Merge branch 'tf2' into 'main'

tf2 detection

See merge request dcutoolkit/deeplearing/dlexamples_new!2

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/distributed_executer.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Interface to run mask rcnn model in different distributed strategies."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import abc
+import os
+import six
+
+import math
+
+import multiprocessing
+
+import tensorflow as tf
+
+from mask_rcnn.utils.logging_formatter import logging
+from mask_rcnn.utils.distributed_utils import MPI_is_distributed
+from mask_rcnn.utils.distributed_utils import MPI_local_rank
+from mask_rcnn.utils.distributed_utils import MPI_rank
+
+from mask_rcnn.hooks.logging_hook import AutoLoggingHook
+
+from mask_rcnn.utils.lazy_imports import LazyImport
+hvd = LazyImport("horovod.tensorflow")
+
+from tensorflow.core.protobuf import rewriter_config_pb2
+
+from mask_rcnn import evaluation
+from mask_rcnn.hyperparameters import params_io
+from mask_rcnn.hooks import CheckpointSaverHook
+from mask_rcnn.hooks import PretrainedWeightsLoadingHook
+
+
+def get_training_hooks(mode, model_dir, checkpoint_path=None, skip_checkpoint_variables=None):
+
+    assert mode in ('train', 'eval')
+
+    training_hooks = [
+        AutoLoggingHook(
+            # log_every_n_steps=RUNNING_CONFIG.display_step,
+            log_every_n_steps=5 if "NGC_JOB_ID" not in os.environ else 100,
+            # warmup_steps=RUNNING_CONFIG.warmup_steps,
+            warmup_steps=100,
+            is_training=True
+        )
+    ]
+
+    if not MPI_is_distributed() or MPI_rank() == 0:
+        training_hooks.append(PretrainedWeightsLoadingHook(
+            prefix="resnet50/",
+            checkpoint_path=checkpoint_path,
+            skip_variables_regex=skip_checkpoint_variables
+        ))
+
+    if MPI_is_distributed() and mode == "train":
+        training_hooks.append(hvd.BroadcastGlobalVariablesHook(root_rank=0))
+
+    if not MPI_is_distributed() or MPI_rank() == 0:
+        training_hooks.append(CheckpointSaverHook(
+            checkpoint_dir=model_dir,
+            checkpoint_basename="model.ckpt"
+        ))
+
+    return training_hooks
+
+
+@six.add_metaclass(abc.ABCMeta)
+class BaseExecuter(object):
+  """Interface to run Mask RCNN model in TPUs/GPUs.
+
+  Arguments:
+    flags: FLAGS object passed from the user.
+    model_config: Model configuration needed to run distribution strategy.
+    model_fn: Model function to be passed to Estimator.
+  """
+
+  def __init__(self, runtime_config, model_fn):
+
+    self._runtime_config = runtime_config
+    self._model_fn = model_fn
+
+    os.environ['CUDA_CACHE_DISABLE'] = '0'
+
+    os.environ['TF_USE_CUDNN_BATCHNORM_SPATIAL_PERSISTENT'] = '1'
+
+    os.environ['TF_ADJUST_HUE_FUSED'] = '1'
+    os.environ['TF_ADJUST_SATURATION_FUSED'] = '1'
+    os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'
+
+    os.environ['TF_AUTOTUNE_THRESHOLD'] = '2'
+
+  @staticmethod
+  def _get_session_config(mode, use_xla, use_amp, use_tf_distributed=False, allow_xla_at_inference=False):
+
+      assert mode in ('train', 'eval')
+
+
+      rewrite_options = rewriter_config_pb2.RewriterConfig(
+          # arithmetic_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+          # arithmetic_optimization=rewriter_config_pb2.RewriterConfig.ON,
+
+          # constant_folding=rewriter_config_pb2.RewriterConfig.OFF,
+          # constant_folding=rewriter_config_pb2.RewriterConfig.ON,           # TO TEST
+
+          # debug_stripper=rewriter_config_pb2.RewriterConfig.OFF,
+          # debug_stripper=rewriter_config_pb2.RewriterConfig.ON,           # TO TEST
+
+          # dependency_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+          # dependency_optimization=rewriter_config_pb2.RewriterConfig.ON,           # TO TEST
+
+          # disable_model_pruning=False,           # INCOMPATIBLE with AMP
+          # function_optimization=True,
+          # implementation_selector=True,
+
+          # loop_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+          # loop_optimization=rewriter_config_pb2.RewriterConfig.ON,           # TO TEST
+
+          # The default setting (SCHEDULING and SWAPPING HEURISTICS only)
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.DEFAULT_MEM_OPT,
+
+          # Disabled in the meta-optimizer.
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.NO_MEM_OPT,
+
+          # Driven by manual op-level annotations.
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.MANUAL,
+
+          # Swapping heuristic will move a tensor from the GPU to the CPU and move it
+          # back when needed to reduce peak memory usage..
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.SWAPPING_HEURISTICS,
+
+          # Recomputation heuristics will recompute ops (such as Relu activation)
+          # during backprop instead of storing them, reducing peak memory usage.
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.RECOMPUTATION_HEURISTICS,
+
+          # Scheduling will split big ops such as AddN and try to enforce a schedule of
+          # the new computations that decreases peak memory usage.
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.SCHEDULING_HEURISTICS,
+
+          # Use any combination of swapping and recomputation heuristics.
+          # memory_optimization=rewriter_config_pb2.RewriterConfig.HEURISTICS,
+
+          meta_optimizer_iterations=rewriter_config_pb2.RewriterConfig.TWO,
+          # meta_optimizer_iterations=rewriter_config_pb2.RewriterConfig.ONE,
+          # meta_optimizer_iterations=rewriter_config_pb2.RewriterConfig.DEFAULT_NUM_ITERS,
+
+          # pin_to_host_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+          # pin_to_host_optimization=rewriter_config_pb2.RewriterConfig.ON,         # TO TEST
+          #
+          # remapping=rewriter_config_pb2.RewriterConfig.OFF,
+          # remapping=rewriter_config_pb2.RewriterConfig.ON,                   # TO TEST
+
+          # scoped_allocator_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+          # scoped_allocator_optimization=rewriter_config_pb2.RewriterConfig.ON,  # TO TEST
+
+          # shape_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+          # shape_optimization=rewriter_config_pb2.RewriterConfig.ON,           # TO TEST
+      )
+
+      if use_amp:
+          logging.info("[%s] AMP is activated - Experiment Feature" % mode)
+          rewrite_options.auto_mixed_precision = True
+
+      config = tf.compat.v1.ConfigProto(
+          allow_soft_placement=True,
+          log_device_placement=False,
+          graph_options=tf.compat.v1.GraphOptions(
+              rewrite_options=rewrite_options,
+              # infer_shapes=True  # Heavily drops throughput by 30%
+          )
+      )
+
+      if use_tf_distributed:
+        config.gpu_options.force_gpu_compatible = False
+
+      else:
+        config.gpu_options.force_gpu_compatible = True  # Force pinned memory
+
+        if MPI_is_distributed():
+            config.gpu_options.visible_device_list = str(MPI_local_rank())
+
+      if use_xla and (mode == "train" or allow_xla_at_inference):
+          logging.info("[%s] XLA is activated - Experiment Feature" % mode)
+          config.graph_options.optimizer_options.global_jit_level = tf.compat.v1.OptimizerOptions.ON_1
+          # config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_2
+
+      if mode == 'train':
+          config.intra_op_parallelism_threads = 1  # Avoid pool of Eigen threads
+
+          if MPI_is_distributed():
+              config.inter_op_parallelism_threads = max(2, multiprocessing.cpu_count() // hvd.local_size())
+
+          elif not use_tf_distributed:
+              config.inter_op_parallelism_threads = 4
+
+      return config
+
+  @abc.abstractmethod
+  def build_strategy_configuration(self, mode):
+    """Builds run configuration for distributed train/eval.
+
+    Returns:
+      RunConfig with distribution strategy configurations
+      to pass to the constructor of TPUEstimator/Estimator.
+    """
+
+    NotImplementedError('Must be implemented in subclass')
+
+  def build_model_parameters(self, mode):
+    """Builds model parameter."""
+
+    assert mode in ('train', 'eval')
+
+    batch_size = self._runtime_config.train_batch_size if mode == 'train' else self._runtime_config.eval_batch_size
+
+    params = dict(
+        self._runtime_config.values(),
+        mode=mode,
+        batch_size=batch_size,
+        model_dir=self._runtime_config.model_dir,
+    )
+
+    if mode == 'eval':
+      params = dict(
+        params,
+        augment_input_data=False,
+      )
+
+    return params
+
+  def build_mask_rcnn_estimator(self, params, run_config, mode):
+    """Creates TPUEstimator/Estimator instance.
+
+    Arguments:
+      params: A dictionary to pass to Estimator `model_fn`.
+      run_config: RunConfig instance specifying distribution strategy
+        configurations.
+      mode: Mode -- one of 'train` or `eval`.
+
+    Returns:
+      TFEstimator or TPUEstimator instance.
+    """
+    assert mode in ('train', 'eval')
+
+    return tf.estimator.Estimator(
+        model_fn=self._model_fn,
+        model_dir=self._runtime_config.model_dir,
+        config=run_config,
+        params=params
+    )
+
+  def _save_config(self):
+    """Save parameters to config files if model_dir is defined."""
+
+    model_dir = self._runtime_config.model_dir
+
+    if model_dir is not None:
+      if not tf.io.gfile.exists(model_dir):
+        tf.io.gfile.makedirs(model_dir)
+
+      params_io.save_hparams_to_yaml(self._runtime_config, model_dir + '/params.yaml')
+
+  def _write_summary(self, summary_dir, eval_results, predictions, current_step):
+
+    if not self._runtime_config.visualize_images_summary:
+      predictions = None
+
+    evaluation.write_summary(eval_results, summary_dir, current_step, predictions=predictions)
+
+  def train(self, train_input_fn, run_eval_after_train=False, eval_input_fn=None):
+    """Run distributed training on Mask RCNN model."""
+
+    self._save_config()
+    train_run_config = self.build_strategy_configuration('train')
+    train_params = self.build_model_parameters('train')
+    train_estimator = self.build_mask_rcnn_estimator(train_params, train_run_config, 'train')
+    print('train_estimator111111111111111111111111111111111111111111111111111111111111111111')
+    train_estimator.train(
+        input_fn=train_input_fn,
+        max_steps=self._runtime_config.total_steps,
+        hooks=get_training_hooks(
+            mode="train",
+            model_dir=self._runtime_config.model_dir,
+            checkpoint_path=self._runtime_config.checkpoint,
+            skip_checkpoint_variables=self._runtime_config.skip_checkpoint_variables
+        )
+    )
+
+    if not run_eval_after_train:
+      return None
+
+    if eval_input_fn is None:
+      raise ValueError('Eval input_fn must be passed to conduct evaluation after training.')
+
+    eval_run_config = self.build_strategy_configuration('eval')
+    eval_params = self.build_model_parameters('eval')
+    eval_estimator = self.build_mask_rcnn_estimator(eval_params, eval_run_config, 'eval')
+
+    last_ckpt = tf.train.latest_checkpoint(self._runtime_config.model_dir, latest_filename=None)
+    logging.info("Restoring parameters from %s\n" % last_ckpt)
+
+    eval_results, predictions = evaluation.evaluate(
+        eval_estimator,
+        eval_input_fn,
+        self._runtime_config.eval_samples,
+        self._runtime_config.eval_batch_size,
+        self._runtime_config.include_mask,
+        self._runtime_config.val_json_file,
+        report_frequency=self._runtime_config.report_frequency
+    )
+
+    output_dir = os.path.join(self._runtime_config.model_dir, 'eval')
+    tf.io.gfile.makedirs(output_dir)
+
+    # Summary writer writes out eval metrics.
+    self._write_summary(output_dir, eval_results, predictions, self._runtime_config.total_steps)
+
+    return eval_results
+
+  def train_and_eval(self, train_input_fn, eval_input_fn):
+    """Run distributed train and eval on Mask RCNN model."""
+
+    self._save_config()
+    output_dir = os.path.join(self._runtime_config.model_dir, 'eval')
+    tf.io.gfile.makedirs(output_dir)
+
+    train_run_config = self.build_strategy_configuration('train')
+    train_params = self.build_model_parameters('train')
+    train_estimator = self.build_mask_rcnn_estimator(train_params, train_run_config, 'train')
+
+    eval_estimator = None
+    eval_results = None
+
+    num_cycles = math.ceil(self._runtime_config.total_steps / self._runtime_config.num_steps_per_eval)
+
+    training_hooks = get_training_hooks(
+        mode="train",
+        model_dir=self._runtime_config.model_dir,
+        checkpoint_path=self._runtime_config.checkpoint,
+        skip_checkpoint_variables=self._runtime_config.skip_checkpoint_variables
+    )
+
+    for cycle in range(1, num_cycles + 1):
+
+      if not MPI_is_distributed() or MPI_rank() == 0:
+
+        print()  # Visual Spacing
+        logging.info("=================================")
+        logging.info('     Start training cycle %02d' % cycle)
+        logging.info("=================================\n")
+
+      max_cycle_step = min(int(cycle * self._runtime_config.num_steps_per_eval), self._runtime_config.total_steps)
+
+      PROFILER_ENABLED = False
+
+      if (not MPI_is_distributed() or MPI_rank() == 0) and PROFILER_ENABLED:
+          profiler_context_manager = tf.contrib.tfprof.ProfileContext
+
+      else:
+          from contextlib import suppress
+          profiler_context_manager = lambda *args, **kwargs: suppress()  # No-Op context manager
+
+      with profiler_context_manager(
+              '/workspace/profiling/',
+              trace_steps=range(100, 200, 3),
+              dump_steps=[200]
+      ) as pctx:
+
+          if (not MPI_is_distributed() or MPI_rank() == 0) and PROFILER_ENABLED:
+            opts = tf.compat.v1.profiler.ProfileOptionBuilder.time_and_memory()
+            pctx.add_auto_profiling('op', opts, [150, 200])
+
+          train_estimator.train(
+              input_fn=train_input_fn,
+              max_steps=max_cycle_step,
+              hooks=training_hooks,
+          )
+
+      if not MPI_is_distributed() or MPI_rank() == 0:
+
+          print()  # Visual Spacing
+          logging.info("=================================")
+          logging.info('    Start evaluation cycle %02d' % cycle)
+          logging.info("=================================\n")
+
+          if eval_estimator is None:
+              eval_run_config = self.build_strategy_configuration('eval')
+              eval_params = self.build_model_parameters('eval')
+              eval_estimator = self.build_mask_rcnn_estimator(eval_params, eval_run_config, 'eval')
+
+          last_ckpt = tf.train.latest_checkpoint(self._runtime_config.model_dir, latest_filename=None)
+          logging.info("Restoring parameters from %s\n" % last_ckpt)
+
+          eval_results, predictions = evaluation.evaluate(
+              eval_estimator,
+              eval_input_fn,
+              self._runtime_config.eval_samples,
+              self._runtime_config.eval_batch_size,
+              self._runtime_config.include_mask,
+              self._runtime_config.val_json_file,
+              report_frequency=self._runtime_config.report_frequency
+          )
+
+          self._write_summary(output_dir, eval_results, predictions, max_cycle_step)
+
+      if MPI_is_distributed():
+          from mpi4py import MPI
+          MPI.COMM_WORLD.Barrier()  # Waiting for all MPI processes to sync
+
+    return eval_results
+
+  def eval(self, eval_input_fn):
+    """Run distributed eval on Mask RCNN model."""
+
+    output_dir = os.path.join(self._runtime_config.model_dir, 'eval')
+    tf.io.gfile.makedirs(output_dir)
+
+    # Summary writer writes out eval metrics.
+    run_config = self.build_strategy_configuration('eval')
+    eval_params = self.build_model_parameters('eval')
+    eval_estimator = self.build_mask_rcnn_estimator(eval_params, run_config, 'eval')
+
+    logging.info('Starting to evaluate.')
+
+    last_ckpt = tf.train.latest_checkpoint(self._runtime_config.model_dir, latest_filename=None)
+
+    if last_ckpt is not None:
+      logging.info("Restoring parameters from %s\n" % last_ckpt)
+      current_step = int(os.path.basename(last_ckpt).split('-')[1])
+
+    else:
+      logging.warning(
+          "Could not find trained model in model_dir: `%s`, running initialization to predict\n" %
+          self._runtime_config.model_dir
+      )
+      current_step = 0
+
+    eval_results, predictions = evaluation.evaluate(
+        eval_estimator,
+        eval_input_fn,
+        self._runtime_config.eval_samples,
+        self._runtime_config.eval_batch_size,
+        self._runtime_config.include_mask,
+        self._runtime_config.val_json_file
+    )
+
+    self._write_summary(output_dir, eval_results, predictions, current_step)
+
+    if current_step >= self._runtime_config.total_steps:
+        logging.info('Evaluation finished after training step %d' % current_step)
+
+    return eval_results
+
+
+class EstimatorExecuter(BaseExecuter):
+  """Interface that runs Mask RCNN model using TPUEstimator."""
+
+  def __init__(self, runtime_config, model_fn):
+    super(EstimatorExecuter, self).__init__(runtime_config, model_fn)
+
+    if MPI_is_distributed():
+      os.environ['HOROVOD_GPU_ALLREDUCE'] = 'NCCL'
+      os.environ['HOROVOD_NUM_NCCL_STREAMS'] = '1'
+      # os.environ['HOROVOD_AUTOTUNE'] = '2'
+
+      hvd.init()
+
+      logging.info("Horovod successfully initialized ...")
+
+    os.environ['TF_GPU_THREAD_MODE'] = 'gpu_private'
+    os.environ['TF_GPU_THREAD_COUNT'] = '1' if not MPI_is_distributed() else str(hvd.size())
+
+    os.environ['TF_SYNC_ON_FINISH'] = '0'
+
+  def build_strategy_configu    ration(self, mode):
+    """Retrieves model configuration for running TF Estimator."""
+
+    run_config = tf.estimator.RunConfig(
+        tf_random_seed=(
+            self._runtime_config.seed
+            if not MPI_is_distributed() or self._runtime_config.seed is None else
+            self._runtime_config.seed + MPI_rank()
+        ),
+        model_dir=self._runtime_config.model_dir,
+        save_summary_steps=None,  # disabled
+        save_checkpoints_steps=None,  # disabled
+        save_checkpoints_secs=None,  # disabled
+        keep_checkpoint_max=20,  # disabled
+        keep_checkpoint_every_n_hours=None,  # disabled
+        log_step_count_steps=None,  # disabled
+        session_config=self._get_session_config(
+            mode=mode,
+            use_xla=self._runtime_config.xla,
+            use_amp=self._runtime_config.amp,
+            use_tf_distributed=False,
+            allow_xla_at_inference=self._runtime_config.allow_xla_at_inference  # TODO: Remove when XLA at inference fixed
+        ),
+        protocol=None,
+        device_fn=None,
+        train_distribute=None,
+        eval_distribute=None,
+        experimental_distribute=None
+    )
+
+    return run_config
+
+
+class TFDistributedExecuter(BaseExecuter):
+  """Interface that runs Mask RCNN model using MultiWorkerMirroredStrategy."""
+
+  @staticmethod
+  def is_eval_task():
+    return tf.distribute.cluster_resolver.TFConfigClusterResolver().task_type == 'evaluator'
+
+  def build_strategy_configuration(self, mode):
+    """Retrieves model configuration for MultiWorkerMirroredStrategy."""
+
+    distributed_strategy = tf.distribute.MirroredStrategy()
+    # distributed_strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy(
+    #     tf.distribute.experimental.CollectiveCommunication.NCCL
+    # )
+
+    run_config = tf.estimator.RunConfig(
+        tf_random_seed=self._runtime_config.seed,
+        model_dir=self._runtime_config.model_dir,
+        save_summary_steps=None,  # disabled
+        save_checkpoints_steps=None,  # disabled
+        save_checkpoints_secs=None,  # disabled
+        keep_checkpoint_max=20,  # disabled
+        keep_checkpoint_every_n_hours=None,  # disabled
+        log_step_count_steps=None,  # disabled
+        session_config=self._(
+            mode=mode,
+            use_xla=self._runtime_config.xla,
+            use_amp=self._runtime_config.amp,
+            use_tf_distributed=True,
+            # TODO: Remove when XLA at inference fixed
+            allow_xla_at_inference=self._runtime_config.allow_xla_at_inference
+        ),
+        protocol=None,
+        device_fn=None,
+        train_distribute=distributed_strategy if mode == "train" else None,
+        eval_distribute=None,
+        experimental_distribute=None
+    )
+
+    return run_config

