new tf branch for dtk21.10.1

TensorFlow/ComputeVision/Classification/benchmarks-master/scripts/tf_cnn_benchmarks/models/experimental/deepspeech.py

+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""DeepSpeech2 model configuration.
+
+References:
+  https://arxiv.org/abs/1512.02595
+  Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import itertools
+
+import numpy as np
+from six.moves import xrange  # pylint: disable=redefined-builtin
+import tensorflow.compat.v1 as tf
+import constants
+from cnn_util import log_fn
+from models import model as model_lib
+from tensorflow.python.ops import variables  # pylint: disable=g-direct-tensorflow-import
+
+
+class DeepSpeechDecoder(object):
+  """Greedy decoder implementation for Deep Speech model."""
+
+  def __init__(self, labels, blank_index=28):
+    """Decoder initialization.
+
+    Args:
+      labels: a string specifying the speech labels for the decoder to use.
+      blank_index: an integer specifying index for the blank character. Defaults
+        to 28.
+    """
+    self.labels = labels
+    self.blank_index = blank_index
+    self.int_to_char = dict([(i, c) for (i, c) in enumerate(labels)])
+
+  def convert_to_string(self, sequence):
+    """Convert a sequence of indexes into corresponding string."""
+    return ''.join([self.int_to_char[i] for i in sequence])
+
+  def wer(self, decode, target):
+    """Computes the Word Error Rate (WER).
+
+    WER is defined as the edit distance between the two provided sentences after
+    tokenizing to words.
+
+    Args:
+      decode: string of the decoded output.
+      target: a string for the ground truth label.
+
+    Returns:
+      A float number for the WER of the current decode-target pair.
+    """
+    try:
+      from nltk.metrics import distance  # pylint: disable=g-import-not-at-top
+    except ImportError as e:
+      if 'nltk.metrics' not in e.message:
+        raise
+      raise ImportError('To use the experimental deepspeech model, you must '
+                        'pip install -U nltk')
+
+    # Map each word to a new char.
+    words = set(decode.split() + target.split())
+    word2char = dict(zip(words, range(len(words))))
+
+    new_decode = [chr(word2char[w]) for w in decode.split()]
+    new_target = [chr(word2char[w]) for w in target.split()]
+
+    return distance.edit_distance(''.join(new_decode), ''.join(new_target))
+
+  def cer(self, decode, target):
+    """Computes the Character Error Rate (CER).
+
+    CER is defined as the edit distance between the two given strings.
+
+    Args:
+      decode: a string of the decoded output.
+      target: a string for the ground truth label.
+
+    Returns:
+      A float number denoting the CER for the current sentence pair.
+    """
+    try:
+      from nltk.metrics import distance  # pylint: disable=g-import-not-at-top
+    except ImportError as e:
+      if 'nltk.metrics' not in e.message:
+        raise
+      raise ImportError('To use the experimental deepspeech model, you must '
+                        'pip install -U nltk')
+    return distance.edit_distance(decode, target)
+
+  def decode(self, char_indexes):
+    """Decode the best guess from logits using greedy algorithm."""
+    # Merge repeated chars.
+    merge = [k for k, _ in itertools.groupby(char_indexes)]
+    # Remove the blank index in the decoded sequence.
+    merge_remove_blank = []
+    for k in merge:
+      if k != self.blank_index:
+        merge_remove_blank.append(k)
+
+    return self.convert_to_string(merge_remove_blank)
+
+  def decode_logits(self, logits):
+    """Decode the best guess from logits using greedy algorithm."""
+    # Choose the class with maximimum probability.
+    best = list(np.argmax(logits, axis=1))
+    return self.decode(best)
+
+
+class DeepSpeech2Model(model_lib.Model):
+  """Define DeepSpeech2 model."""
+
+  # Supported rnn cells.
+  SUPPORTED_RNNS = {
+      'lstm': tf.nn.rnn_cell.BasicLSTMCell,
+      'rnn': tf.nn.rnn_cell.RNNCell,
+      'gru': tf.nn.rnn_cell.GRUCell,
+  }
+
+  # Parameters for batch normalization.
