

PiperOrigin-RevId: 213065869

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

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

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

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

PiperOrigin-RevId: 213345395

* Add release note for iNaturalist Species trained models.

PiperOrigin-RevId: 213347179

* Fix the bug of uninitialized gt_is_crowd_list variable.

PiperOrigin-RevId: 213364858

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

PiperOrigin-RevId: 213554260

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

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

--
207758641  by Zhichao Lu:

    Require tensorflow version 1.9+ for running object detection API.

--
207641470  by Zhichao Lu:

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

--
207621728  by Zhichao Lu:

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

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

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

--
207610524  by Zhichao Lu:

    Update anchor generators and box predictors for python3 compatibility.

--
207585122  by Zhichao Lu:

    Refactoring convolutional box predictor into separate prediction heads.

--
207549305  by Zhichao Lu:

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

--
207336575  by Zhichao Lu:

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

--
207327862  by Zhichao Lu:

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

--
207323154  by Zhichao Lu:

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

--
207301109  by Zhichao Lu:

    Integrating expected_classification_loss_under_sampling op as an option in the ssd_meta_arch

--
207286221  by Zhichao Lu:

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

--
207231739  by Zhichao Lu:

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

--
207206356  by Zhichao Lu:

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

--
207135930  by Zhichao Lu:

    Internal change.

--
206862541  by Zhichao Lu:

    Do not unpad the outputs from batch_non_max_suppression before sampling.

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

--

PiperOrigin-RevId: 207771702

* Remove unused doc.

* 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.

Add option to override base feature extractor hyperparams in SSD models. This would allow us to use the same set of hyperparams for the complete feature extractor (base + new layers) if desired.

PiperOrigin-RevId: 191787921

191649512  by Zhichao Lu:

    Introduce two parameters in ssd.proto - freeze_batchnorm, inplace_batchnorm_update - and set up slim arg_scopes in ssd_meta_arch.py such that applies it to all batchnorm ops in the predict() method.

    This centralizes the control of freezing and doing inplace batchnorm updates.

--
191620303  by Zhichao Lu:

    Modifications to the preprocessor to support multiclass scores

--
191610773  by Zhichao Lu:

    Adding multiclass_scores to InputDataFields and adding padding for multiclass_scores.

--
191595011  by Zhichao Lu:

    Contains implementation of the detection metric for the Open Images Challenge.

--
191449408  by Zhichao Lu:

    Change hyperparams_builder to return a callable so the users can inherit values from outer arg_scopes. This allows us to easily set batch_norm parameters like "is_training" and "inplace_batchnorm_update" for all feature extractors from the base class and propagate it correctly to the nested scopes.

--
191437008  by Zhichao Lu:

    Contains implementation of the Recall@N and MedianRank@N metrics.

--
191385254  by Zhichao Lu:

    Add config rewrite flag to eval.py

--
191382500  by Zhichao Lu:

    Fix bug for config_util.

--

PiperOrigin-RevId: 191649512

187187978  by Zhichao Lu:

    Only updating hyperparameters if they have non-null values.

--
187097690  by Zhichao Lu:

    Rewrite some conditions a bit more clearly.

--
187085190  by Zhichao Lu:

    More informative error message.

--
186935376  by Zhichao Lu:

    Added option to evaluator.evaluate to use custom evaluator objects.

--
186808249  by Zhichao Lu:

    Fix documentation re: number of stages.

--
186775014  by Zhichao Lu:

    Change anchor generator interface to return a list of BoxLists containing anchors for different feature map layers.

--
186729028  by Zhichao Lu:

    Minor fixes to object detection.

--
186723716  by Zhichao Lu:

    Fix tf_example_decoder.py initailization issue.

--
186668505  by Zhichao Lu:

    Remove unused import.

--
186475361  by Zhichao Lu:

    Update the box predictor interface to return list of predictions - one from each feature map - instead of stacking them into one large tensor.

--
186410844  by Zhichao Lu:

    Fix PythonPath Dependencies.

--
186365384  by Zhichao Lu:

    Made some of the functions in exporter public so they can be reused.

--
186341438  by Zhichao Lu:

    Re-introducing check that label-map-path must be a valid (non-empty) string prior to overwriting pipeline config.

--
186036984  by Zhichao Lu:

    Adding default hyperparameters and allowing for overriding them via flags.

--
186026006  by Zhichao Lu:

    Strip `eval_` prefix from name argument give to TPUEstimator.evaluate since it adds the same prefix internally.

--
186016042  by Zhichao Lu:

    Add an option to evaluate models on training data.

--
185944986  by Zhichao Lu:

    let _update_label_map_path go through even if the path is empty

--
185860781  by Zhichao Lu:

    Add random normal initializer option to hyperparams builder.

    Scale the regression losses outside of the box encoder by adjusting huber loss delta and regression loss weight.

--
185846325  by Zhichao Lu:

    Add an option to normalize localization loss by the code size(number of box coordinates) in SSD Meta architecture.

--
185761217  by Zhichao Lu:

    Change multiscale_grid_anchor_generator to return anchors in normalized coordinates by default and add option to configure it.

    In SSD meta architecture, TargetAssigner operates in normalized coordinate space (i.e, groundtruth boxes are in normalized coordinates) hence we need the option to generate anchors in normalized coordinates.

--
185747733  by Zhichao Lu:

    Change the smooth L1 localization implementationt to use tf.losses.huber_loss and expose the delta parameter in the proto.

--
185715309  by Zhichao Lu:

    Obviates the need for prepadding on mobilenet v1 and v2 for fully convolutional models.

--
185685695  by Zhichao Lu:

    Fix manual stepping schedule to return first rate when there are no boundaries

--
185621650  by Zhichao Lu:

    Added target assigner proto for configuring negative class weights.

--

PiperOrigin-RevId: 187187978

This works around a bug in earlier proto versions
that automatically infer these values to be integer
instead of float.

For details see our paper:
"Speed/accuracy trade-offs for modern convolutional object detectors."
Huang J, Rathod V, Sun C, Zhu M, Korattikara A, Fathi A, Fischer I,
Wojna Z, Song Y, Guadarrama S, Murphy K, CVPR 2017
https://arxiv.org/abs/1611.10012