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/evaluation.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Functions to perform COCO evaluation."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import copy
+import operator
+import pprint
+import six
+import time
+
+import io
+from PIL import Image
+
+import numpy as np
+import tensorflow as tf
+
+from mask_rcnn.utils.logging_formatter import logging
+
+from mask_rcnn import coco_metric
+from mask_rcnn.utils import coco_utils
+
+from mask_rcnn.object_detection import visualization_utils
+
+import dllogger
+from dllogger import Verbosity
+
+
+def process_prediction_for_eval(prediction):
+    """Process the model prediction for COCO eval."""
+    image_info = prediction['image_info']
+    box_coordinates = prediction['detection_boxes']
+    processed_box_coordinates = np.zeros_like(box_coordinates)
+
+    for image_id in range(box_coordinates.shape[0]):
+        scale = image_info[image_id][2]
+
+        for box_id in range(box_coordinates.shape[1]):
+            # Map [y1, x1, y2, x2] -> [x1, y1, w, h] and multiply detections
+            # Map [y1, x1, y2, x2] -> [x1, y1, w, h] and multiply detections
+            # by image scale.
+            y1, x1, y2, x2 = box_coordinates[image_id, box_id, :]
+            new_box = scale * np.array([x1, y1, x2 - x1, y2 - y1])
+            processed_box_coordinates[image_id, box_id, :] = new_box
+
+    prediction['detection_boxes'] = processed_box_coordinates
+    return prediction
+
+
+def compute_coco_eval_metric(predictor,
+                             num_batches=-1,
+                             include_mask=True,
+                             annotation_json_file="",
+                             eval_batch_size=-1,
+                             report_frequency=None):
+    """Compute COCO eval metric given a prediction generator.
+
+    Args:
+    predictor: a generator that iteratively pops a dictionary of predictions
+      with the format compatible with COCO eval tool.
+    num_batches: the number of batches to be aggregated in eval. This is how
+      many times that the predictor gets pulled.
+    include_mask: a boolean that indicates whether we include the mask eval.
+    annotation_json_file: the annotation json file of the eval dataset.
+
+    Returns:
+    eval_results: the aggregated COCO metric eval results.
+    """
+
+    if annotation_json_file == "":
+        annotation_json_file = None
+
+    use_groundtruth_from_json = (annotation_json_file is not None)
+
+    predictions = dict()
+    batch_idx = 0
+
+    if use_groundtruth_from_json:
+        eval_metric = coco_metric.EvaluationMetric(annotation_json_file, include_mask=include_mask)
+
+    else:
+        eval_metric = coco_metric.EvaluationMetric(filename=None, include_mask=include_mask)
+
+    def evaluation_preds(preds):
+
+        # Essential to avoid modifying the source dict
+        _preds = copy.deepcopy(preds)
+
+        for k, v in six.iteritems(_preds):
+            _preds[k] = np.concatenate(_preds[k], axis=0)
+
+        if 'orig_images' in _preds and _preds['orig_images'].shape[0] > 10:
+            # Only samples a few images for visualization.
+            _preds['orig_images'] = _preds['orig_images'][:10]
+
+        if use_groundtruth_from_json:
+            eval_results = eval_metric.predict_metric_fn(_preds)
+
+        else:
+            images, annotations = coco_utils.extract_coco_groundtruth(_preds, include_mask)
+            coco_dataset = coco_utils.create_coco_format_dataset(images, annotations)
+            eval_results = eval_metric.predict_metric_fn(_preds, groundtruth_data=coco_dataset)
+
+        return eval_results
+
+    # Take into account cuDNN & Tensorflow warmup
+    # Drop N first steps for avg throughput calculation
+    BURNIN_STEPS = 100
+    model_throughput_list = list()
+    inference_time_list = list()
+
+    while num_batches < 0 or batch_idx < num_batches:
+
+        try:
+            step_t0 = time.time()
+            step_predictions = six.next(predictor)
+            batch_time = time.time() - step_t0
+
+            throughput = eval_batch_size / batch_time
+            model_throughput_list.append(throughput)
+            inference_time_list.append(batch_time)
+
+            logging.info('Running inference on batch %03d/%03d... - Step Time: %.4fs - Throughput: %.1f imgs/s' % (
+                batch_idx + 1,
+                num_batches,
+                batch_time,
+                throughput
+            ))
+
+        except StopIteration:
+            logging.info('Get StopIteration at %d batch.' % (batch_idx + 1))
+            break
+
+        step_predictions = process_prediction_for_eval(step_predictions)
+
+        for k, v in step_predictions.items():
+
+            if k not in predictions:
+                predictions[k] = [v]
+
+            else:
+                predictions[k].append(v)
+
+        batch_idx = batch_idx + 1
+
+        # If you want the report to happen each report_frequency to happen each report_frequency batches.
+        # Thus, each report is of eval_batch_size * report_frequency
+        if report_frequency and batch_idx % report_frequency == 0:
+            eval_results = evaluation_preds(preds=predictions)
+            logging.info('Eval results: %s' % pprint.pformat(eval_results, indent=4))
+
+    inference_time_list.sort()
+    eval_results = evaluation_preds(preds=predictions)
+
+    average_time = np.mean(inference_time_list)
+    latency_50 = max(inference_time_list[:int(len(inference_time_list) * 0.5)])
+    latency_90 = max(inference_time_list[:int(len(inference_time_list) * 0.90)])
+    latency_95 = max(inference_time_list[:int(len(inference_time_list) * 0.95)])
+    latency_99 = max(inference_time_list[:int(len(inference_time_list) * 0.99)])
+    latency_100 = max(inference_time_list[:int(len(inference_time_list) * 1)])
+
+    print()  # Visual Spacing
+    logging.info("# @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ #")
+    logging.info("         Evaluation Performance Summary          ")
+    logging.info("# @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ #")
+
+    total_processing_hours, rem = divmod(np.sum(model_throughput_list), 3600)
+    total_processing_minutes, total_processing_seconds = divmod(rem, 60)
+
+    if len(model_throughput_list) > BURNIN_STEPS:
+        # Take into account cuDNN & Tensorflow warmup
+        # Drop N first steps for avg throughput calculation
+        # Also drop last step which may have a different batch size
+        avg_throughput = np.mean(model_throughput_list[BURNIN_STEPS:-1])
+    else:
+        avg_throughput = -1.
+
+    print()  # Visual Spacing
+    logging.info("Average throughput: {throughput:.1f} samples/sec".format(throughput=avg_throughput))
+    logging.info("Inference Latency Average (s) = {avg:.4f}".format(avg=average_time))
+    logging.info("Inference Latency 50% (s) = {cf_50:.4f}".format(cf_50=latency_50))
+    logging.info("Inference Latency 90%  (s) = {cf_90:.4f}".format(cf_90=latency_90))
+    logging.info("Inference Latency 95%  (s) = {cf_95:.4f}".format(cf_95=latency_95))
+    logging.info("Inference Latency 99%  (s) = {cf_99:.4f}".format(cf_99=latency_99))
+    logging.info("Inference Latency 100%  (s) = {cf_100:.4f}".format(cf_100=latency_100))
+    logging.info("Total processed steps: {total_steps}".format(total_steps=len(model_throughput_list)))
+    logging.info(
+        "Total processing time: {hours}h {minutes:02d}m {seconds:02d}s".format(
+            hours=total_processing_hours,
+            minutes=int(total_processing_minutes),
+            seconds=int(total_processing_seconds)
+        )
+    )
+    dllogger.log(step=(), data={"avg_inference_throughput": avg_throughput}, verbosity=Verbosity.DEFAULT)
+    avg_inference_time = float(total_processing_hours * 3600 + int(total_processing_minutes) * 60 +
+        int(total_processing_seconds))
+    dllogger.log(step=(), data={"avg_inference_time": avg_inference_time}, verbosity=Verbosity.DEFAULT)
+    logging.info("==================== Metrics ====================")
+
+    # logging.info('Eval Epoch results: %s' % pprint.pformat(eval_results, indent=4))
+    for key, value in sorted(eval_results.items(), key=operator.itemgetter(0)):
+        logging.info("%s: %.9f" % (key, value))
+    print()  # Visual Spacing
+
+    return eval_results, predictions
+
+
+def evaluate(eval_estimator,
+             input_fn,
+             num_eval_samples,
+             eval_batch_size,
+             include_mask=True,
+             validation_json_file="",
+             report_frequency=None):
+
+    """Runs COCO evaluation once."""
+    predictor = eval_estimator.predict(
+        input_fn=input_fn,
+        yield_single_examples=False
+    )
+
+    # Every predictor.next() gets a batch of prediction (a dictionary).
+    num_eval_times = num_eval_samples // eval_batch_size
+    assert num_eval_times > 0, 'num_eval_samples must be >= eval_batch_size!'
+
+    eval_results, predictions = compute_coco_eval_metric(
+        predictor,
+        num_eval_times,
+        include_mask,
+        validation_json_file,
+        eval_batch_size=eval_batch_size,
+        report_frequency=report_frequency
+    )
+
+    return eval_results, predictions
+
+
+def write_summary(eval_results, summary_dir, current_step, predictions=None):
+    """Write out eval results for the checkpoint."""
+    with tf.Graph().as_default():
+        summaries = []
+
+        # Summary writer writes out eval metrics.
+        try:
+            # Tensorflow 1.x
+            summary_writer = tf.compat.v1.summary.FileWriter(summary_dir)
+        except AttributeError:
+            # Tensorflow 2.x
+            summary_writer = tf.summary.create_file_writer(summary_dir)
+            summary_writer.as_default()
+
+        eval_results_dict = {}
+        for metric in eval_results:
+            try:
+                summaries.append(tf.compat.v1.Summary.Value(tag=metric, simple_value=eval_results[metric]))
+                eval_results_dict[metric] = float(eval_results[metric])
+
+            except AttributeError:
+                tf.summary.scalar(name=metric, data=eval_results[metric], step=current_step)
+                eval_results_dict[metric] = float(eval_results[metric])
+        dllogger.log(step=(), data=eval_results_dict, verbosity=Verbosity.DEFAULT)
+
+        if isinstance(predictions, dict) and predictions:
+            images_summary = get_image_summary(predictions, current_step)
+
+            try:
+                summaries += images_summary
+            except TypeError:
+                summaries.append(images_summary)
+
+        try:
+            # tf_summaries = tf.compat.v1.Summary(value=list(summaries))
+            tf_summaries = tf.compat.v1.Summary(value=summaries)
+            summary_writer.add_summary(tf_summaries, current_step)
+            summary_writer.flush()
+
+        except AttributeError:
+            tf.summary.flush(summary_writer)
+
+
+def generate_image_preview(image, boxes, scores, classes, gt_boxes=None, segmentations=None):
+    """Creates an image summary given predictions."""
+    max_boxes_to_draw = 100
+    min_score_thresh = 0.1
+
+    # Visualizes the predicitons.
+    image_with_detections = visualization_utils.visualize_boxes_and_labels_on_image_array(
+        image,
+        boxes,
+        classes=classes,
+        scores=scores,
+        category_index={},
+        instance_masks=segmentations,
+        use_normalized_coordinates=False,
+        max_boxes_to_draw=max_boxes_to_draw,
+        min_score_thresh=min_score_thresh,
+        agnostic_mode=False
+    )
+
+    if gt_boxes is not None:
+        # Visualizes the groundtruth boxes. They are in black by default.
+        image_with_detections = visualization_utils.visualize_boxes_and_labels_on_image_array(
+            image_with_detections,
+            gt_boxes,
+            classes=None,
+            scores=None,
+            category_index={},
+            use_normalized_coordinates=False,
+            max_boxes_to_draw=max_boxes_to_draw,
+            agnostic_mode=True
+        )
+
+    return image_with_detections
+
+
+def generate_image_buffer(input_image):
+    buf = io.BytesIO()
+    w, h = input_image.shape[:2]
+    ratio = 1024 / w
+    new_size = [int(w * ratio), int(h * ratio)]
+
+    image = Image.fromarray(input_image.astype(np.uint8))
+    image.thumbnail(new_size)
+    image.save(buf, format='png')
+
+    return buf.getvalue()
+
+
+def get_image_summary(predictions, current_step, max_images=10):
+    """Write out image and prediction for summary."""
+
+    if 'orig_images' not in predictions:
+        logging.info('Missing orig_images in predictions: %s', predictions.keys())
+        return
+
+    max_images = min(
+        len(predictions['orig_images']) * predictions['orig_images'][0].shape[0],
+        max_images
+    )
+
+    _detection_boxes = np.concatenate(predictions['detection_boxes'], axis=0)
+    _detection_scores = np.concatenate(predictions['detection_scores'], axis=0)
+    _detection_classes = np.concatenate(predictions['detection_classes'], axis=0)
+    _image_info = np.concatenate(predictions['image_info'], axis=0)
+    _num_detections = np.concatenate(predictions['num_detections'], axis=0)
+    _orig_images = np.concatenate(predictions['orig_images'], axis=0)
+
+    if 'detection_masks' in predictions:
+        _detection_masks = np.concatenate(predictions['detection_masks'], axis=0)
+    else:
+        _detection_masks = None
+
+    if 'groundtruth_boxes' in predictions:
+        _groundtruth_boxes = np.concatenate(predictions['groundtruth_boxes'], axis=0)
+    else:
+        _groundtruth_boxes = None
+
+    _orig_images = _orig_images * 255
+    _orig_images = _orig_images.astype(np.uint8)
+
+    image_previews = []
+
+    for i in range(max_images):
+        num_detections = min(len(_detection_boxes[i]), int(_num_detections[i]))
+
+        detection_boxes = _detection_boxes[i][:num_detections]
+        detection_scores = _detection_scores[i][:num_detections]
+        detection_classes = _detection_classes[i][:num_detections]
+
+        image = _orig_images[i]
+        image_height = image.shape[0]
+        image_width = image.shape[1]
+
+        # Rescale the box to fit the visualization image.
+        h, w = _image_info[i][3:5]
+        detection_boxes = detection_boxes / np.array([w, h, w, h])
+        detection_boxes = detection_boxes * np.array([image_width, image_height, image_width, image_height])
+
+        if _groundtruth_boxes is not None:
+            gt_boxes = _groundtruth_boxes[i]
+            gt_boxes = gt_boxes * np.array([image_height, image_width, image_height, image_width])
+        else:
+            gt_boxes = None
+
+        if _detection_masks is not None:
+            instance_masks = _detection_masks[i][0:num_detections]
+            segmentations = coco_metric.generate_segmentation_from_masks(
+                instance_masks,
+                detection_boxes,
+                image_height,
+                image_width
+            )
+        else:
+            segmentations = None
+
+        # From [x, y, w, h] to [x1, y1, x2, y2] and
+        # process_prediction_for_eval() set the box to be [x, y] format, need to
+        # reverted them to [y, x] format.
+        xmin, ymin, w, h = np.split(detection_boxes, 4, axis=-1)
+        xmax = xmin + w
+        ymax = ymin + h
+
+        boxes_to_visualize = np.concatenate([ymin, xmin, ymax, xmax], axis=-1)
+
+        image_preview = generate_image_preview(
+            image,
+            boxes=boxes_to_visualize,
+            scores=detection_scores,
+            classes=detection_classes.astype(np.int32),
+            gt_boxes=gt_boxes,
+            segmentations=segmentations
+        )
+        image_previews.append(image_preview)
+
+    try:
+        summaries = []
+
+        for i, image_preview in enumerate(image_previews):
+            image_buffer = generate_image_buffer(image_preview)
+            image_summary = tf.compat.v1.Summary.Image(encoded_image_string=image_buffer)
+            image_value = tf.compat.v1.Summary.Value(tag='%d_input' % i, image=image_summary)
+
+            summaries.append(image_value)
+
+    except AttributeError:
+        image_previews = np.array(image_previews)
+        summaries = tf.summary.image(
+            name='image_summary',
+            data=image_previews,
+            step=current_step,
+            max_outputs=max_images
+        )
+
+    return summaries

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hooks/__init__.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from mask_rcnn.hooks.ckpt_hook import CheckpointSaverHook
+from mask_rcnn.hooks.pretrained_restore_hook import PretrainedWeightsLoadingHook
+
+__all__ = [
+    "CheckpointSaverHook",
+    "PretrainedWeightsLoadingHook",
+]