+  BATCH_NORM_EPSILON = 1e-5
+  BATCH_NORM_DECAY = 0.997
+
+  # Filters of convolution layer
+  CONV_FILTERS = 32
+
+  def __init__(self,
+               num_rnn_layers=5,
+               rnn_type='lstm',
+               is_bidirectional=True,
+               rnn_hidden_size=800,
+               use_bias=True,
+               params=None):
+    """Initialize DeepSpeech2 model.
+
+    Args:
+      num_rnn_layers: an integer, the number of rnn layers (default: 5).
+      rnn_type: a string, one of the supported rnn cells: gru, rnn or lstm.
+      is_bidirectional: a boolean to indicate if the rnn layer is bidirectional.
+      rnn_hidden_size: an integer for the number of hidden units in the RNN
+        cell.
+      use_bias: a boolean specifying whether to use a bias in the last fc layer.
+      params: the params from BenchmarkCNN.
+    """
+    super(DeepSpeech2Model, self).__init__(
+        'deepspeech2',
+        batch_size=128,
+        learning_rate=0.0005,
+        fp16_loss_scale=128,
+        params=params)
+    self.num_rnn_layers = num_rnn_layers
+    self.rnn_type = rnn_type
+    self.is_bidirectional = is_bidirectional
+    self.rnn_hidden_size = rnn_hidden_size
+    self.use_bias = use_bias
+    self.num_feature_bins = 161
+    self.max_time_steps = 3494
+    self.max_label_length = 576
+
+  def _batch_norm(self, inputs, training):
+    """Batch normalization layer.
+
+    Note that the momentum to use will affect validation accuracy over time.
+    Batch norm has different behaviors during training/evaluation. With a large
+    momentum, the model takes longer to get a near-accurate estimation of the
+    moving mean/variance over the entire training dataset, which means we need
+    more iterations to see good evaluation results. If the training data is
+    evenly distributed over the feature space, we can also try setting a smaller
+    momentum (such as 0.1) to get good evaluation result sooner.
+
+    Args:
+      inputs: input data for batch norm layer.
+      training: a boolean to indicate if it is in training stage.
+
+    Returns:
+      tensor output from batch norm layer.
+    """
+    return tf.layers.batch_normalization(
+        inputs=inputs,
+        momentum=DeepSpeech2Model.BATCH_NORM_DECAY,
+        epsilon=DeepSpeech2Model.BATCH_NORM_EPSILON,
+        fused=True,
+        training=training)
+
+  def _conv_bn_layer(self, inputs, padding, filters, kernel_size, strides,
+                     layer_id, training):
+    """Defines 2D convolutional + batch normalization layer.
+
+    Args:
+      inputs: input data for convolution layer.
+      padding: padding to be applied before convolution layer.
+      filters: an integer, number of output filters in the convolution.
+      kernel_size: a tuple specifying the height and width of the 2D convolution
+        window.
+      strides: a tuple specifying the stride length of the convolution.
+      layer_id: an integer specifying the layer index.
+      training: a boolean to indicate which stage we are in (training/eval).
+
+    Returns:
+      tensor output from the current layer.
+    """
+    # Perform symmetric padding on the feature dimension of time_step
+    # This step is required to avoid issues when RNN output sequence is shorter
+    # than the label length.
+    inputs = tf.pad(
+        inputs,
+        [[0, 0], [padding[0], padding[0]], [padding[1], padding[1]], [0, 0]])
+    inputs = tf.layers.conv2d(
+        inputs=inputs,
+        filters=filters,
+        kernel_size=kernel_size,
+        strides=strides,
+        padding='valid',
+        use_bias=False,
+        activation=tf.nn.relu6,
+        name='cnn_{}'.format(layer_id))
+    return self._batch_norm(inputs, training)
+
+  def _rnn_layer(self, inputs, rnn_cell, rnn_hidden_size, layer_id,
+                 use_batch_norm, is_bidirectional, training):
+    """Defines a batch normalization + rnn layer.
+
+    Args:
+      inputs: input tensors for the current layer.
+      rnn_cell: RNN cell instance to use.
+      rnn_hidden_size: an integer for the dimensionality of the rnn output
+        space.
+      layer_id: an integer for the index of current layer.
+      use_batch_norm: a boolean specifying whether to perform batch
+        normalization on input states.
+      is_bidirectional: a boolean specifying whether the rnn layer is
+        bi-directional.
+      training: a boolean to indicate which stage we are in (training/eval).
+
+    Returns:
+      tensor output for the current layer.
+    """
+    if use_batch_norm:
+      inputs = self._batch_norm(inputs, training)
+
+    # Construct forward/backward RNN cells.
+    fw_cell = rnn_cell(
+        num_units=rnn_hidden_size, name='rnn_fw_{}'.format(layer_id))
+
+    if is_bidirectional:
+      bw_cell = rnn_cell(
+          num_units=rnn_hidden_size, name='rnn_bw_{}'.format(layer_id))
+      outputs, _ = tf.nn.bidirectional_dynamic_rnn(
+          cell_fw=fw_cell,
+          cell_bw=bw_cell,
+          inputs=inputs,
+          dtype=tf.float32,
+          swap_memory=True)
+      rnn_outputs = tf.concat(outputs, -1)
+    else:
+      rnn_outputs = tf.nn.dynamic_rnn(
+          fw_cell, inputs, dtype=tf.float32, swap_memory=True)
+
+    return rnn_outputs
+
+  def get_input_data_types(self, subset):
+    """Returns the list of data types of the inputs."""
+    del subset  # Same data types for both train and validation subsets.
+    return [self.data_type, tf.int32, tf.int32, tf.int32]
+
+  def get_input_shapes(self, subset):
+    """Returns the list of shapes of the padded inputs."""
+    del subset  # Same shapes for both train and validation subsets
+    return [
+        [self.batch_size, self.max_time_steps, self.num_feature_bins, 1],
+        [self.batch_size, self.max_label_length],
+        [self.batch_size, 1],
+        [self.batch_size, 1],
+    ]
+
+  def get_synthetic_inputs(self, input_name, nclass):
+    inputs = tf.random_uniform(self.get_input_shapes('train')[0],
+                               dtype=self.get_input_data_types('train')[0])
+    inputs = variables.VariableV1(inputs, trainable=False,
+                                  collections=[tf.GraphKeys.LOCAL_VARIABLES],
+                                  name=input_name)
+    labels = tf.convert_to_tensor(
+        np.random.randint(28, size=[self.batch_size, self.max_label_length]))
+    input_lengths = tf.convert_to_tensor(
+        [self.max_time_steps] * self.batch_size)
+    label_lengths = tf.convert_to_tensor(
+        [self.max_label_length] * self.batch_size)
+    return [inputs, labels, input_lengths, label_lengths]
+
+  # TODO(laigd): support fp16.
+  # TODO(laigd): support multiple gpus.
+  def build_network(self, inputs, phase_train=True, nclass=29):
+    """Builds the forward pass of the deepspeech2 model.
+
+    Args:
+      inputs: The input list of the model.
+      phase_train: True during training. False during evaluation.
+      nclass: Number of classes that the input spectrogram can belong to.
+
+    Returns:
+      A BuildNetworkResult which contains the logits and model-specific extra
+        information.
+    """
+    inputs = inputs[0]  # Get the spectrogram feature.
+
+    # Two cnn layers.
+    inputs = self._conv_bn_layer(
+        inputs,
+        padding=(20, 5),
+        filters=DeepSpeech2Model.CONV_FILTERS,
+        kernel_size=(41, 11),
+        strides=(2, 2),
+        layer_id=1,
+        training=phase_train)
+
+    inputs = self._conv_bn_layer(
+        inputs,
+        padding=(10, 5),
+        filters=DeepSpeech2Model.CONV_FILTERS,
+        kernel_size=(21, 11),
+        strides=(2, 1),
+        layer_id=2,
+        training=phase_train)
+
+    # output of conv_layer2 with the shape of
+    # [batch_size (N), times (T), features (F), channels (C)].
+    # Convert the conv output to rnn input.
+
+    # batch_size = tf.shape(inputs)[0]
+    feat_size = inputs.get_shape().as_list()[2]
+    inputs = tf.reshape(
+        inputs,
+        [self.batch_size, -1, feat_size * DeepSpeech2Model.CONV_FILTERS])
+
+    # RNN layers.
+    rnn_cell = DeepSpeech2Model.SUPPORTED_RNNS[self.rnn_type]
+    for layer_counter in xrange(self.num_rnn_layers):
+      # No batch normalization on the first layer.
+      use_batch_norm = (layer_counter != 0)
+      inputs = self._rnn_layer(inputs, rnn_cell, self.rnn_hidden_size,
+                               layer_counter + 1, use_batch_norm,
+                               self.is_bidirectional, phase_train)
+
+    # FC layer with batch norm.