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hooks/ckpt_hook.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+import tensorflow as tf
+
+from mask_rcnn.utils.logging_formatter import logging
+
+__all__ = ["CheckpointSaverHook"]
+
+
+class CheckpointSaverHook(tf.estimator.SessionRunHook):
+    """Saves checkpoints every N steps or seconds."""
+
+    def __init__(self, checkpoint_dir, checkpoint_basename="model.ckpt"):
+        """Initializes a `CheckpointSaverHook`.
+        Args:
+          checkpoint_dir: `str`, base directory for the checkpoint files.
+          checkpoint_basename: `str`, base name for the checkpoint files.
+        Raises:
+          ValueError: One of `save_steps` or `save_secs` should be set.
+          ValueError: At most one of `saver` or `scaffold` should be set.
+        """
+        logging.info("Create CheckpointSaverHook.")
+
+        self._saver = None
+        self._checkpoint_dir = checkpoint_dir
+        self._save_path = os.path.join(checkpoint_dir, checkpoint_basename)
+
+        self._steps_per_run = 1
+
+        self._is_initialized = False
+
+        self._global_step_tensor = None
+        self._summary_writer = None
+
+    def _set_steps_per_run(self, steps_per_run):
+        self._steps_per_run = steps_per_run
+
+    def begin(self):
+        self._global_step_tensor = tf.compat.v1.train.get_or_create_global_step()
+
+        self._saver = tf.compat.v1.train.Saver()
+
+        from tensorflow.python.training import summary_io
+        self._summary_writer = summary_io.SummaryWriterCache.get(self._checkpoint_dir)
+
+        if self._global_step_tensor is None:
+            raise RuntimeError(
+                "Global step should be created to use CheckpointSaverHook."
+            )
+
+    def after_create_session(self, session, coord):
+
+        if not self._is_initialized:
+            global_step = session.run(self._global_step_tensor)
+            from tensorflow.python.keras.backend import get_graph
+            default_graph = get_graph()
+
+            # We do write graph and saver_def at the first call of before_run.
+            # We cannot do this in begin, since we let other hooks to change graph and
+            # add variables in begin. Graph is finalized after all begin calls.
+            tf.io.write_graph(
+                default_graph.as_graph_def(add_shapes=True),
+                self._checkpoint_dir,
+                "graph.pbtxt"
+            )
+
+            saver_def = self._saver.saver_def
+
+            from tensorflow.python.framework import meta_graph
+
+            meta_graph_def = meta_graph.create_meta_graph_def(
+                graph_def=default_graph.as_graph_def(add_shapes=True),
+                saver_def=saver_def
+            )
+
+            self._summary_writer.add_graph(default_graph)
+            self._summary_writer.add_meta_graph(meta_graph_def)
+
+            # The checkpoint saved here is the state at step "global_step".
+            self._save(session, global_step)
+
+            self._is_initialized = True
+
+    def end(self, session):
+        last_step = session.run(self._global_step_tensor)
+
+        self._save(session, last_step)
+
+    def _save(self, session, step):
+        """Saves the latest checkpoint, returns should_stop."""
+        logging.info("Saving checkpoints for %d into %s.", step, self._save_path)
+
+        self._saver.save(session, self._save_path, global_step=step)
+        self._summary_writer.add_session_log(
+            tf.compat.v1.SessionLog(status=tf.compat.v1.SessionLog.CHECKPOINT, checkpoint_path=self._save_path),
+            step
+        )

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hooks/logging_hook.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import operator
+import time
+
+import numpy as np
+import tensorflow as tf
+
+from distutils.version import LooseVersion
+
+from mask_rcnn.utils.logging_formatter import logging
+
+from mask_rcnn.utils import meters
+
+from mask_rcnn.utils.decorators import atexit_hook
+
+from mask_rcnn.utils.distributed_utils import MPI_is_distributed
+from mask_rcnn.utils.distributed_utils import MPI_rank_and_size
+from mask_rcnn.utils.distributed_utils import MPI_size
+
+from mask_rcnn.utils.logging_backend import LoggingBackend
+from mask_rcnn.utils.logging_backend import RuntimeMode
+
+from mask_rcnn.utils.metric_tracking import clear_registered_metrics
+
+from mask_rcnn.utils.metric_tracking import TF_METRICS
+from mask_rcnn.utils.metric_tracking import KERAS_MODELS
+
+from mask_rcnn.utils.lazy_imports import LazyImport
+hvd = LazyImport("horovod.tensorflow")
+
+__all__ = ["AutoLoggingHook"]
+
+
+@atexit_hook
+class _AutoLoggingHook(tf.estimator.SessionRunHook):
+
+    def __init__(self, log_every_n_steps=200, warmup_steps=500, is_training=True):
+        """
+        AutoLogging Hook for Tensorflow
+
+        :param log_every_n_steps: log will be output on the console every N steps
+        :param warmup_steps: integers, numbers of steps considered as warmup
+        :param is_training: boolean
+        """
+
+        self._logging_proxy = LoggingBackend()
+
+        self._initialized = False
+        self._metrics = copy.copy(TF_METRICS)
+
+        self._batch_size_tensor = None
+
+        self._AMP_steps_since_last_loss_scale = None
+        self._AMP_loss_scale_tensor = None
+
+        self._current_step = None
+        self._amp_steps_non_skipped = None
+
+        self._warmup_steps = warmup_steps
+
+        self._log_every_n_steps = log_every_n_steps
+
+        self._step_t0 = None
+        self._session_t0 = None
+        self._session_run_times = list()
+
+        self._global_step_tensor = None
+
+        self._is_training = is_training
+        self._runtime_mode = RuntimeMode.TRAIN if is_training else RuntimeMode.VALIDATION
+
+        self._model_throughput = meters.MovingAverageMeter(window_size=1000)
+        self._model_stats = None
+
+        self._n_gpus = None
+
+    def __atexit__(self):
+
+        if self._initialized:
+
+            total_processing_time = int(np.sum(self._session_run_times))
+
+            try:
+                avg_throughput = self._model_throughput.read()
+            except ValueError:
+                avg_throughput = -1
+
+            self._logging_proxy.log_summary(
+                is_train=self._is_training,
+                total_steps=self._current_step,
+                total_processing_time=total_processing_time,
+                avg_throughput=avg_throughput
+            )
+
+            metric_data = dict()
+
+            for key, value in self._metrics.items():
+                try:
+                    metric_data[key] = value["aggregator"].read()
+
+                except ValueError:
+                    pass
+
+            self._logging_proxy.log_final_metrics(metric_data=metric_data, runtime_mode=self._runtime_mode)
+
+    def begin(self):
+        """Called once before using the session.
+        When called, the default graph is the one that will be launched in the
+        session.  The hook can modify the graph by adding new operations to it.
+        After the `begin()` call the graph will be finalized and the other callbacks
+        can not modify the graph anymore. Second call of `begin()` on the same
+        graph, should not change the graph.
+        """
+
+        from tensorflow.python.keras.backend import get_graph
+        _graph = get_graph()
+
+        try:
+            self._batch_size_tensor = None
+
+            for tensor in _graph.as_graph_def().node:
+                if "IteratorGetNext" in tensor.name:
+                    _input_tensor = _graph.get_tensor_by_name(tensor.name + ":0")
+                    try:
+                        self._batch_size_tensor = tf.shape(input=_input_tensor)[0]
+                    except TypeError:  # Ragged Tensor
+                        self._batch_size_tensor = _input_tensor.bounding_shape()[0]
+                    break
+            else:
+                raise RuntimeError(
+                    "Tensor `{}` could not be found. "
+                    "Make sure you are using tf.data API".format("IteratorGetNext")
+                )
+
+        except RuntimeError:
+            raise
+
+        except Exception as e:
+            raise RuntimeError(
+                "Impossible to fetch the tensor: `IteratorGetNext`. Make sure you are using tf.data API."
+            ) from e
+
+        self._global_step_tensor = tf.compat.v1.train.get_or_create_global_step()
+
+        try:
+            self._AMP_loss_scale_tensor = _graph.get_tensor_by_name("current_loss_scale/Read/ReadVariableOp:0")
+            self._AMP_steps_since_last_loss_scale = _graph.get_tensor_by_name("current_loss_scale/Read/ReadVariableOp:0")
+
+        except RuntimeError:
+            raise
+
+        # TF-AMP is not activated
+        except Exception:
+            pass
+
+        # if self._is_training:
+        #     self.runtime_data["params_count"] = tf.reduce_sum(
+        #         [tf.reduce_prod(v.shape) for v in tf.trainable_variables()]
+        #     )
+
+    def end(self, session):  # pylint: disable=unused-argument
+        """Called at the end of session.
+        The `session` argument can be used in case the hook wants to run final ops,
+        such as saving a last checkpoint.
+        If `session.run()` raises exception other than OutOfRangeError or
+        StopIteration then `end()` is not called.
+        Note the difference between `end()` and `after_run()` behavior when
+        `session.run()` raises OutOfRangeError or StopIteration. In that case
+        `end()` is called but `after_run()` is not called.
+        Args:
+          session: A TensorFlow Session that will be soon closed.
+        """
+
+        self._session_run_times.append(time.time() - self._session_t0)
+
+    def after_create_session(self, session, coord):  # pylint: disable=unused-argument3
+        """Called when new TensorFlow session is created.
+        This is called to signal the hooks that a new session has been created. This
+        has two essential differences with the situation in which `begin` is called:
+        * When this is called, the graph is finalized and ops can no longer be added
+            to the graph.
+        * This method will also be called as a result of recovering a wrapped
+            session, not only at the beginning of the overall session.
+        Args:
+          session: A TensorFlow Session that has been created.
+          coord: A Coordinator object which keeps track of all threads.
+        """
+
+        # ========= Collect the number of GPUs ======== #
+        if self._is_training:
+
+            if MPI_is_distributed():
+                self._n_gpus = MPI_size()
+
+            elif tf.distribute.has_strategy():
+                self._n_gpus = tf.distribute.get_strategy().num_replicas_in_sync
+
+            else:
+                self._n_gpus = 1
+
+        else:
+            self._n_gpus = 1
+
+        # =========== TensorFlow Hook Setup =========== #
+        _global_step, _metrics = setup_tensorflow_hook(
+            sess=session,
+            logging_proxy=self._logging_proxy,
+            is_training=self._is_training,
+            is_initialized=self._initialized
+        )
+
+        if _global_step >= 0:
+            self._current_step = self._amp_steps_non_skipped = _global_step
+
+        self._metrics.update(_metrics)
+
+        if not self._is_training:
+
+            for metric_name in self._metrics.keys():
+                self._metrics[metric_name]["aggregator"].reset()
+
+        self._initialized = True
+        # %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% #
+
+        self._session_t0 = time.time()
+
+    def before_run(self, run_context):  # pylint: disable=unused-argument
+        """Called before each call to run().
+        You can return from this call a `SessionRunArgs` object indicating ops or
+        tensors to add to the upcoming `run()` call.  These ops/tensors will be run
+        together with the ops/tensors originally passed to the original run() call.
+        The run args you return can also contain feeds to be added to the run()
+        call.
+        The `run_context` argument is a `SessionRunContext` that provides
+        information about the upcoming `run()` call: the originally requested
+        op/tensors, the TensorFlow Session.
+        At this point graph is finalized and you can not add ops.
+        Args:
+          run_context: A `SessionRunContext` object.
+        Returns:
+          None or a `SessionRunArgs` object.
+        """
+
+        self._current_step += 1
+
+        request_fetches = {
+            "global_step": self._global_step_tensor, "metrics": dict(), "batch_size": self._batch_size_tensor
+        }
+
+        if self._is_training and self._AMP_steps_since_last_loss_scale is not None:
+            request_fetches["AMP"] = {
+                "steps_since_last_loss_scale": self._AMP_steps_since_last_loss_scale,
+                "current_loss_scale": self._AMP_loss_scale_tensor,
+            }
+
+        if self._current_step % self._log_every_n_steps == 0:
+            for key, value in self._metrics.items():
+                request_fetches["metrics"][key] = value["tensor"]
+
+        self._step_t0 = time.time()
+
+        return tf.estimator.SessionRunArgs(request_fetches)
+
+    def after_run(self, run_context, run_values):  # pylint: disable=unused-argument
+        """Called after each call to run().
+        The `run_values` argument contains results of requested ops/tensors by
+        `before_run()`.
+        The `run_context` argument is the same one send to `before_run` call.
+        `run_context.request_stop()` can be called to stop the iteration.
+        If `session.run()` raises any exceptions then `after_run()` is not called.
+        Args:
+          run_context: A `SessionRunContext` object.
+          run_values: A SessionRunValues object.
+        """
+
+        batch_time = time.time() - self._step_t0
+
+        _global_step = run_values.results["global_step"]
+
+        if self._is_training and self._AMP_steps_since_last_loss_scale is not None:
+
+            try:
+                AMP_steps_since_last_loss_scale = run_values.results["AMP"]["steps_since_last_loss_scale"]
+                AMP_loss_scale = run_values.results["AMP"]["current_loss_scale"]
+
+            except KeyError:
+                AMP_steps_since_last_loss_scale = None
+                AMP_loss_scale = None
+
+            if AMP_steps_since_last_loss_scale is not None:
+
+                # Step has been skipped
+                if _global_step != (self._amp_steps_non_skipped + 1):
+                    logging.warning(
+                        "AMP - Training iteration `#{step}` has been skipped and loss rescaled. "
+                        "New Loss Scale: {loss_scale}\n".format(step=self._current_step, loss_scale=AMP_loss_scale)
+                    )
+
+                else:
+                    self._amp_steps_non_skipped += 1
+
+                    if AMP_steps_since_last_loss_scale == 0:
+                        logging.warning(
+                            "AMP - Training iteration `#{step}` - Loss scale has been automatically increased. "
+                            "New Loss Scale: {loss_scale}\n".format(step=self._current_step, loss_scale=AMP_loss_scale)
+                        )
+
+        else:
+            AMP_steps_since_last_loss_scale = None
+            AMP_loss_scale = None
+
+        def get_model_throughput():
+            gpu_batch_size = run_values.results["batch_size"]
+            return gpu_batch_size / batch_time * self._n_gpus
+
+        # def get_model_stats():
+        #     return get_tf_model_statistics(batch_size=run_values.results["batch_size"], scope_name=None)
+        #
+        # if self._model_stats is None:
+        #     self._model_stats = get_model_stats()
+
+        is_log_step = self._current_step % self._log_every_n_steps == 0
+
+        if is_log_step:
+
+            if self._current_step > self._warmup_steps:
+
+                try:
+                    model_throughput = self._model_throughput.read()
+                except ValueError:
+                    model_throughput = get_model_throughput()
+
+            else:
+                model_throughput = get_model_throughput()
+
+            self._logging_proxy.log_step(iteration=self._current_step, throughput=model_throughput, gpu_stats=[])
+
+            self._logging_proxy.log_amp_runtime(
+                current_loss_scale=AMP_loss_scale,
+                steps_non_skipped=_global_step,
+                steps_since_last_scale=AMP_steps_since_last_loss_scale,
+            )
+
+            metric_data = dict()
+
+            for name, value in sorted(run_values.results["metrics"].items(), key=operator.itemgetter(0)):
+                self._metrics[name]["aggregator"].record(value)
+
+                metric_data[name] = self._metrics[name]["aggregator"].read()
+
+            self._logging_proxy.log_metrics(
+                metric_data=metric_data, iteration=self._current_step, runtime_mode=self._runtime_mode
+            )
+
+            print()  # Visual Spacing
+
+        elif self._current_step > self._warmup_steps:
+            # Do not store speed for log step due to additional fetches
+            self._model_throughput.record(get_model_throughput())
+
+
+class _SlaveGPUsHook(tf.estimator.SessionRunHook):
+
+    def after_create_session(self, session, coord):
+
+        with logging.temp_verbosity(logging.INFO):  # Do not warn user about metric cleaning
+            clear_registered_metrics()
+
+
+def real_autologging_hook(*args, **kwargs):
+
+    replica_id = tf.distribute.get_replica_context().replica_id_in_sync_group
+
+    # Do not set a logging hook for GPUs != 0
+    if MPI_rank_and_size()[0] != 0 or (isinstance(replica_id, tf.Tensor) and tf.get_static_value(replica_id) != 0):
+        return _SlaveGPUsHook()
+
+    else:
+        _ = LoggingBackend()  # Making sure the backend is defined before any hook due to __atexit__ hook
+        return _AutoLoggingHook(*args, **kwargs)
+
+
+def collect_registered_metrics():
+
+    if TF_METRICS:  # if not empty
+
+        metrics = copy.copy(TF_METRICS)
+
+        # Do not warn user about metric cleaning
+        with logging.temp_verbosity(logging.INFO):
+            clear_registered_metrics()
+
+        return metrics
+
+    else:
+        return dict()
+
+
+def get_model_variables():
+    """return model variables: global variables without optimizer's variables"""
+
+    return [
+        # yapf: disable
+        var for var in tf.compat.v1.global_variables() if (
+            var.name[-11:] not in "/Momentum:0" and
+            var.name[-11:] not in "/Adadelta:0" and
+            var.name[-13:] not in "/Adadelta_1:0" and
+            var.name[-7:] not in "/Adam:0" and
+            var.name[-9:] not in "/Adam_1:0" and
+            var.name[-10:] not in "/Adagrad:0" and
+            var.name[10:] not in "/RMSProp:0" and
+            var.name[-12:] not in "/RMSProp_1:0" and
+            var.name[-16:] not in "/LARSOptimizer:0"
+        )
+        # yapf: enable
+    ]
+
+
+def get_trainable_variables():
+    """Get a list of trainable TensorFlow variables.
+
+    Parameters
+    ----------
+    train_only : boolean
+        If True, only get the trainable variables.
+
+    Returns
+    -------
+    list of Tensor
+        A list of trainable TensorFlow variables
+
+    Examples
+    --------
+
+    """
+    if KERAS_MODELS or LooseVersion(tf.__version__) >= LooseVersion("2.0.0"):
+        logging.warning(
+            "In TF2.x, only trainable variables created with Keras Models are captured for logging.\n"
+            "In TF1.x, if any keras model is defined. Only variables created inside Keras Models will be logged."
+        )
+
+        var_list = list()
+
+        for model in KERAS_MODELS:
+            var_list.extend(model.trainable_variables)
+
+        # Keep only a list of unique variables (remove potential duplicates)
+        var_list = list(set(var_list))
+
+        # clearing the list of Keras Model to avoid memory leaks
+        KERAS_MODELS.clear()
+
+        return [var for var in sorted(var_list, key=lambda v: v.name)]
+
+    else:
+        # return tf.trainable_variables()  # deprecated in TF2.x
+        from tensorflow.python.keras.backend import get_graph
+        return get_graph().get_collection('trainable_variables')
+
+
+def setup_tensorflow_hook(sess, logging_proxy, is_training, is_initialized):
+
+    global_step = -1
+
+    if is_training:
+
+        if not is_initialized:
+
+            _global_step_tensor = tf.compat.v1.train.get_or_create_global_step()
+
+            global_step = sess.run(_global_step_tensor)
+
+            trainable_variables = get_trainable_variables()
+
+            def count_weights_in_varlist(var_list):
+                return np.sum([np.prod(s.get_shape()) for s in var_list])
+
+            logging_proxy.log_git_status()
+
+            logging_proxy.log_model_statistics(
+                model_statistics={
+                    "# Trainable Weights": "{:,}".format(int(count_weights_in_varlist(trainable_variables))),
+                    "# Model Weights": "{:,}".format(int(count_weights_in_varlist(get_model_variables()))),
+                }
+            )
+
+            logging_proxy.log_trainable_variables([(var.name, var.get_shape()) for var in trainable_variables])
+
+    else:
+
+        if not is_initialized:
+            global_step = 0
+
+    metrics = collect_registered_metrics()
+
+    logging_proxy.log_runtime(is_train=is_training)
+
+    return global_step, metrics
+
+
+AutoLoggingHook = lambda *args, **kwargs: real_autologging_hook(*args, **kwargs)

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hooks/pretrained_restore_hook.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import sys
+import re
+import tensorflow as tf
+
+from mask_rcnn.utils.logging_formatter import logging
+from mask_rcnn.utils.distributed_utils import MPI_rank
+
+__all__ = ["PretrainedWeightsLoadingHook"]
+
+
+# pylint: disable=protected-access
+# Currently variable_scope doesn't provide very good APIs to access
+# all variables under scope and retrieve and check existing scopes.
+def get_variable_full_name(var):
+    """Returns the full name of a variable.
+    For normal Variables, this is the same as the var.op.name.  For
+    sliced or PartitionedVariables, this name is the same for all the
+    slices/partitions. In both cases, this is normally the name used in
+    a checkpoint file.
+    Args:
+    var: A `Variable` object.
+    Returns:
+    A string that is the full name.
+    """
+    if var._save_slice_info:
+        return var._save_slice_info.full_name
+    else:
+        return var.op.name
+
+
+def assign_from_checkpoint(model_path, var_list, ignore_missing_vars=False):
+    """Creates an operation to assign specific variables from a checkpoint.
+    Args:
+    model_path: The full path to the model checkpoint. To get latest checkpoint
+      use `model_path = tf.train.latest_checkpoint(checkpoint_dir)`
+    var_list: A list of (possibly partitioned) `Variable` objects or a
+      dictionary mapping names in the checkpoint to the corresponding variables
+      or list of variables to initialize from that checkpoint value. For
+      partitioned Variables, the name in the checkpoint must be the full
+      variable, not the name of the partitioned variable, eg. "my_var" rather
+      than "my_var/part_4". If empty, returns no_op(), {}.
+    ignore_missing_vars: Boolean, if True ignore variables missing in the
+      checkpoint with a warning instead of failing.
+    Returns:
+    the restore_op and the feed_dict that need to be run to restore var_list.
+    Raises:
+    ValueError: If `ignore_missing_vars` is False and the checkpoint specified
+        at `model_path` is missing one of the variables in `var_list`.
+  """
+    # Normalize var_list into a dictionary mapping names in the
+    # checkpoint to the list of variables to initialize from that
+    # checkpoint variable. Sliced (including partitioned) variables will
+    # end up under the same key.
+    grouped_vars = {}
+    if isinstance(var_list, (tuple, list)):
+        for var in var_list:
+            ckpt_name = get_variable_full_name(var)
+            if ckpt_name not in grouped_vars:
+                grouped_vars[ckpt_name] = []
+            grouped_vars[ckpt_name].append(var)
+
+    else:
+        for ckpt_name, value in var_list.items():
+            if isinstance(value, (tuple, list)):
+                grouped_vars[ckpt_name] = value
+            else:
+                grouped_vars[ckpt_name] = [value]
+
+    # Read each checkpoint entry. Create a placeholder variable and
+    # add the (possibly sliced) data from the checkpoint to the feed_dict.
+    reader = tf.compat.v1.train.NewCheckpointReader(model_path)
+    feed_dict = {}
+    assign_ops = []
+    for ckpt_name in grouped_vars:
+        if not reader.has_tensor(ckpt_name):
+            log_str = 'Checkpoint is missing variable [%s]' % ckpt_name
+            if ignore_missing_vars:
+                logging.warning(log_str)
+                continue
+            else:
+                raise ValueError(log_str)
+        ckpt_value = reader.get_tensor(ckpt_name)
+
+        for var in grouped_vars[ckpt_name]:
+            placeholder_tensor = tf.compat.v1.placeholder(
+                dtype=var.dtype.base_dtype,
+                shape=var.get_shape(),
+                name='placeholder/' + var.op.name
+            )
+
+            assign_ops.append(var.assign(placeholder_tensor))
+
+            if not var._save_slice_info:
+                if var.get_shape() != ckpt_value.shape:
+                    raise ValueError(
+                        'Total size of new array must be unchanged for %s '
+                        'lh_shape: [%s], rh_shape: [%s]' %
+                        (ckpt_name, str(ckpt_value.shape), str(var.get_shape())))
+
+                feed_dict[placeholder_tensor] = ckpt_value.reshape(ckpt_value.shape)
+
+            else:
+                slice_dims = zip(var._save_slice_info.var_offset,
+                                 var._save_slice_info.var_shape)
+
+                slice_dims = [(start, start + size) for (start, size) in slice_dims]
+                slice_dims = [slice(*x) for x in slice_dims]
+
+                slice_value = ckpt_value[slice_dims]
+                slice_value = slice_value.reshape(var._save_slice_info.var_shape)
+
+                feed_dict[placeholder_tensor] = slice_value
+
+    print_op = tf.print(
+        "[GPU %02d] Restoring pretrained weights (%d Tensors) from: %s" % (
+            MPI_rank(),
+            len(assign_ops),
+            model_path
+        ),
+        output_stream=sys.stdout
+    )
+
+    with tf.control_dependencies([print_op]):
+        assign_op = tf.group(*assign_ops)
+
+    return assign_op, feed_dict
+
+
+def build_assigment_map(prefix=None, skip_variables_regex=None):
+    """Generate assigment map for loading checkpoints."""
+
+    all_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.GLOBAL_VARIABLES, scope=prefix)
+
+    if not prefix:
+        prefix = ''
+
+    assignment_map = {}
+
+    for var in all_vars:
+
+        var_name = var.name
+
+        if (
+                var_name[-11:] in "/Momentum:0" or
+                var_name[-11:] in "/Adadelta:0" or
+                var_name[-13:] in "/Adadelta_1:0" or
+                var_name[-7:] in "/Adam:0" or
+                var_name[-9:] in "/Adam_1:0" or
+                var_name[-10:] in "/Adagrad:0" or
+                var_name[-10:] in "/RMSProp:0" or
+                var_name[-12:] in "/RMSProp_1:0" or
+                var_name[-16:] in "/LARSOptimizer:0"
+        ):
+            continue
+
+        # Trim the index of the variable.
+        if ':' in var_name:
+            var_name = var_name[:var_name.rindex(':')]
+
+        if skip_variables_regex and re.match(skip_variables_regex, var_name[len(prefix):]):
+            continue
+
+        assignment_map[var_name[len(prefix):]] = var
+        # assignment_map[var_name] = var
+
+    return assignment_map
+
+
+class PretrainedWeightsLoadingHook(tf.estimator.SessionRunHook):
+
+    def __init__(self, prefix, checkpoint_path, skip_variables_regex=None):
+        self._prefix = prefix
+        self._checkpoint_path = checkpoint_path
+        self._skip_variables_regex = skip_variables_regex
+
+        self._is_initialized = False
+
+        self._init_op = None
+        self._init_feed_dict = None
+
+    def begin(self):
+        vars_to_load = build_assigment_map(
+            prefix=self._prefix,
+            skip_variables_regex=self._skip_variables_regex
+        )
+
+        self._init_op, self._init_feed_dict = assign_from_checkpoint(
+            model_path=self._checkpoint_path,
+            var_list=vars_to_load,
+            ignore_missing_vars=False
+        )
+
+    def after_create_session(self, session, coord=None):
+        if not self._is_initialized:
+            session.run(self._init_op, feed_dict=self._init_feed_dict)
+            logging.info("Pretrained weights loaded with success...\n")
+
+            self._is_initialized = True

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hyperparameters/cmdline_utils.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Defining common model params used across all the models."""
+
+from absl import flags
+
+
+def define_hparams_flags():
+
+    flags.DEFINE_string(
+        'log_path',
+        default="./mrcnn.json",
+        help=(
+            'The path where dllogger json file will be saved. Please include the'
+            ' name of the json file as well.'
+        )
+    )
+
+    flags.DEFINE_string(
+        'data_dir',
+        default=None,
+        help=(
+            'The directory where the input data is stored. Please see the model'
+            ' specific README.md for the expected data format.'
+        )
+    )
+
+    flags.DEFINE_string('checkpoint', default='', help='Checkpoint filepath')
+
+    flags.DEFINE_integer(
+        'eval_batch_size',
+        default=8,
+        help='Batch size for evaluation.'
+    )
+
+    flags.DEFINE_bool(
+        'eval_after_training',
+        default=True,
+        help='Run one eval after the training finishes.'
+    )
+
+    flags.DEFINE_integer('eval_samples', default=5000, help='Number of training steps')
+
+    flags.DEFINE_bool(
+        'include_groundtruth_in_features',
+        default=False,
+        help=(
+            'If `val_json_file` is not provided, one can also read groundtruth'
+            ' from input by setting `include_groundtruth_in_features`=True'
+        )
+    )
+
+    # Gradient clipping is a fairly coarse heuristic to stabilize training.
+    # This model clips the gradient by its L2 norm globally (i.e., across
+    # all variables), using a threshold obtained from multiplying this
+    # parameter with sqrt(number_of_weights), to have a meaningful value
+    # across both training phases and different sizes of imported modules.
+    # Refer value: 0.02, for 25M weights, yields clip norm 10.
+    # Zero or negative number means no clipping.
+    flags.DEFINE_float("global_gradient_clip_ratio", default=-1.0, help="Global Gradient Clipping Ratio")
+
+    flags.DEFINE_float("init_learning_rate", default=2.5e-3, help="Initial Learning Rate")
+
+    flags.DEFINE_float("warmup_learning_rate", default=0., help="Warmup Learning Rate Decay Factor")
+
+    flags.DEFINE_bool('finetune_bn', False, 'is batchnorm training mode')
+
+    flags.DEFINE_float("l2_weight_decay", default=1e-4, help="l2 regularization weight")
+
+    flags.DEFINE_string('mode', default='train_and_eval', help='Mode to run: train or eval')
+
+    flags.DEFINE_string(
+        'model_dir',
+        default=None,
+        help='The directory where the model and training/evaluation summaries are stored.'
+    )
+
+    flags.DEFINE_float("momentum", default=0.9, help="Optimizer Momentum")
+
+    flags.DEFINE_integer('num_steps_per_eval', default=2500, help='Number of steps per evaluation epoch.')
+
+    flags.DEFINE_integer('save_checkpoints_steps', default=2500, help='Save a checkpoint every N steps.')
+
+    flags.DEFINE_integer('seed', default=None, help='Set a debug seed for reproducibility.')
+
+    flags.DEFINE_integer('train_batch_size', default=2, help='Batch size for training.')
+
+    flags.DEFINE_integer(
+        'total_steps',
+        default=938240,
+        help=(
+            'The number of steps to use for training. This flag'
+            ' should be adjusted according to the --train_batch_size flag.'
+        )
+    )
+
+    flags.DEFINE_list(
+        'learning_rate_decay_levels',
+        default=['0.1', '0.01'],
+        help=(
+            'The learning rate decay levels which modify the learning rate using the formula:'
+            ' `lr = decay * init_lr`. Decay factor applied at learning_rate_steps.'
+        )
+    )
+    flags.DEFINE_list(
+        'learning_rate_steps',
+        default=['480000', '640000'],
+        help=(
+            'The steps at which learning rate changes. This flag'
+            ' should be adjusted according to the --train_batch_size flag.'
+        )
+    )
+    flags.DEFINE_integer('warmup_steps', default=1000, help='The number of steps to use warmup learning rate for')
+
+    flags.DEFINE_bool('amp', default=False, help='Enable automatic mixed precision')
+
+    flags.DEFINE_bool(
+        'use_batched_nms',
+        default=False,
+        help='Enable Batched NMS at inference.'
+    )
+
+    flags.DEFINE_bool(
+        'use_custom_box_proposals_op',
+        default=False,
+        help='Use GenerateBoundingBoxProposals op.'
+    )
+
+    flags.DEFINE_bool('use_fake_data', False, 'Use fake input.')
+
+    flags.DEFINE_bool(
+        'use_tf_distributed',
+        default=False,
+        help='Use tensorflow distributed API'
+    )
+
+    flags.DEFINE_bool('xla', default=False, help='Enable XLA JIT Compiler.')
+
+    flags.DEFINE_string('training_file_pattern', default="", help='TFRecords file pattern for the training files')
+
+    flags.DEFINE_string('validation_file_pattern', default="", help='TFRecords file pattern for the validation files')
+
+    flags.DEFINE_string('val_json_file', default="", help='Filepath for the validation json file')
+
+    ############################# TO BE REMOVED ###################################
+
+    flags.DEFINE_integer(
+        'report_frequency',
+        default=None,
+        help='The amount of batches in between accuracy reports at evaluation time'
+    )
+
+    ############################# TO BE REMOVED ###################################
+
+    ############################### ISSUES TO FIX - FLAGS #############################"
+
+    # TODO: Remove when XLA at inference fixed
+    flags.DEFINE_bool(
+        'allow_xla_at_inference',
+        default=False,
+        help='Enable XLA JIT Compiler at Inference'
+    )
+
+    return flags.FLAGS

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hyperparameters/flags_to_params.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Functions to override model parameters from command-line flags."""
+
+from mask_rcnn.hyperparameters import params_dict
+
+ESSENTIAL_FLAGS = ['tpu', 'data_dir', 'model_dir']
+
+
+def override_params_from_input_flags(params, input_flags):
+    """Update params dictionary with input flags.
+
+  Args:
+    params: ParamsDict object containing dictionary of model parameters.
+    input_flags: All the flags with non-null value of overridden model
+    parameters.
+
+  Returns:
+    ParamsDict object containing dictionary of model parameters.
+  """
+    if params is None:
+        raise ValueError('Input dictionary is empty. It is expected to be loaded with default ' 'values')
+
+    if not isinstance(params, params_dict.ParamsDict):
+        raise ValueError('The base parameter set must be a ParamsDict, was: {}'.format(type(params)))
+
+    essential_flag_dict = {}
+    for key in ESSENTIAL_FLAGS:
+        flag_value = input_flags.get_flag_value(key, None)
+
+        if flag_value is None:
+            raise ValueError('Flag {} could not be None.'.format(key))
+        else:
+            essential_flag_dict[key] = flag_value
+
+    params_dict.override_params_dict(params, essential_flag_dict, is_strict=False)
+
+    normal_flag_dict = get_dictionary_from_flags(params.as_dict(), input_flags)
+
+    params_dict.override_params_dict(params, normal_flag_dict, is_strict=False)
+
+    return params
+
+
+def get_dictionary_from_flags(params, input_flags):
+    """Generate dictionary from non-null flags.
+
+  Args:
+    params: Python dictionary of model parameters.
+    input_flags: All the flags with non-null value of overridden model
+    parameters.
+
+  Returns:
+    Python dict of overriding model parameters.
+  """
+    flag_dict = {}
+    for k, v in params.items():
+        if isinstance(v, dict):
+            d = get_dictionary_from_flags(v, input_flags)
+            flag_dict[k] = d
+        else:
+            try:
+                flag_value = input_flags.get_flag_value(k, None)
+                if flag_value is not None:
+                    flag_dict[k] = flag_value
+            except AttributeError:
+                flag_dict[k] = v
+
+    return flag_dict

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hyperparameters/hyperparameters.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+import os
+import warnings
+import six
+
+import yaml
+
+import tensorflow as tf
+
+from mask_rcnn.utils.logging_formatter import logging
+
+
+class _Hyperparameters(object):
+    """_Hyperparameters class to generate final hparams from various inputs."""
+
+    def __init__(self, default_hparams_file, specific_hparams_file, input_flags, hparams_overrides):
+        """Initialze and load parameter dictionary with different input sources.
+
+    Args:
+      default_hparams_file: YAML storing default values of all hyperparameters.
+      specific_hparams_file: YAML file storing accelerator specific values of
+      hyperparameters to override the default values.
+      input_flags: Command line flags values for hyperparameters. [This is
+      for backward compatibility, so that users passing hyperparameters as
+      regular flags should not run into trouble].
+      hparams_overrides: A kv string representing which hyperparameters need to
+      be override from the command-line.
+
+    Raises:
+      ValueError: Raised when 'default_hparams_file' is not readable.
+    """
+        if not tf.io.gfile.exists(default_hparams_file):
+            raise ValueError(
+                'Expected a valid path to a YAML file, which represents the default '
+                'hyperparameters file. {}'.format(default_hparams_file)
+            )
+
+        self._params = {}
+        self._params_source = {}
+        self._default_hparams_file = default_hparams_file
+        self._specific_hparams_file = specific_hparams_file
+        self._input_flags = input_flags
+        self._hparams_overrides = hparams_overrides
+
+    def get_parameters(self, log_params):
+        """Returns the dictionary loaded with final values of all hyperparameters.
+
+    Args:
+      log_params: Bool to specify if the hyperparameters final value need to be
+      logged or not.
+
+    Returns:
+      Python dictionary with all the final hyperparameters.
+    """
+        self._params, self._params_source = load_from_file(
+            self._params, self._params_source, self._default_hparams_file
+        )
+        self._params, self._params_source = load_from_file(
+            self._params, self._params_source, self._specific_hparams_file
+        )
+        self._params, self._params_source = load_from_input_flags(self._params, self._params_source, self._input_flags)
+        self._params, self._params_source = load_from_hparams_overrides(
+            self._params, self._params_source, self._hparams_overrides
+        )
+
+        if log_params:
+            self.log_parameters()
+
+        return self._params
+
+    def log_parameters(self):
+        """Log the hyperparameters value along with the source of those values.
+    """
+        params_log = ''
+
+        for k in self._params:
+            params_log += k + ': \t' + str(self._params[k])
+            params_log += ' \t[' + self._params_source[k] + ']\n'
+
+        logging.info('\nModel hyperparameters [source]:\n%s', params_log)
+
+
+def load_from_file(params, params_source, file_path):
+    """Given a path to a YAML file, read the file and load it to dictionary.
+
+  Args:
+    params: Python dictionary of hyperparameters.
+    params_source: Python dictionary to record source of hyperparameters.
+    file_path: Python string containing path to file.
+
+  Returns:
+    Python dict of hyperparameters.
+  """
+    if file_path is None:
+        return params, params_source
+
+    if not tf.io.gfile.exists(file_path):
+        warnings.warn('Could not read Hyperparameter file : ' + file_path, RuntimeWarning)
+        return params, params_source
+
+    with tf.io.gfile.GFile(file_path, 'r') as f:
+        overrides = yaml.load(f)
+    for key, value in six.iteritems(overrides):
+        params[key] = value
+        params_source[key] = os.path.basename(file_path)
+
+    return params, params_source
+
+
+# TODO(amangu): Once global hyperparameter flags will be removed, we won't need
+# this function. Remove this functions after implementing this.
+def load_from_input_flags(params, params_source, input_flags):
+    """Update params dictionary with input flags.
+
+  Args:
+    params: Python dictionary of hyperparameters.
+    params_source: Python dictionary to record source of hyperparameters.
+    input_flags: All the flags with non-null value of overridden
+    hyperparameters.
+
+  Returns:
+    Python dict of hyperparameters.
+  """
+    if params is None:
+        raise ValueError('Input dictionary is empty. It is expected to be loaded with default ' 'values')
+
+    if not isinstance(params, dict):
+        raise ValueError('The base parameter set must be a Python dict, was: {}'.format(type(params)))
+
+    for key in params:
+        flag_value = input_flags.get_flag_value(key, None)
+
+        if flag_value is not None:
+            params[key] = flag_value
+            params_source[key] = 'Command-line flags'
+
+    return params, params_source
+
+
+# TODO(amangu): Add tests to verify different dtypes of params.
+def load_from_hparams_overrides(params, params_source, hparams_overrides):
+    """Given a dictionary of hyperparameters and a list of overrides, merge them.
+
+  Args:
+    params: Python dict containing a base hyperparameters set.
+    params_source: Python dictionary to record source of hyperparameters.
+    hparams_overrides: Python list of strings. This is a set of k=v overrides
+    for the hyperparameters in `params`; if `k=v1` in `params` but `k=v2` in
+    `hparams_overrides`, the second value wins and the value for `k` is `v2`.
+
+  Returns:
+    Python dict of hyperparameters.
+  """
+    if params is None:
+        raise ValueError('Input dictionary is empty. It is expected to be loaded with default ' 'values')
+
+    if not isinstance(params, dict):
+        raise ValueError('The base hyperparameters set must be a Python dict, was: {}'.format(type(params)))
+
+    if hparams_overrides is None:
+        return params, params_source
+
+    if isinstance(hparams_overrides, six.string_types):
+        hparams_overrides = [hparams_overrides]
+
+    if not isinstance(hparams_overrides, list):
+        raise ValueError(
+            'Expected that hparams_overrides would be `None`, a single string, or a'
+            ' list of strings, was: {}'.format(type(hparams_overrides))
+        )
+
+    for kv_pair in hparams_overrides:
+        if not isinstance(kv_pair, six.string_types):
+            raise ValueError(
+                'Expected that hparams_overrides would contain Python list of strings,'
+                ' but encountered an item: {}'.format(type(kv_pair))
+            )
+        key, value = kv_pair.split('=')
+        parser = type(params[key])
+        if parser is bool:
+            params[key] = value not in ('0', 'False', 'false')
+        else:
+            params[key] = parser(value)
+        params_source[key] = 'Command-line `hparams` flag'
+
+    return params, params_source
+
+
+def get_hyperparameters(default_hparams_file, specific_hparams_file, input_flags, hparams_overrides, log_params=True):
+    """Single function to get hparams for any model using different sources.
+
+  Args:
+    default_hparams_file: YAML storing default values of all hyperparameters.
+    specific_hparams_file: YAML file storing accelerator specific values of
+    hyperparameters to override the default values.
+    input_flags: Command line flags values for hyperparameters. [This is
+    for backward compatibility, so that users passing hyperparameters as
+    regular flags should not run into trouble].
+    hparams_overrides: A kv string representing which hyperparameters need to
+    be override from the command-line.
+    log_params: Bool to specify if the hyperparameters final value need to be
+    logged or not.
+
+  Returns:
+    Python dictionary with all the final hyperparameters.
+  """
+    parameter = _Hyperparameters(default_hparams_file, specific_hparams_file, input_flags, hparams_overrides)
+
+    return parameter.get_parameters(log_params)

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hyperparameters/mask_rcnn_params.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Parameters used to build Mask-RCNN model."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from argparse import Namespace
+
+
+class _Namespace(Namespace):
+    def values(self):
+        return self.__dict__
+
+
+def default_config():
+    return _Namespace(**dict(
+        # input pre-processing parameters
+        image_size=(832, 1344),
+        augment_input_data=True,
+        gt_mask_size=112,
+
+        # dataset specific parameters
+        num_classes=91,
+        # num_classes=81,
+        skip_crowd_during_training=True,
+        use_category=True,
+
+        # Region Proposal Network
+        rpn_positive_overlap=0.7,
+        rpn_negative_overlap=0.3,
+        rpn_batch_size_per_im=256,
+        rpn_fg_fraction=0.5,
+        rpn_min_size=0.,
+
+        # Proposal layer.
+        batch_size_per_im=512,
+        fg_fraction=0.25,
+        fg_thresh=0.5,
+        bg_thresh_hi=0.5,
+        bg_thresh_lo=0.,
+
+        # Faster-RCNN heads.
+        fast_rcnn_mlp_head_dim=1024,
+        bbox_reg_weights=(10., 10., 5., 5.),
+
+        # Mask-RCNN heads.
+        include_mask=True,  # whether or not to include mask branch.   # ===== Not existing in MLPerf ===== #
+        mrcnn_resolution=28,
+
+        # training
+        train_rpn_pre_nms_topn=2000,
+        train_rpn_post_nms_topn=1000,
+        train_rpn_nms_threshold=0.7,
+
+        # evaluation
+        test_detections_per_image=100,
+        test_nms=0.5,
+        test_rpn_pre_nms_topn=1000,
+        test_rpn_post_nms_topn=1000,
+        test_rpn_nms_thresh=0.7,
+
+        # model architecture
+        min_level=2,
+        max_level=6,
+        num_scales=1,
+        aspect_ratios=[(1.0, 1.0), (1.4, 0.7), (0.7, 1.4)],
+        anchor_scale=8.0,
+
+        # localization loss
+        rpn_box_loss_weight=1.0,
+        fast_rcnn_box_loss_weight=1.0,
+        mrcnn_weight_loss_mask=1.0,
+
+        # ---------- Training configurations ----------
+
+        # Skips loading variables from the resnet checkpoint. It is used for
+        # skipping nonexistent variables from the constructed graph. The list
+        # of loaded variables is constructed from the scope 'resnetX', where 'X'
+        # is depth of the resnet model. Supports regular expression.
+        skip_checkpoint_variables='^NO_SKIP$',
+
+        # ---------- Eval configurations ----------
+        # Visualizes images and detection boxes on TensorBoard.
+        visualize_images_summary=False,
+    ))

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hyperparameters/params_dict.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""A parameter dictionary class which supports the nest structure."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import copy
+import re
+import six
+import yaml
+
+import tensorflow as tf
+
+# regex pattern that matches on key-value pairs in a comma-separated
+# key-value pair string. It splits each k-v pair on the = sign, and
+# matches on values that are within single quotes, double quotes, single
+# values (e.g. floats, ints, etc.), and a lists within brackets.
+_PARAM_RE = re.compile(
+    r"""
+  (?P<name>[a-zA-Z][\w\.]*)    # variable name: "var" or "x"
+  \s*=\s*
+  ((?P<val>\'[^\]]*\'          # single quote
+  |
+  \"[^\]]*\"                   # double quote
+  |
+  [^,\[]*                      # single value
+  |
+  \[[^\]]*\]))                 # list of values
+  ($|,\s*)""", re.VERBOSE
+)
+
+
+class ParamsDict(object):
+    """A hyperparameter container class."""
+
+    RESERVED_ATTR = ['_locked', '_restrictions']
+
+    def __init__(self, default_params=None, restrictions=None):
+        """Instantiate a ParamsDict.
+
+    Instantiate a ParamsDict given a set of default parameters and a list of
+    restrictions. Upon initialization, it validates itself by checking all the
+    defined restrictions, and raise error if it finds inconsistency.
+
+    Args:
+      default_params: a Python dict or another ParamsDict object including the
+        default parameters to initialize.
+      restrictions: a list of strings, which define a list of restrictions to
+        ensure the consistency of different parameters internally. Each
+        restriction string is defined as a binary relation with a set of
+        operators, including {'==', '!=',  '<', '<=', '>', '>='}.
+    """
+        self._locked = False
+        self._restrictions = []
+        if restrictions:
+            self._restrictions = restrictions
+        if default_params is None:
+            default_params = {}
+        self.override(default_params, is_strict=False)
+        self.validate()
+
+    def _set(self, k, v):
+        if isinstance(v, dict):
+            self.__dict__[k] = ParamsDict(v)
+        else:
+            self.__dict__[k] = copy.deepcopy(v)
+
+    def __setattr__(self, k, v):
+        """Sets the value of the existing key.
+
+    Note that this does not allow directly defining a new key. Use the
+    `override` method with `is_strict=False` instead.
+
+    Args:
+      k: the key string.
+      v: the value to be used to set the key `k`.
+
+    Raises:
+      KeyError: if k is not defined in the ParamsDict.
+    """
+        if k not in ParamsDict.RESERVED_ATTR:
+            if k not in self.__dict__.keys():
+                raise KeyError(
+                    'The key `%{}` does not exist. '
+                    'To extend the existing keys, use '
+                    '`override` with `is_strict` = True.'.format(k)
+                )
+            if self._locked:
+                raise ValueError('The ParamsDict has been locked. ' 'No change is allowed.')
+        self._set(k, v)
+
+    def __getattr__(self, k):
+        """Gets the value of the existing key.
+
+    Args:
+      k: the key string.
+
+    Returns:
+      the value of the key.
+
+    Raises:
+      KeyError: if k is not defined in the ParamsDict.
+    """
+        if k not in self.__dict__.keys():
+            raise KeyError('The key `{}` does not exist. '.format(k))
+        return self.__dict__[k]
+
+    def override(self, override_params, is_strict=True):
+        """Override the ParamsDict with a set of given params.
+
+    Args:
+      override_params: a dict or a ParamsDict specifying the parameters to
+        be overridden.
+      is_strict: a boolean specifying whether override is strict or not. If
+        True, keys in `override_params` must be present in the ParamsDict.
+        If False, keys in `override_params` can be different from what is
+        currently defined in the ParamsDict. In this case, the ParamsDict will
+        be extended to include the new keys.
+    """
+        if self._locked:
+            raise ValueError('The ParamsDict has been locked. No change is allowed.')
+        if isinstance(override_params, ParamsDict):
+            override_params = override_params.as_dict()
+        self._override(override_params, is_strict)  # pylint: disable=protected-access
+
+    def _override(self, override_dict, is_strict=True):
+        """The implementation of `override`."""
+        for k, v in six.iteritems(override_dict):
+            if k in ParamsDict.RESERVED_ATTR:
+                raise KeyError('The key `%{}` is internally reserved. ' 'Can not be overridden.')
+            if k not in self.__dict__.keys():
+                if is_strict:
+                    raise KeyError(
+                        'The key `{}` does not exist. '
+                        'To extend the existing keys, use '
+                        '`override` with `is_strict` = False.'.format(k)
+                    )
+                else:
+                    self._set(k, v)
+            else:
+                if isinstance(v, dict):
+                    self.__dict__[k]._override(v, is_strict)  # pylint: disable=protected-access
+                elif isinstance(v, ParamsDict):
+                    self.__dict__[k]._override(v.as_dict(), is_strict)  # pylint: disable=protected-access
+                else:
+                    self.__dict__[k] = copy.deepcopy(v)
+
+    def lock(self):
+        """Makes the ParamsDict immutable."""
+        self._locked = True
+
+    def as_dict(self):
+        """Returns a dict representation of ParamsDict.
+
+    For the nested ParamsDict, a nested dict will be returned.
+    """
+        params_dict = {}
+        for k, v in six.iteritems(self.__dict__):
+            if k not in ParamsDict.RESERVED_ATTR:
+                if isinstance(v, ParamsDict):
+                    params_dict[k] = v.as_dict()
+                else:
+                    params_dict[k] = copy.deepcopy(v)
+        return params_dict
+
+    def validate(self):
+        """Validate the parameters consistency based on the restrictions.
+
+    This method validates the internal consistency using the pre-defined list of
+    restrictions. A restriction is defined as a string which specfiies a binary
+    operation. The supported binary operations are {'==', '!=', '<', '<=', '>',
+    '>='}. Note that the meaning of these operators are consistent with the
+    underlying Python immplementation. Users should make sure the define
+    restrictions on their type make sense.
+
+    For example, for a ParamsDict like the following
+    ```
+    a:
+      a1: 1
+      a2: 2
+    b:
+      bb:
+        bb1: 10
+        bb2: 20
+      ccc:
+        a1: 1
+        a3: 3
+    ```
+    one can define two restrictions like this
+    ['a.a1 == b.ccc.a1', 'a.a2 <= b.bb.bb2']
+
+    What it enforces are:
+     - a.a1 = 1 == b.ccc.a1 = 2
+     - a.a2 = 2 <= b.bb.bb2 = 20
+
+    Raises:
+      KeyError: if any of the following happens
+        (1) any of parameters in any of restrictions is not defined in
+            ParamsDict,
+        (2) any inconsistency violating the restriction is found.
+      ValueError: if the restriction defined in the string is not supported.
+    """
+
+        def _get_kv(dotted_string, params_dict):
+            tokenized_params = dotted_string.split('.')
+            v = params_dict
+            for t in tokenized_params:
+                v = v[t]
+            return tokenized_params[-1], v
+
+        def _get_kvs(tokens, params_dict):
+            if len(tokens) != 2:
+                raise ValueError('Only support binary relation in restriction.')
+            stripped_tokens = [t.strip() for t in tokens]
+            left_k, left_v = _get_kv(stripped_tokens[0], params_dict)
+            right_k, right_v = _get_kv(stripped_tokens[1], params_dict)
+            return left_k, left_v, right_k, right_v
+
+        params_dict = self.as_dict()
+        for restriction in self._restrictions:
+            if '==' in restriction:
+                tokens = restriction.split('==')
+                _, left_v, _, right_v = _get_kvs(tokens, params_dict)
+                if left_v != right_v:
+                    raise KeyError('Found inconsistncy between key `{}` and key `{}`.'.format(tokens[0], tokens[1]))
+            elif '!=' in restriction:
+                tokens = restriction.split('!=')
+                _, left_v, _, right_v = _get_kvs(tokens, params_dict)
+                if left_v == right_v:
+                    raise KeyError('Found inconsistncy between key `{}` and key `{}`.'.format(tokens[0], tokens[1]))
+            elif '<' in restriction:
+                tokens = restriction.split('<')
+                _, left_v, _, right_v = _get_kvs(tokens, params_dict)
+                if left_v >= right_v:
+                    raise KeyError('Found inconsistncy between key `{}` and key `{}`.'.format(tokens[0], tokens[1]))
+            elif '<=' in restriction:
+                tokens = restriction.split('<=')
+                _, left_v, _, right_v = _get_kvs(tokens, params_dict)
+                if left_v > right_v:
+                    raise KeyError('Found inconsistncy between key `{}` and key `{}`.'.format(tokens[0], tokens[1]))
+            elif '>' in restriction:
+                tokens = restriction.split('>')
+                _, left_v, _, right_v = _get_kvs(tokens, params_dict)
+                if left_v <= right_v:
+                    raise KeyError('Found inconsistncy between key `{}` and key `{}`.'.format(tokens[0], tokens[1]))
+            elif '>=' in restriction:
+                tokens = restriction.split('>=')
+                _, left_v, _, right_v = _get_kvs(tokens, params_dict)
+                if left_v < right_v:
+                    raise KeyError('Found inconsistncy between key `{}` and key `{}`.'.format(tokens[0], tokens[1]))
+            else:
+                raise ValueError('Unsupported relation in restriction.')
+
+
+def read_yaml_to_params_dict(file_path):
+    """Reads a YAML file to a ParamsDict."""
+    with tf.io.gfile.GFile(file_path, 'r') as f:
+        params_dict = yaml.load(f)
+        return ParamsDict(params_dict)
+
+
+def save_params_dict_to_yaml(params, file_path):
+    """Saves the input ParamsDict to a YAML file."""
+    with tf.io.gfile.GFile(file_path, 'w') as f:
+
+        def _my_list_rep(dumper, data):
+            # u'tag:yaml.org,2002:seq' is the YAML internal tag for sequence.
+            return dumper.represent_sequence(u'tag:yaml.org,2002:seq', data, flow_style=True)
+
+        yaml.add_representer(list, _my_list_rep)
+        yaml.dump(params.as_dict(), f, default_flow_style=False)
+
+
+def nested_csv_str_to_json_str(csv_str):
+    """Converts a nested (using '.') comma-separated k=v string to a JSON string.
+
+  Converts a comma-separated string of key/value pairs that supports
+  nesting of keys to a JSON string. Nesting is implemented using
+  '.' between levels for a given key.
+
+  Spacing between commas and = is supported (e.g. there is no difference between
+  "a=1,b=2", "a = 1, b = 2", or "a=1, b=2") but there should be no spaces before
+  keys or after values (e.g. " a=1,b=2" and "a=1,b=2 " are not supported).
+
+  Note that this will only support values supported by CSV, meaning
+  values such as nested lists (e.g. "a=[[1,2,3],[4,5,6]]") are not
+  supported. Strings are supported as well, e.g. "a='hello'".
+
+  An example conversion would be:
+
+  "a=1, b=2, c.a=2, c.b=3, d.a.a=5"
+
+  to
+
+  "{ a: 1, b : 2, c: {a : 2, b : 3}, d: {a: {a : 5}}}"
+
+  Args:
+    csv_str: the comma separated string.
+
+  Returns:
+    the converted JSON string.
+
+  Raises:
+    ValueError: If csv_str is not in a comma separated string or
+      if the string is formatted incorrectly.
+  """
+    if not csv_str:
+        return ''
+
+    formatted_entries = []
+    nested_map = collections.defaultdict(list)
+    pos = 0
+    while pos < len(csv_str):
+        m = _PARAM_RE.match(csv_str, pos)
+        if not m:
+            raise ValueError('Malformed hyperparameter value while parsing ' 'CSV string: %s' % csv_str[pos:])
+        pos = m.end()
+        # Parse the values.
+        m_dict = m.groupdict()
+        name = m_dict['name']
+        v = m_dict['val']
+
+        name_nested = name.split('.')
+        if len(name_nested) > 1:
+            grouping = name_nested[0]
+            value = '.'.join(name_nested[1:]) + '=' + v
+            nested_map[grouping].append(value)
+        else:
+            formatted_entries.append('%s : %s' % (name, v))
+
+    for grouping, value in nested_map.items():
+        value = ','.join(value)
+        value = nested_csv_str_to_json_str(value)
+        formatted_entries.append('%s : %s' % (grouping, value))
+    return '{' + ', '.join(formatted_entries) + '}'
+
+
+def override_params_dict(params, dict_or_string_or_yaml_file, is_strict):
+    """Override a given ParamsDict using a dict, JSON/YAML/CSV string or YAML file.
+
+  The logic of the function is outlined below:
+  1. Test that the input is a dict. If not, proceed to 2.
+  2. Tests that the input is a string. If not, raise unknown ValueError
+  2.1. Test if the string is in a CSV format. If so, parse.
+  If not, proceed to 2.2.
+  2.2. Try loading the string as a YAML/JSON. If successful, parse to
+  dict and use it to override. If not, proceed to 2.3.
+  2.3. Try using the string as a file path and load the YAML file.
+
+  Args:
+    params: a ParamsDict object to be overridden.
+    dict_or_string_or_yaml_file: a Python dict, JSON/YAML/CSV string or
+      path to a YAML file specifying the parameters to be overridden.
+    is_strict: a boolean specifying whether override is strict or not.
+
+  Returns:
+    params: the overridden ParamsDict object.
+
+  Raises:
+    ValueError: if failed to override the parameters.
+  """
+    if not dict_or_string_or_yaml_file:
+        return params
+    if isinstance(dict_or_string_or_yaml_file, dict):
+        params.override(dict_or_string_or_yaml_file, is_strict)
+    elif isinstance(dict_or_string_or_yaml_file, six.string_types):
+        try:
+            dict_or_string_or_yaml_file = (nested_csv_str_to_json_str(dict_or_string_or_yaml_file))
+        except ValueError:
+            pass
+        params_dict = yaml.load(dict_or_string_or_yaml_file)
+        if isinstance(params_dict, dict):
+            params.override(params_dict, is_strict)
+        else:
+            with tf.io.gfile.GFile(dict_or_string_or_yaml_file) as f:
+                params.override(yaml.load(f), is_strict)
+    else:
+        raise ValueError('Unknown input type to parse.')
+    return params

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/hyperparameters/params_io.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#============================================================================
+
+"""Utils to handle parameters IO."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import yaml
+
+import tensorflow as tf
+
+
+def save_hparams_to_yaml(hparams, file_path):
+    with tf.io.gfile.GFile(file_path, 'w') as f:
+        try:
+            hparams_val = hparams.values()
+        except AttributeError:
+            hparams_val = hparams.__dict__
+        yaml.dump(hparams_val, f)
+
+
+def override_hparams(hparams, dict_or_string_or_yaml_file):
+    """Override a given hparams using a dict or a string or a JSON file.
+
+  Args:
+    hparams: a HParams object to be overridden.
+    dict_or_string_or_yaml_file: a Python dict, or a comma-separated string,
+      or a path to a YAML file specifying the parameters to be overridden.
+
+  Returns:
+    hparams: the overridden HParams object.
+
+  Raises:
+    ValueError: if failed to override the parameters.
+  """
+    if not dict_or_string_or_yaml_file:
+        return hparams
+
+    if isinstance(dict_or_string_or_yaml_file, dict):
+
+        for key, val in dict_or_string_or_yaml_file.items():
+
+            if key not in hparams:
+                try:  # TF 1.x
+                    hparams.add_hparam(key, val)
+                except AttributeError:  # TF 2.x
+                    try:  # Dict
+                        hparams[key] = val
+                    except TypeError:  # Namespace
+                        setattr(hparams, key, val)
+            else:
+                raise ValueError("Parameter `%s` is already defined" % key)
+
+        # hparams.override_from_dict(dict_or_string_or_yaml_file)
+
+    elif isinstance(dict_or_string_or_yaml_file, six.string_types):
+        try:
+            hparams.parse(dict_or_string_or_yaml_file)
+
+        except ValueError as parse_error:
+            try:
+                with tf.io.gfile.GFile(dict_or_string_or_yaml_file) as f:
+                    hparams.override_from_dict(yaml.load(f))
+
+            except Exception as read_error:
+                parse_message = ('Failed to parse config string: %s\n' % parse_error.message)
+                read_message = ('Failed to parse yaml file provided. %s' % read_error.message)
+                raise ValueError(parse_message + read_message)
+
+    else:
+        raise ValueError('Unknown input type to parse.')
+    return hparams

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/mask_rcnn_model.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Model definition for the Mask-RCNN Model.
+
+Defines model_fn of Mask-RCNN for TF Estimator. The model_fn includes Mask-RCNN
+model architecture, loss function, learning rate schedule, and evaluation
+procedure.
+
+"""
+
+import itertools
+
+import tensorflow as tf
+
+from mask_rcnn import anchors
+
+from mask_rcnn.models import fpn
+from mask_rcnn.models import heads
+from mask_rcnn.models import resnet
+
+from mask_rcnn.training import losses, learning_rates
+
+from mask_rcnn.ops import postprocess_ops
+from mask_rcnn.ops import roi_ops
+from mask_rcnn.ops import spatial_transform_ops
+from mask_rcnn.ops import training_ops
+
+from mask_rcnn.utils.logging_formatter import logging
+
+from mask_rcnn.utils.distributed_utils import MPI_is_distributed
+from mask_rcnn.utils.distributed_utils import MPI_local_rank
+
+from mask_rcnn.utils.meters import StandardMeter
+from mask_rcnn.utils.metric_tracking import register_metric
+
+from mask_rcnn.utils.lazy_imports import LazyImport
+hvd = LazyImport("horovod.tensorflow")
+
+MODELS = dict()
+
+
+
+
+def create_optimizer(learning_rate, params):
+    """Creates optimized based on the specified flags."""
+
+    optimizer = tf.compat.v1.train.MomentumOptimizer(learning_rate, momentum=params['momentum'])
+
+    if MPI_is_distributed():
+        optimizer = hvd.DistributedOptimizer(
+            optimizer,
+            name=None,
+            device_dense='/gpu:0',
+            device_sparse='',
+            # compression=hvd.Compression.fp16,
+            compression=hvd.Compression.none,
+            sparse_as_dense=False
+        )
+
+    if params["amp"]:
+        loss_scale = tf.train.experimental.DynamicLossScale(
+            initial_loss_scale=(2 ** 12),
+            increment_period=2000,
+            multiplier=2.0
+        )
+        optimizer = tf.compat.v1.train.experimental.MixedPrecisionLossScaleOptimizer(optimizer, loss_scale=loss_scale)
+
+    return optimizer
+
+
+def compute_model_statistics(batch_size, is_training=True):
+    """Compute number of parameters and FLOPS."""
+    options = tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()
+    options['output'] = 'none'
+
+    from tensorflow.python.keras.backend import get_graph
+    flops = tf.compat.v1.profiler.profile(get_graph(), options=options).total_float_ops
+    flops_per_image = flops / batch_size
+
+    logging.info('[%s Compute Statistics] %.1f GFLOPS/image' % (
+        "Training" if is_training else "Inference",
+        flops_per_image/1e9
+    ))
+
+
+def build_model_graph(features, labels, is_training, params):
+    """Builds the forward model graph."""
+    model_outputs = {}
+    is_gpu_inference = not is_training and params['use_batched_nms']
+
+    batch_size, image_height, image_width, _ = features['images'].get_shape().as_list()
+
+    if 'source_ids' not in features:
+        features['source_ids'] = -1 * tf.ones([batch_size], dtype=tf.float32)
+
+    all_anchors = anchors.Anchors(params['min_level'], params['max_level'],
+                                  params['num_scales'], params['aspect_ratios'],
+                                  params['anchor_scale'],
+                                  (image_height, image_width))
+
+    MODELS["backbone"] = resnet.Resnet_Model(
+        "resnet50",
+        data_format='channels_last',
+        trainable=is_training,
+        finetune_bn=params['finetune_bn']
+    )
+
+    backbone_feats = MODELS["backbone"](
+        features['images'],
+        training=is_training,
+    )
+
+    MODELS["FPN"] = fpn.FPNNetwork(params['min_level'], params['max_level'], trainable=is_training)
+    fpn_feats = MODELS["FPN"](backbone_feats, training=is_training)
+
+    model_outputs.update({'fpn_features': fpn_feats})
+
+    def rpn_head_fn(features, min_level=2, max_level=6, num_anchors=3):
+        """Region Proposal Network (RPN) for Mask-RCNN."""
+        scores_outputs = dict()
+        box_outputs = dict()
+
+        MODELS["RPN_Heads"] = heads.RPN_Head_Model(name="rpn_head", num_anchors=num_anchors, trainable=is_training)
+
+        for level in range(min_level, max_level + 1):
+            scores_outputs[level], box_outputs[level] = MODELS["RPN_Heads"](features[level], training=is_training)
+
+        return scores_outputs, box_outputs
+
+    rpn_score_outputs, rpn_box_outputs = rpn_head_fn(
+        features=fpn_feats,
+        min_level=params['min_level'],
+        max_level=params['max_level'],
+        num_anchors=len(params['aspect_ratios'] * params['num_scales'])
+    )
+
+    if is_training:
+        rpn_pre_nms_topn = params['train_rpn_pre_nms_topn']
+        rpn_post_nms_topn = params['train_rpn_post_nms_topn']
+        rpn_nms_threshold = params['train_rpn_nms_threshold']
+
+    else:
+        rpn_pre_nms_topn = params['test_rpn_pre_nms_topn']
+        rpn_post_nms_topn = params['test_rpn_post_nms_topn']
+        rpn_nms_threshold = params['test_rpn_nms_thresh']
+
+    if params['use_custom_box_proposals_op']:
+        rpn_box_scores, rpn_box_rois = roi_ops.custom_multilevel_propose_rois(
+            scores_outputs=rpn_score_outputs,
+            box_outputs=rpn_box_outputs,
+            all_anchors=all_anchors,
+            image_info=features['image_info'],
+            rpn_pre_nms_topn=rpn_pre_nms_topn,
+            rpn_post_nms_topn=rpn_post_nms_topn,
+            rpn_nms_threshold=rpn_nms_threshold,
+            rpn_min_size=params['rpn_min_size']
+        )
+
+    else:
+        rpn_box_scores, rpn_box_rois = roi_ops.multilevel_propose_rois(
+            scores_outputs=rpn_score_outputs,
+            box_outputs=rpn_box_outputs,
+            all_anchors=all_anchors,
+            image_info=features['image_info'],
+            rpn_pre_nms_topn=rpn_pre_nms_topn,
+            rpn_post_nms_topn=rpn_post_nms_topn,
+            rpn_nms_threshold=rpn_nms_threshold,
+            rpn_min_size=params['rpn_min_size'],
+            bbox_reg_weights=None,
+            use_batched_nms=params['use_batched_nms']
+        )
+
+
+    rpn_box_rois = tf.cast(rpn_box_rois, dtype=tf.float32)
+
+    if is_training:
+        rpn_box_rois = tf.stop_gradient(rpn_box_rois)
+        rpn_box_scores = tf.stop_gradient(rpn_box_scores)  # TODO Jonathan: Unused => Shall keep ?
+
+        # Sampling
+        box_targets, class_targets, rpn_box_rois, proposal_to_label_map = training_ops.proposal_label_op(
+            rpn_box_rois,
+            labels['gt_boxes'],
+            labels['gt_classes'],
+            batch_size_per_im=params['batch_size_per_im'],
+            fg_fraction=params['fg_fraction'],
+            fg_thresh=params['fg_thresh'],
+            bg_thresh_hi=params['bg_thresh_hi'],
+            bg_thresh_lo=params['bg_thresh_lo']
+        )
+
+    # Performs multi-level RoIAlign.
+    box_roi_features = spatial_transform_ops.multilevel_crop_and_resize(
+        features=fpn_feats,
+        boxes=rpn_box_rois,
+        output_size=7,
+        is_gpu_inference=is_gpu_inference
+    )
+
+    MODELS["Box_Head"] = heads.Box_Head_Model(
+        num_classes=params['num_classes'],
+        mlp_head_dim=params['fast_rcnn_mlp_head_dim'],
+        trainable=is_training
+    )
+
+    class_outputs, box_outputs, _ = MODELS["Box_Head"](inputs=box_roi_features)
+
+    if not is_training:
+        if params['use_batched_nms']:
+            generate_detections_fn = postprocess_ops.generate_detections_gpu
+
+        else:
+            generate_detections_fn = postprocess_ops.generate_detections_tpu
+
+        detections = generate_detections_fn(
+            class_outputs=class_outputs,
+            box_outputs=box_outputs,
+            anchor_boxes=rpn_box_rois,
+            image_info=features['image_info'],
+            pre_nms_num_detections=params['test_rpn_post_nms_topn'],
+            post_nms_num_detections=params['test_detections_per_image'],
+            nms_threshold=params['test_nms'],
+            bbox_reg_weights=params['bbox_reg_weights']
+        )
+
+        model_outputs.update({
+            'num_detections': detections[0],
+            'detection_boxes': detections[1],
+            'detection_classes': detections[2],
+            'detection_scores': detections[3],
+        })
+
+    else:  # is training
+        encoded_box_targets = training_ops.encode_box_targets(
+            boxes=rpn_box_rois,
+            gt_boxes=box_targets,
+            gt_labels=class_targets,
+            bbox_reg_weights=params['bbox_reg_weights']
+        )
+
+        model_outputs.update({
+            'rpn_score_outputs': rpn_score_outputs,
+            'rpn_box_outputs': rpn_box_outputs,
+            'class_outputs': class_outputs,
+            'box_outputs': box_outputs,
+            'class_targets': class_targets,
+            'box_targets': encoded_box_targets,
+            'box_rois': rpn_box_rois,
+        })
+
+    # Faster-RCNN mode.
+    if not params['include_mask']:
+        return model_outputs
+
+    # Mask sampling
+    if not is_training:
+        selected_box_rois = model_outputs['detection_boxes']
+        class_indices = model_outputs['detection_classes']
+
+        # If using GPU for inference, delay the cast until when Gather ops show up
+        # since GPU inference supports float point better.
+        # TODO(laigd): revisit this when newer versions of GPU libraries is
+        # released.
+        if not params['use_batched_nms']:
+            class_indices = tf.cast(class_indices, dtype=tf.int32)
+
+    else:
+        selected_class_targets, selected_box_targets, \
+        selected_box_rois, proposal_to_label_map = training_ops.select_fg_for_masks(
+            class_targets=class_targets,
+            box_targets=box_targets,
+            boxes=rpn_box_rois,
+            proposal_to_label_map=proposal_to_label_map,
+            max_num_fg=int(params['batch_size_per_im'] * params['fg_fraction'])
+        )
+
+        class_indices = tf.cast(selected_class_targets, dtype=tf.int32)
+
+    mask_roi_features = spatial_transform_ops.multilevel_crop_and_resize(
+        features=fpn_feats,
+        boxes=selected_box_rois,
+        output_size=14,
+        is_gpu_inference=is_gpu_inference
+    )
+
+    MODELS["Mask_Head"] = heads.Mask_Head_Model(
+        class_indices,
+        num_classes=params['num_classes'],
+        mrcnn_resolution=params['mrcnn_resolution'],
+        is_gpu_inference=is_gpu_inference,
+        trainable=is_training,
+        name="mask_head"
+    )
+
+    mask_outputs = MODELS["Mask_Head"](inputs=mask_roi_features)
+
+    if MPI_local_rank() == 0:
+        # Print #FLOPs in model.
+        compute_model_statistics(batch_size, is_training=is_training)
+
+    if is_training:
+        mask_targets = training_ops.get_mask_targets(
+
+            fg_boxes=selected_box_rois,
+            fg_proposal_to_label_map=proposal_to_label_map,
+            fg_box_targets=selected_box_targets,
+            mask_gt_labels=labels['cropped_gt_masks'],
+            output_size=params['mrcnn_resolution']
+        )
+
+        model_outputs.update({
+            'mask_outputs': mask_outputs,
+            'mask_targets': mask_targets,
+            'selected_class_targets': selected_class_targets,
+        })
+
+    else:
+        model_outputs.update({
+            'detection_masks': tf.nn.sigmoid(mask_outputs),
+        })
+
+    return model_outputs
+
+
+def _model_fn(features, labels, mode, params):
+    """Model defination for the Mask-RCNN model based on ResNet.
+
+    Args:
+    features: the input image tensor and auxiliary information, such as
+      `image_info` and `source_ids`. The image tensor has a shape of
+      [batch_size, height, width, 3]. The height and width are fixed and equal.
+    labels: the input labels in a dictionary. The labels include score targets
+      and box targets which are dense label maps. The labels are generated from
+      get_input_fn function in data/dataloader.py
+    mode: the mode of TPUEstimator including TRAIN, EVAL, and PREDICT.
+    params: the dictionary defines hyperparameters of model. The default
+      settings are in default_hparams function in this file.
+    Returns:
+    tpu_spec: the TPUEstimatorSpec to run training, evaluation, or prediction.
+    """
+
+    # Set up training loss and learning rate.
+    global_step = tf.compat.v1.train.get_or_create_global_step()
+
+    if mode == tf.estimator.ModeKeys.PREDICT:
+
+        if params['include_groundtruth_in_features'] and 'labels' in features:
+            # In include groundtruth for eval.
+            labels = features['labels']
+
+        else:
+            labels = None
+
+        if 'features' in features:
+            features = features['features']
+            # Otherwise, it is in export mode, the features is past in directly.
+
+    model_outputs = build_model_graph(features, labels, mode == tf.estimator.ModeKeys.TRAIN, params)
+
+    model_outputs.update({
+        'source_id': features['source_ids'],
+        'image_info': features['image_info'],
+    })
+
+    if mode == tf.estimator.ModeKeys.PREDICT and 'orig_images' in features:
+        model_outputs['orig_images'] = features['orig_images']
+
+    # First check if it is in PREDICT mode or EVAL mode to fill out predictions.
+    # Predictions are used during the eval step to generate metrics.
+    if mode in [tf.estimator.ModeKeys.PREDICT, tf.estimator.ModeKeys.EVAL]:
+        predictions = {}
+
+        try:
+            model_outputs['orig_images'] = features['orig_images']
+        except KeyError:
+            pass
+
+        if labels and params['include_groundtruth_in_features']:
+            # Labels can only be embedded in predictions. The prediction cannot output
+            # dictionary as a value.
+            predictions.update(labels)
+
+        model_outputs.pop('fpn_features', None)
+        predictions.update(model_outputs)
+
+        if mode == tf.estimator.ModeKeys.PREDICT:
+            # If we are doing PREDICT, we can return here.
+            return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
+
+    # score_loss and box_loss are for logging. only total_loss is optimized.
+    total_rpn_loss, rpn_score_loss, rpn_box_loss = losses.rpn_loss(
+        score_outputs=model_outputs['rpn_score_outputs'],
+        box_outputs=model_outputs['rpn_box_outputs'],
+        labels=labels,
+        params=params
+    )
+
+    total_fast_rcnn_loss, fast_rcnn_class_loss, fast_rcnn_box_loss = losses.fast_rcnn_loss(
+        class_outputs=model_outputs['class_outputs'],
+        box_outputs=model_outputs['box_outputs'],
+        class_targets=model_outputs['class_targets'],
+        box_targets=model_outputs['box_targets'],
+        params=params
+    )
+
+    # Only training has the mask loss.
+    # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/modeling/model_builder.py
+    if mode == tf.estimator.ModeKeys.TRAIN and params['include_mask']:
+        mask_loss = losses.mask_rcnn_loss(
+            mask_outputs=model_outputs['mask_outputs'],
+            mask_targets=model_outputs['mask_targets'],
+            select_class_targets=model_outputs['selected_class_targets'],
+            params=params
+        )
+
+    else:
+        mask_loss = 0.
+
+    trainable_variables = list(itertools.chain.from_iterable([model.trainable_variables for model in MODELS.values()]))
+
+    l2_regularization_loss = params['l2_weight_decay'] * tf.add_n([
+        tf.nn.l2_loss(v)
+        for v in trainable_variables
+        if not any([pattern in v.name for pattern in ["batch_normalization", "bias", "beta"]])
+    ])
+
+    total_loss = total_rpn_loss + total_fast_rcnn_loss + mask_loss + l2_regularization_loss
+
+    if mode == tf.estimator.ModeKeys.EVAL:
+        # Predictions can only contain a dict of tensors, not a dict of dict of
+        # tensors. These outputs are not used for eval purposes.
+        del predictions['rpn_score_outputs']
+        del predictions['rpn_box_outputs']
+
+        return tf.estimator.EstimatorSpec(
+            mode=mode,
+            predictions=predictions,
+            loss=total_loss
+        )
+
+    if mode == tf.estimator.ModeKeys.TRAIN:
+
+        learning_rate = learning_rates.step_learning_rate_with_linear_warmup(
+            global_step=global_step,
+            init_learning_rate=params['init_learning_rate'],
+            warmup_learning_rate=params['warmup_learning_rate'],
+            warmup_steps=params['warmup_steps'],
+            learning_rate_levels=params['learning_rate_levels'],
+            learning_rate_steps=params['learning_rate_steps']
+        )
+
+        optimizer = create_optimizer(learning_rate, params)
+
+        grads_and_vars = optimizer.compute_gradients(total_loss, trainable_variables, colocate_gradients_with_ops=True)
+
+        gradients, variables = zip(*grads_and_vars)
+        grads_and_vars = []
+
+        # Special treatment for biases (beta is named as bias in reference model)
+        # Reference: https://github.com/ddkang/Detectron/blob/80f3295308/lib/modeling/optimizer.py#L109
+        for grad, var in zip(gradients, variables):
+
+            if grad is not None and any([pattern in var.name for pattern in ["bias", "beta"]]):
+                grad = 2.0 * grad
+
+            grads_and_vars.append((grad, var))
+
+        train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)
+
+    else:
+        train_op = None
+        learning_rate = None
+
+    replica_id = tf.distribute.get_replica_context().replica_id_in_sync_group
+
+    if not isinstance(replica_id, tf.Tensor) or tf.get_static_value(replica_id) == 0:
+
+        register_metric(name="L2 loss", tensor=l2_regularization_loss, aggregator=StandardMeter())
+        register_metric(name="Mask loss", tensor=mask_loss, aggregator=StandardMeter())
+        register_metric(name="Total loss", tensor=total_loss, aggregator=StandardMeter())
+
+        register_metric(name="RPN box loss", tensor=rpn_box_loss, aggregator=StandardMeter())
+        register_metric(name="RPN score loss", tensor=rpn_score_loss, aggregator=StandardMeter())
+        register_metric(name="RPN total loss", tensor=total_rpn_loss, aggregator=StandardMeter())
+
+        register_metric(name="FastRCNN class loss", tensor=fast_rcnn_class_loss, aggregator=StandardMeter())
+        register_metric(name="FastRCNN box loss", tensor=fast_rcnn_box_loss, aggregator=StandardMeter())
+        register_metric(name="FastRCNN total loss", tensor=total_fast_rcnn_loss, aggregator=StandardMeter())
+
+        register_metric(name="Learning rate", tensor=learning_rate, aggregator=StandardMeter())
+        pass
+    return tf.estimator.EstimatorSpec(
+        mode=mode,
+        loss=total_loss,
+        train_op=train_op,
+    )
+
+
+def mask_rcnn_model_fn(features, labels, mode, params):
+    """Mask-RCNN model."""
+
+    return _model_fn(
+        features,
+        labels,
+        mode,
+        params
+    )

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/models/fpn.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Feature Pyramid Network.
+
+Feature Pyramid Networks were proposed in:
+[1] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan,
+    , and Serge Belongie
+    Feature Pyramid Networks for Object Detection. CVPR 2017.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+from mask_rcnn.ops import spatial_transform_ops
+
+
+class FPNNetwork(tf.keras.models.Model):
+    def __init__(self, min_level=3, max_level=7, filters=256, trainable=True):
+        """Generates multiple scale feature pyramid (FPN).
+
+        Args:
+        feats_bottom_up: a dictionary of tensor with level as keys and bottom up
+          feature tensors as values. They are the features to generate FPN features.
+        min_level: the minimum level number to generate FPN features.
+        max_level: the maximum level number to generate FPN features.
+        filters: the FPN filter size.
+
+        Returns:
+        feats: a dictionary of tensor with level as keys and the generated FPN
+          features as values.
+        """
+        super(FPNNetwork, self).__init__(name="fpn", trainable=trainable)
+
+        self._local_layers = dict()
+
+        self._min_level = min_level
+        self._max_level = max_level
+
+        self._filters = filters
+
+        self._backbone_max_level = 5  # max(feats_bottom_up.keys())
+        self._upsample_max_level = (
+            self._backbone_max_level if self._max_level > self._backbone_max_level else self._max_level
+        )
+
+        self._local_layers["stage1"] = dict()
+        for level in range(self._min_level, self._upsample_max_level + 1):
+            self._local_layers["stage1"][level] = tf.keras.layers.Conv2D(
+                filters=self._filters,
+                kernel_size=(1, 1),
+                padding='same',
+                name='l%d' % level,
+                trainable=trainable
+            )
+
+        self._local_layers["stage2"] = dict()
+        # add post-hoc 3x3 convolution kernel
+        for level in range(self._min_level, self._upsample_max_level + 1):
+            self._local_layers["stage2"][level] = tf.keras.layers.Conv2D(
+                filters=self._filters,
+                strides=(1, 1),
+                kernel_size=(3, 3),
+                padding='same',
+                name='post_hoc_d%d' % level,
+                trainable=trainable
+            )
+
+        self._local_layers["stage3_1"] = dict()
+        self._local_layers["stage3_2"] = dict()
+
+        if self._max_level == self._upsample_max_level + 1:
+            self._local_layers["stage3_1"] = tf.keras.layers.MaxPool2D(
+                pool_size=1,
+                strides=2,
+                padding='valid',
+                name='p%d' % self._max_level,
+                trainable=trainable
+            )
+
+        else:
+            for level in range(self._upsample_max_level + 1, self._max_level + 1):
+                self._local_layers["stage3_2"][level] = tf.keras.layers.Conv2D(
+                    filters=self._filters,
+                    strides=(2, 2),
+                    kernel_size=(3, 3),
+                    padding='same',
+                    name='p%d' % level,
+                    trainable=trainable
+                )
+
+    def call(self, inputs, *args, **kwargs):
+
+        feats_bottom_up = inputs
+
+        # lateral connections
+        feats_lateral = {}
+
+        for level in range(self._min_level, self._upsample_max_level + 1):
+            feats_lateral[level] = self._local_layers["stage1"][level](feats_bottom_up[level])
+
+        # add top-down path
+        feats = {self._upsample_max_level: feats_lateral[self._upsample_max_level]}
+
+        for level in range(self._upsample_max_level - 1, self._min_level - 1, -1):
+            feats[level] = spatial_transform_ops.nearest_upsampling(
+                feats[level + 1], 2
+            ) + feats_lateral[level]
+
+        # add post-hoc 3x3 convolution kernel
+        for level in range(self._min_level, self._upsample_max_level + 1):
+            feats[level] = self._local_layers["stage2"][level](feats[level])
+
+        if self._max_level == self._upsample_max_level + 1:
+            feats[self._max_level] = self._local_layers["stage3_1"](feats[self._max_level - 1])
+
+        else:
+            for level in range(self._upsample_max_level + 1, self._max_level + 1):
+                feats[level] = self._local_layers["stage3_2"][level](feats[level - 1])
+
+        return feats

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/models/heads.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Functions to build various prediction heads in Mask-RCNN."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+__all__ = ["RPN_Head_Model", "Box_Head_Model", "Mask_Head_Model"]
+
+
+class RPN_Head_Model(tf.keras.models.Model):
+
+        def __init__(self, name, num_anchors, trainable, *args, **kwargs):
+            super(RPN_Head_Model, self).__init__(name=name, trainable=trainable, *args, **kwargs)
+            """Shared RPN heads."""
+            self._local_layers = dict()
+
+            # TODO(chiachenc): check the channel depth of the first convolution.
+            self._local_layers["conv1"] = tf.keras.layers.Conv2D(
+                256,
+                kernel_size=(3, 3),
+                strides=(1, 1),
+                activation=tf.nn.relu,
+                bias_initializer=tf.keras.initializers.Zeros(),
+                kernel_initializer=tf.random_normal_initializer(stddev=0.01),
+                padding='same',
+                trainable=trainable,
+                name='rpn'
+            )
+
+            # Proposal classification scores
+            # scores = tf.keras.layers.Conv2D(
+            self._local_layers["conv2"] = tf.keras.layers.Conv2D(
+                num_anchors,
+                kernel_size=(1, 1),
+                strides=(1, 1),
+                bias_initializer=tf.keras.initializers.Zeros(),
+                kernel_initializer=tf.random_normal_initializer(stddev=0.01),
+                padding='valid',
+                trainable=trainable,
+                name='rpn-class'
+            )
+
+            # Proposal bbox regression deltas
+            # bboxes = tf.keras.layers.Conv2D(
+            self._local_layers["conv3"] = tf.keras.layers.Conv2D(
+                4 * num_anchors,
+                kernel_size=(1, 1),
+                strides=(1, 1),
+                bias_initializer=tf.keras.initializers.Zeros(),
+                kernel_initializer=tf.random_normal_initializer(stddev=0.01),
+                padding='valid',
+                trainable=trainable,
+                name='rpn-box'
+            )
+
+        def call(self, inputs, *args, **kwargs):
+            net = self._local_layers["conv1"](inputs)
+            scores = self._local_layers["conv2"](net)
+            bboxes = self._local_layers["conv3"](net)
+
+            return scores, bboxes
+
+
+class Box_Head_Model(tf.keras.Model):
+
+    def __init__(self, num_classes=91, mlp_head_dim=1024, name="box_head", trainable=True, *args, **kwargs):
+        """Box and class branches for the Mask-RCNN model.
+
+        Args:
+        roi_features: A ROI feature tensor of shape
+          [batch_size, num_rois, height_l, width_l, num_filters].
+        num_classes: a integer for the number of classes.
+        mlp_head_dim: a integer that is the hidden dimension in the fully-connected
+          layers.
+        """
+        super(Box_Head_Model, self).__init__(name=name, trainable=trainable, *args, **kwargs)
+
+        self._num_classes = num_classes
+        self._mlp_head_dim = mlp_head_dim
+
+        self._dense_fc6 = tf.keras.layers.Dense(
+            units=mlp_head_dim,
+            activation=tf.nn.relu,
+            trainable=trainable,
+            name='fc6'
+        )
+
+        self._dense_fc7 = tf.keras.layers.Dense(
+            units=mlp_head_dim,
+            activation=tf.nn.relu,
+            trainable=trainable,
+            name='fc7'
+        )
+
+        self._dense_class = tf.keras.layers.Dense(
+            num_classes,
+            kernel_initializer=tf.random_normal_initializer(stddev=0.01),
+            bias_initializer=tf.keras.initializers.Zeros(),
+            trainable=trainable,
+            name='class-predict'
+        )
+
+        self._dense_box = tf.keras.layers.Dense(
+            num_classes * 4,
+            kernel_initializer=tf.random_normal_initializer(stddev=0.001),
+            bias_initializer=tf.keras.initializers.Zeros(),
+            trainable=trainable,
+            name='box-predict'
+        )
+
+    def call(self, inputs, **kwargs):
+        """
+        Returns:
+        class_outputs: a tensor with a shape of
+          [batch_size, num_rois, num_classes], representing the class predictions.
+        box_outputs: a tensor with a shape of
+          [batch_size, num_rois, num_classes * 4], representing the box predictions.
+        box_features: a tensor with a shape of
+          [batch_size, num_rois, mlp_head_dim], representing the box features.
+        """
+
+        # reshape inputs before FC.
+        batch_size, num_rois, height, width, filters = inputs.get_shape().as_list()
+
+        net = tf.reshape(inputs, [batch_size, num_rois, height * width * filters])
+
+        net = self._dense_fc6(net)
+
+        box_features = self._dense_fc7(net)
+
+        class_outputs = self._dense_class(box_features)
+
+        box_outputs = self._dense_box(box_features)
+
+        return class_outputs, box_outputs, box_features
+
+
+class Mask_Head_Model(tf.keras.Model):
+
+    @staticmethod
+    def _get_stddev_equivalent_to_msra_fill(kernel_size, fan_out):
+        """Returns the stddev of random normal initialization as MSRAFill."""
+        # Reference: https://github.com/pytorch/pytorch/blob/master/caffe2/operators/filler_op.h#L445-L463
+        # For example, kernel size is (3, 3) and fan out is 256, stddev is 0.029.
+        # stddev = (2/(3*3*256))^0.5 = 0.029
+        return (2 / (kernel_size[0] * kernel_size[1] * fan_out)) ** 0.5
+
+    def __init__(
+            self,
+            class_indices,
+            num_classes=91,
+            mrcnn_resolution=28,
+            is_gpu_inference=False,
+            name="mask_head",
+            trainable=True,
+            *args,
+            **kwargs
+    ):
+        """Mask branch for the Mask-RCNN model.
+
+        Args:
+        roi_features: A ROI feature tensor of shape
+          [batch_size, num_rois, height_l, width_l, num_filters].
+        class_indices: a Tensor of shape [batch_size, num_rois], indicating
+          which class the ROI is.
+        num_classes: an integer for the number of classes.
+        mrcnn_resolution: an integer that is the resolution of masks.
+        is_gpu_inference: whether to build the model for GPU inference.
+        """
+        super(Mask_Head_Model, self).__init__(name=name, trainable=trainable, *args, **kwargs)
+
+        self._class_indices = class_indices
+        self._num_classes = num_classes
+        self._mrcnn_resolution = mrcnn_resolution
+        self._is_gpu_inference = is_gpu_inference
+
+        self._conv_stage1 = list()
+        kernel_size = (3, 3)
+        fan_out = 256
+
+        init_stddev = Mask_Head_Model._get_stddev_equivalent_to_msra_fill(kernel_size, fan_out)
+
+        for conv_id in range(4):
+            self._conv_stage1.append(tf.keras.layers.Conv2D(
+                fan_out,
+                kernel_size=kernel_size,
+                strides=(1, 1),
+                padding='same',
+                dilation_rate=(1, 1),
+                activation=tf.nn.relu,
+                kernel_initializer=tf.random_normal_initializer(stddev=init_stddev),
+                bias_initializer=tf.keras.initializers.Zeros(),
+                trainable=trainable,
+                name='mask-conv-l%d' % conv_id
+            ))
+
+        kernel_size = (2, 2)
+        fan_out = 256
+
+        init_stddev = Mask_Head_Model._get_stddev_equivalent_to_msra_fill(kernel_size, fan_out)
+
+        self._conv_stage2 = tf.keras.layers.Conv2DTranspose(
+            fan_out,
+            kernel_size=kernel_size,
+            strides=(2, 2),
+            padding='valid',
+            activation=tf.nn.relu,
+            kernel_initializer=tf.random_normal_initializer(stddev=init_stddev),
+            bias_initializer=tf.keras.initializers.Zeros(),
+            trainable=trainable,
+            name='conv5-mask'
+        )
+
+        kernel_size = (1, 1)
+        fan_out = self._num_classes
+
+        init_stddev = Mask_Head_Model._get_stddev_equivalent_to_msra_fill(kernel_size, fan_out)
+
+        self._conv_stage3 = tf.keras.layers.Conv2D(
+            fan_out,
+            kernel_size=kernel_size,
+            strides=(1, 1),
+            padding='valid',
+            kernel_initializer=tf.random_normal_initializer(stddev=init_stddev),
+            bias_initializer=tf.keras.initializers.Zeros(),
+            trainable=trainable,
+            name='mask_fcn_logits'
+        )
+
+    def call(self, inputs, **kwargs):
+        """
+        Returns:
+        mask_outputs: a tensor with a shape of
+          [batch_size, num_masks, mask_height, mask_width],
+          representing the mask predictions.
+        fg_gather_indices: a tensor with a shape of [batch_size, num_masks, 2],
+          representing the fg mask targets.
+        Raises:
+        ValueError: If boxes is not a rank-3 tensor or the last dimension of
+          boxes is not 4.
+        """
+
+        batch_size, num_rois, height, width, filters = inputs.get_shape().as_list()
+
+        net = tf.reshape(inputs, [-1, height, width, filters])
+
+        for conv_id in range(4):
+            net = self._conv_stage1[conv_id](net)
+
+        net = self._conv_stage2(net)
+
+        mask_outputs = self._conv_stage3(net)
+
+        mask_outputs = tf.reshape(
+            mask_outputs,
+            [-1, num_rois, self._mrcnn_resolution, self._mrcnn_resolution, self._num_classes]
+        )
+
+        with tf.name_scope('masks_post_processing'):
+
+            mask_outputs = tf.transpose(a=mask_outputs, perm=[0, 1, 4, 2, 3])
+
+            indices_dtype = tf.float32 if self._is_gpu_inference else tf.int32
+
+            if batch_size == 1:
+                indices = tf.reshape(
+                    tf.reshape(
+                        tf.range(num_rois, dtype=indices_dtype),
+                        [batch_size, num_rois, 1]
+                    ) * self._num_classes + tf.expand_dims(self._class_indices, axis=-1),
+                    [batch_size, -1]
+                )
+                indices = tf.cast(indices, tf.int32)
+
+                mask_outputs = tf.gather(
+                    tf.reshape(mask_outputs, [batch_size, -1, self._mrcnn_resolution, self._mrcnn_resolution]),
+                    indices,
+                    axis=1
+                )
+
+                mask_outputs = tf.squeeze(mask_outputs, axis=1)
+                mask_outputs = tf.reshape(
+                    mask_outputs,
+                    [batch_size, num_rois, self._mrcnn_resolution, self._mrcnn_resolution])
+
+            else:
+                batch_indices = (
+                        tf.expand_dims(tf.range(batch_size, dtype=indices_dtype), axis=1) *
+                        tf.ones([1, num_rois], dtype=indices_dtype)
+                )
+
+                mask_indices = (
+                        tf.expand_dims(tf.range(num_rois, dtype=indices_dtype), axis=0) *
+                        tf.ones([batch_size, 1], dtype=indices_dtype)
+                )
+
+                gather_indices = tf.stack([batch_indices, mask_indices, self._class_indices], axis=2)
+
+                if self._is_gpu_inference:
+                    gather_indices = tf.cast(gather_indices, dtype=tf.int32)
+
+                mask_outputs = tf.gather_nd(mask_outputs, gather_indices)
+
+        return mask_outputs

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/models/keras_utils.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import itertools
+
+import tensorflow as tf
+
+
+__all__ = ["KerasMockLayer"]
+
+
+class KerasMockLayer(tf.Module):
+    """
+    This class reproduces the Keras Layer important APIs without enforcing a variable scope.
+    """
+    def __init__(self, trainable=True, *args, **kwargs):
+        super(KerasMockLayer, self).__init__(*args, **kwargs)
+        self._local_layers = dict()
+        self._trainable = trainable
+
+    @property
+    def trainable(self):
+        return self._trainable
+
+    @trainable.setter
+    def trainable(self, value):
+        self._trainable = value
+        for layer in getattr(self, '_layers', []):
+            layer.trainable = value
+
+    @property
+    def variables(self):
+        """Returns the list of all layer variables/weights.
+        Alias of `self.weights`.
+        Returns:
+          A list of variables.
+        """
+        return self.weights
+
+    @property
+    def trainable_variables(self):
+        return self.trainable_weights
+
+    @property
+    def non_trainable_variables(self):
+        return self.non_trainable_weights
+
+    @property
+    def weights(self):
+        """Returns the list of all layer variables/weights.
+        Returns:
+          A list of variables.
+        """
+        return self.trainable_weights + self.non_trainable_weights
+
+    @property
+    def name(self):
+        return self._name
+
+    @property
+    def trainable_weights(self):
+        layers = list()
+
+        for layer in self._local_layers.values():
+            if not isinstance(layer, dict):
+                layers.append(layer)
+            else:
+                for sublayer in layer.values():
+                    layers.append(sublayer)
+
+        return list(itertools.chain.from_iterable([layer.trainable_variables for layer in layers]))
+
+    @property
+    def non_trainable_weights(self):
+        layers = list()
+
+        for layer in self._local_layers.values():
+            if not isinstance(layer, dict):
+                layers.append(layer)
+            else:
+                for sublayer in layer.values():
+                    layers.append(sublayer)
+
+        return list(itertools.chain.from_iterable([layer.non_trainable_weights for layer in layers]))

TensorFlow2x/ComputeVision/Detection/MaskRCNN/mask_rcnn/models/resnet.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Resnet."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+from tensorflow.python.ops import variable_scope
+from tensorflow.python.keras import backend
+
+from mask_rcnn.models.keras_utils import KerasMockLayer
+
+_BATCH_NORM_DECAY = 0.997
+_BATCH_NORM_EPSILON = 1e-4
+
+
+class BNReLULayer(KerasMockLayer):
+    def __init__(self, trainable, relu=True, init_zero=False, data_format='channels_last'):
+        """Performs a batch normalization followed by a ReLU.
+
+        Args:
+        inputs: `Tensor` of shape `[batch, channels, ...]`.
+        trainable: `bool` for whether to finetune the batchnorm layer.
+        relu: `bool` if False, omits the ReLU operation.
+        init_zero: `bool` if True, initializes scale parameter of batch
+            normalization with 0 instead of 1 (default).
+        data_format: `str` either "channels_first" for `[batch, channels, height,
+            width]` or "channels_last for `[batch, height, width, channels]`.
+        name: the name of the batch normalization layer
+
+        Returns:
+        A normalized `Tensor` with the same `data_format`.
+        """
+        super(BNReLULayer, self).__init__(trainable=trainable)
+
+        if init_zero:
+            gamma_initializer = tf.keras.initializers.Zeros()
+        else:
+            gamma_initializer = tf.keras.initializers.Ones()
+
+        if data_format == 'channels_first':
+            axis = 1
+        else:
+            axis = 3
+
+        self._local_layers = dict()
+        self._local_layers["batchnorm"] = tf.keras.layers.BatchNormalization(
+            axis=axis,
+            momentum=_BATCH_NORM_DECAY,
+            epsilon=_BATCH_NORM_EPSILON,
+            center=True,
+            scale=True,
+            trainable=self._trainable,
+            gamma_initializer=gamma_initializer,
+            fused=True,
+            name="batch_normalization"
+        )
+
+        if relu:
+            self._local_layers["relu"] = tf.keras.layers.ReLU()
+
+    def __call__(self, inputs, training=False, *args, **kwargs):
+
+        net = self._local_layers["batchnorm"](inputs, training=training and self._trainable)
+
+        try:
+            return self._local_layers["relu"](net)
+        except KeyError:
+            return net
+
+
+class FixedPaddingLayer(KerasMockLayer):
+    def __init__(self, kernel_size, data_format='channels_last', trainable=True):
+        """Pads the input along the spatial dimensions independently of input size.
+
+        Args:
+        kernel_size: `int` kernel size to be used for `conv2d` or max_pool2d`
+            operations. Should be a positive integer.
+        data_format: `str` either "channels_first" for `[batch, channels, height,
+            width]` or "channels_last for `[batch, height, width, channels]`.
+        """
+        super(FixedPaddingLayer, self).__init__(trainable=trainable)
+
+        pad_total = kernel_size - 1
+        pad_beg = pad_total // 2
+        pad_end = pad_total - pad_beg
+
+        if data_format == 'channels_first':
+            self._paddings = [[0, 0], [0, 0], [pad_beg, pad_end], [pad_beg, pad_end]]
+        else:
+            self._paddings = [[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]]
+
+    def __call__(self, inputs, *args, **kwargs):
+        """
+      Args:
+        inputs: `Tensor` of size `[batch, channels, height, width]` or
+            `[batch, height, width, channels]` depending on `data_format`.
+      Returns:
+        A padded `Tensor` of the same `data_format` with size either intact
+        (if `kernel_size == 1`) or padded (if `kernel_size > 1`).
+        :param **kwargs:
+      """
+
+        return tf.pad(tensor=inputs, paddings=self._paddings)
+
+
+class Conv2dFixedPadding(KerasMockLayer):
+    def __init__(self, filters, kernel_size, strides, data_format='channels_last', trainable=False):
+        """Strided 2-D convolution with explicit padding.
+
+        The padding is consistent and is based only on `kernel_size`, not on the
+        dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
+
+        Args:
+        inputs: `Tensor` of size `[batch, channels, height_in, width_in]`.
+        filters: `int` number of filters in the convolution.
+        kernel_size: `int` size of the kernel to be used in the convolution.
+        strides: `int` strides of the convolution.
+        data_format: `str` either "channels_first" for `[batch, channels, height,
+            width]` or "channels_last for `[batch, height, width, channels]`.
+
+        Returns:
+            A `Tensor` of shape `[batch, filters, height_out, width_out]`.
+        """
+        super(Conv2dFixedPadding, self).__init__(trainable=trainable)
+
+        if strides > 1:
+            self._local_layers["fixed_padding"] = FixedPaddingLayer(kernel_size=kernel_size, data_format=data_format)
+
+        self._local_layers["conv2d"] = tf.keras.layers.Conv2D(
+            filters=filters,
+            kernel_size=kernel_size,
+            strides=strides,
+            padding=('SAME' if strides == 1 else 'VALID'),
+            use_bias=False,
+            kernel_initializer=tf.keras.initializers.VarianceScaling(),
+            data_format=data_format,
+            trainable=self._trainable,
+            name="conv2d"
+        )
+
+    def __call__(self, inputs, *args, **kwargs):
+
+        try:
+            net = self._local_layers["fixed_padding"](inputs)
+        except KeyError:
+            net = inputs
+
+        return self._local_layers["conv2d"](net)
+
+
+class ResidualBlock(KerasMockLayer):
+    def __init__(self, filters, trainable, finetune_bn, strides, use_projection=False, data_format='channels_last'):
+        """Standard building block for residual networks with BN after convolutions.
+
+        Args:
+        filters: `int` number of filters for the first two convolutions. Note that
+            the third and final convolution will use 4 times as many filters.
+        finetune_bn: `bool` for whether the model is in training.
+        strides: `int` block stride. If greater than 1, this block will ultimately downsample the input.
+        use_projection: `bool` for whether this block should use a projection
+            shortcut (versus the default identity shortcut). This is usually `True`
+            for the first block of a block group, which may change the number of
+            filters and the resolution.
+        data_format: `str` either "channels_first" for `[batch, channels, height, width]`
+            or "channels_last for `[batch, height, width, channels]`.
+        """
+        super(ResidualBlock, self).__init__(trainable=trainable)
+
+        self._finetune_bn = finetune_bn
+
+        if use_projection:
+            self._local_layers["projection"] = dict()
+
+            self._local_layers["projection"]["conv2d"] = Conv2dFixedPadding(
+                filters=filters,
+                kernel_size=1,
+                strides=strides,
+                data_format=data_format,
+                trainable=trainable
+            )
+
+            self._local_layers["projection"]["batchnorm"] = BNReLULayer(
+                trainable=finetune_bn and trainable,
+                relu=False,
+                init_zero=False,
+                data_format=data_format,
+            )
+
+        self._local_layers["conv2d_1"] = Conv2dFixedPadding(
+            trainable=trainable,
+            filters=filters,
+            kernel_size=3,
+            strides=strides,
+            data_format=data_format,
+        )
+
+        self._local_layers["conv2d_2"] = Conv2dFixedPadding(
+            trainable=trainable,
+            filters=filters,
+            kernel_size=3,
+            strides=1,
+            data_format=data_format,
+        )
+
+        self._local_layers["batchnorm_1"] = BNReLULayer(
+            trainable=finetune_bn and trainable,
+            relu=True,
+            init_zero=False,
+            data_format=data_format,
+        )
+
+        self._local_layers["batchnorm_2"] = BNReLULayer(
+            trainable=finetune_bn and trainable,
+            relu=False,
+            init_zero=True,
+            data_format=data_format,
+        )
+
+        self._local_layers["activation"] = tf.keras.layers.ReLU()
+
+    def __call__(self, inputs, training=False):
+        """
+        Args:
+        inputs: `Tensor` of size `[batch, channels, height, width]`.
+
+        Returns:
+        The output `Tensor` of the block.
+        """
+
+        try:
+            # Projection shortcut in first layer to match filters and strides
+            shortcut = self._local_layers["projection"]["conv2d"](inputs=inputs)
+
+            shortcut = self._local_layers["projection"]["batchnorm"](
+                inputs=shortcut,
+                training=training and self._trainable and self._finetune_bn
+            )
+
+        except KeyError:
+            shortcut = inputs
+
+        net = inputs
+
+        for i in range(1, 3):
+            net = self._local_layers["conv2d_%d" % i](inputs=net)
+
+            net = self._local_layers["batchnorm_%d" % i](
+                inputs=net,
+                training=training and self._trainable and self._finetune_bn
+            )
+
+        return self._local_layers["activation"](net + shortcut)
+
+
+class BottleneckBlock(KerasMockLayer):
+    def __init__(self, filters, trainable, finetune_bn, strides, use_projection=False, data_format='channels_last'):
+        """Bottleneck block variant for residual networks with BN after convolutions.
+
+        Args:
+        filters: `int` number of filters for the first two convolutions. Note that
+            the third and final convolution will use 4 times as many filters.
+        finetune_bn: `bool` for whether the model is in training.
+        strides: `int` block stride. If greater than 1, this block will ultimately downsample the input.
+        use_projection: `bool` for whether this block should use a projection
+            shortcut (versus the default identity shortcut). This is usually `True`
+            for the first block of a block group, which may change the number of
+            filters and the resolution.
+        data_format: `str` either "channels_first" for `[batch, channels, height, width]`
+            or "channels_last for `[batch, height, width, channels]`.
+        """
+        super(BottleneckBlock, self).__init__(trainable=trainable)
+
+        self._finetune_bn = finetune_bn
+
+        if use_projection:
+            # Projection shortcut only in first block within a group. Bottleneck blocks
+            # end with 4 times the number of filters.
+            filters_out = 4 * filters
+
+            self._local_layers["projection"] = dict()
+
+            self._local_layers["projection"]["conv2d"] = Conv2dFixedPadding(
+                filters=filters_out,
+                kernel_size=1,
+                strides=strides,
+                data_format=data_format,
+                trainable=trainable
+            )
+
+            self._local_layers["projection"]["batchnorm"] = BNReLULayer(
+                trainable=finetune_bn and trainable,
+                relu=False,
+                init_zero=False,
+                data_format=data_format,
+            )
+
+        self._local_layers["conv2d_1"] = Conv2dFixedPadding(
+            filters=filters,
+            kernel_size=1,
+            strides=1,
+            data_format=data_format,
+            trainable=trainable
+        )
+
+        self._local_layers["conv2d_2"] = Conv2dFixedPadding(
+            filters=filters,
+            kernel_size=3,
+            strides=strides,
+            data_format=data_format,
+            trainable=trainable
+        )
+
+        self._local_layers["conv2d_3"] = Conv2dFixedPadding(
+            filters=4 * filters,
+            kernel_size=1,
+            strides=1,
+            data_format=data_format,
+            trainable=trainable
+        )
+
+        self._local_layers["batchnorm_1"] = BNReLULayer(
+            trainable=finetune_bn and trainable,
+            relu=True,
+            init_zero=False,
+            data_format=data_format,
+        )
+
+        self._local_layers["batchnorm_2"] = BNReLULayer(
+            trainable=finetune_bn and trainable,
+            relu=True,
+            init_zero=False,
+            data_format=data_format,
+        )
+
+        self._local_layers["batchnorm_3"] = BNReLULayer(
+            trainable=finetune_bn and trainable,
+            relu=False,
+            init_zero=True,
+            data_format=data_format,
+        )
+
+        self._local_layers["activation"] = tf.keras.layers.ReLU()
+
+    def __call__(self, inputs, training=False):
+        """
+        Args:
+        inputs: `Tensor` of size `[batch, channels, height, width]`.
+
+        Returns:
+        The output `Tensor` of the block.
+        """
+
+        try:
+            # Projection shortcut in first layer to match filters and strides
+            shortcut = self._local_layers["projection"]["conv2d"](inputs=inputs)
+
+            shortcut = self._local_layers["projection"]["batchnorm"](
+                inputs=shortcut,
+                training=training and self._trainable and self._finetune_bn
+            )
+
+        except KeyError:
+            shortcut = inputs
+
+        net = inputs
+
+        for i in range(1, 4):
+            net = self._local_layers["conv2d_%d" % i](inputs=net)
+
+            net = self._local_layers["batchnorm_%d" % i](
+                inputs=net,
+                training=training and self._trainable and self._finetune_bn
+            )
+
+        return self._local_layers["activation"](net + shortcut)
+
+
+class BlockGroup(KerasMockLayer):
+    def __init__(self, filters, block_layer, n_blocks, strides, trainable, finetune_bn, data_format='channels_last'):
+        """Creates one group of blocks for the ResNet model.
+
+        Args:
+        inputs: `Tensor` of size `[batch, channels, height, width]`.
+        filters: `int` number of filters for the first convolution of the layer.
+        block_layer: `layer` for the block to use within the model
+        n_blocks: `int` number of blocks contained in the layer.
+        strides: `int` stride to use for the first convolution of the layer. If
+            greater than 1, this layer will downsample the input.
+        finetune_bn: `bool` for whether the model is training.
+        name: `str`name for the Tensor output of the block layer.
+        data_format: `str` either "channels_first" for `[batch, channels, height,
+            width]` or "channels_last for `[batch, height, width, channels]`.
+
+        Returns:
+        The output `Tensor` of the block layer.
+        """
+        super(BlockGroup, self).__init__(trainable=trainable)
+
+        self._finetune_bn = finetune_bn
+
+        self._n_blocks = n_blocks
+
+        for block_id in range(self._n_blocks):
+            # Only the first block per block_group uses projection shortcut and strides.
+            self._local_layers["block_%d" % (block_id + 1)] = block_layer(
+                filters=filters,
+                finetune_bn=finetune_bn,
+                trainable=trainable,
+                strides=strides if block_id == 0 else 1,
+                use_projection=block_id == 0,
+                data_format=data_format
+            )
+
+    def __call__(self, inputs, training=False):
+
+        net = inputs
+
+        for block_id in range(self._n_blocks):
+            net = self._local_layers["block_%d" % (block_id + 1)](net, training=training and self._trainable)
+
+        return net
+
+
+class Resnet_Model(KerasMockLayer, tf.keras.models.Model):
+    def __init__(self, resnet_model, data_format='channels_last', trainable=True, finetune_bn=False, *args, **kwargs):
+        """
+        Our actual ResNet network.  We return the output of c2, c3,c4,c5
+        N.B. batch norm is always run with trained parameters, as we use very small
+        batches when training the object layers.
+
+        Args:
+        resnet_model: model type. Authorized Values: (resnet18, resnet34, resnet50, resnet101, resnet152, resnet200)
+        data_format: `str` either "channels_first" for
+          `[batch, channels, height, width]` or "channels_last for `[batch, height, width, channels]`.
+        finetune_bn: `bool` for whether the model is training.
+
+        Returns the ResNet model for a given size and number of output classes.
+        """
+        model_params = {
+            'resnet18': {'block': ResidualBlock, 'layers': [2, 2, 2, 2]},
+            'resnet34': {'block': ResidualBlock, 'layers': [3, 4, 6, 3]},
+            'resnet50': {'block': BottleneckBlock, 'layers': [3, 4, 6, 3]},
+            'resnet101': {'block': BottleneckBlock, 'layers': [3, 4, 23, 3]},
+            'resnet152': {'block': BottleneckBlock, 'layers': [3, 8, 36, 3]},
+            'resnet200': {'block': BottleneckBlock, 'layers': [3, 24, 36, 3]}
+        }
+
+        if resnet_model not in model_params:
+            raise ValueError('Not a valid resnet_model: %s' % resnet_model)
+
+        super(Resnet_Model, self).__init__(trainable=trainable, name=resnet_model, *args, **kwargs)
+
+        self._finetune_bn = finetune_bn
+
+        self._data_format = data_format
+        self._block_layer = model_params[resnet_model]['block']
+        self._n_layers = model_params[resnet_model]['layers']
+
+        self._local_layers["conv2d"] = Conv2dFixedPadding(
+            filters=64,
+            kernel_size=7,
+            strides=2,
+            data_format=self._data_format,
+            # Freeze at conv2d and batchnorm first 11 layers based on reference model.
+            # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py#L194
+            trainable=False
+        )
+
+        self._local_layers["batchnorm"] = BNReLULayer(
+            relu=True,
+            init_zero=False,
+            data_format=self._data_format,
+            # Freeze at conv2d and batchnorm first 11 layers based on reference model.
+            # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py#L194
+            trainable=False
+        )
+
+        self._local_layers["maxpool2d"] = tf.keras.layers.MaxPool2D(
+            pool_size=3,
+            strides=2,
+            padding='SAME',
+            data_format=self._data_format
+        )
+
+        self._local_layers["block_1"] = BlockGroup(
+            filters=64,
+            strides=1,
+            n_blocks=self._n_layers[0],
+            block_layer=self._block_layer,
+            data_format=self._data_format,
+            # Freeze at conv2d and batchnorm first 11 layers based on reference model.
+            # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py#L194
+            trainable=False,
+            finetune_bn=False
+        )
+
+        self._local_layers["block_2"] = BlockGroup(
+            filters=128,
+            strides=2,
+            n_blocks=self._n_layers[1],
+            block_layer=self._block_layer,
+            data_format=self._data_format,
+            # Freeze at conv2d and batchnorm first 11 layers based on reference model.
+            # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py#L194
+            trainable=self._trainable,
+            finetune_bn=self._finetune_bn
+        )
+
+        self._local_layers["block_3"] = BlockGroup(
+            filters=256,
+            strides=2,
+            n_blocks=self._n_layers[2],
+            block_layer=self._block_layer,
+            data_format=self._data_format,
+            # Freeze at conv2d and batchnorm first 11 layers based on reference model.
+            # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py#L194
+            trainable=self._trainable,
+            finetune_bn=self._finetune_bn
+        )
+
+        self._local_layers["block_4"] = BlockGroup(
+            filters=512,
+            strides=2,
+            n_blocks=self._n_layers[3],
+            block_layer=self._block_layer,
+            data_format=self._data_format,
+            # Freeze at conv2d and batchnorm first 11 layers based on reference model.
+            # Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py#L194
+            trainable=self._trainable,
+            finetune_bn=self._finetune_bn
+        )
+
+    def call(self, inputs, training=True, *args, **kwargs):
+        """Creation of the model graph."""
+        net = self._local_layers["conv2d"](inputs=inputs)
+
+        net = self._local_layers["batchnorm"](
+            inputs=net,
+            training=False
+        )
+
+        net = self._local_layers["maxpool2d"](net)
+
+        c2 = self._local_layers["block_1"](
+            inputs=net,
+            training=False,
+        )
+
+        c3 = self._local_layers["block_2"](
+            inputs=c2,
+            training=training,
+        )
+
+        c4 = self._local_layers["block_3"](
+            inputs=c3,
+            training=training,
+        )
+
+        c5 = self._local_layers["block_4"](
+            inputs=c4,
+            training=training,
+        )
+
+        return {2: c2, 3: c3, 4: c4, 5: c5}