+    inputs = self._batch_norm(inputs, phase_train)
+    logits = tf.layers.dense(inputs, nclass, use_bias=self.use_bias)
+
+    return model_lib.BuildNetworkResult(logits=logits, extra_info=None)
+
+  def loss_function(self, inputs, build_network_result):
+    """Computes the ctc loss for the current batch of predictions.
+
+    Args:
+      inputs: the input list of the model.
+      build_network_result: a BuildNetworkResult returned by build_network().
+
+    Returns:
+      The loss tensor of the model.
+    """
+    logits = build_network_result.logits
+    actual_time_steps = inputs[2]
+    probs = tf.nn.softmax(logits)
+    ctc_time_steps = tf.shape(probs)[1]
+    ctc_input_length = tf.to_float(
+        tf.multiply(actual_time_steps, ctc_time_steps))
+    ctc_input_length = tf.to_int32(
+        tf.floordiv(ctc_input_length, tf.to_float(self.max_time_steps)))
+
+    label_length = inputs[3]
+    label_length = tf.to_int32(tf.squeeze(label_length))
+    ctc_input_length = tf.to_int32(tf.squeeze(ctc_input_length))
+
+    labels = inputs[1]
+    sparse_labels = tf.to_int32(
+        tf.keras.backend.ctc_label_dense_to_sparse(labels, label_length))
+    y_pred = tf.log(
+        tf.transpose(probs, perm=[1, 0, 2]) + tf.keras.backend.epsilon())
+
+    losses = tf.expand_dims(
+        tf.nn.ctc_loss(
+            labels=sparse_labels,
+            inputs=y_pred,
+            sequence_length=ctc_input_length,
+            ignore_longer_outputs_than_inputs=True),
+        axis=1)
+    loss = tf.reduce_mean(losses)
+    return loss
+
+  PROBABILITY_TENSOR = 'deepspeech2_prob'
+  LABEL_TENSOR = 'deepspeech2_label'
+
+  def accuracy_function(self, inputs, logits):
+    """Returns the ops to evaluate the model performance."""
+    # Get probabilities of each predicted class
+    probs = tf.nn.softmax(logits)
+    assert probs.shape.as_list()[0] == self.batch_size
+    return {
+        (constants.UNREDUCED_ACCURACY_OP_PREFIX + self.PROBABILITY_TENSOR):
+            probs,
+        (constants.UNREDUCED_ACCURACY_OP_PREFIX + self.LABEL_TENSOR):
+            inputs[1],
+    }
+
+  def postprocess(self, results):
+    """Postprocess results returned from model in Python."""
+    probs = results[self.PROBABILITY_TENSOR]
+
+    total_wer, total_cer = 0, 0
+    speech_labels = " abcdefghijklmnopqrstuvwxyz'-"
+    greedy_decoder = DeepSpeechDecoder(speech_labels)
+
+    # Evaluate the performance using WER (Word Error Rate) and CER (Character
+    # Error Rate) as metrics.
+    targets = results[self.LABEL_TENSOR]  # The ground truth transcript
+    for i in range(self.batch_size):
+      # Decode string.
+      predicted_str = greedy_decoder.decode_logits(probs[i])
+      expected_str = greedy_decoder.decode(targets[i])
+      # Compute CER.
+      total_cer += (greedy_decoder.cer(predicted_str, expected_str) /
+                    len(expected_str))
+      # Compute WER.
+      total_wer += (greedy_decoder.wer(predicted_str, expected_str) /
+                    len(expected_str.split()))
+
+    # Get mean value
+    total_cer /= self.batch_size
+    total_wer /= self.batch_size
+
+    log_fn('total CER: {:f}; total WER: {:f}; total example: {:d}.'.format(
+        total_cer, total_wer, self.batch_size))
+    # TODO(laigd): get rid of top_N_accuracy bindings in benchmark_cnn.py
+    return {'top_1_accuracy': 0., 'top_5_accuracy': 0.}

TensorFlow/ComputeVision/Classification/benchmarks-master/scripts/tf_cnn_benchmarks/models/official_resnet_model.py

+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Import official resnet models."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow.compat.v1 as tf
+import datasets
+from models import model as model_lib
+
+
+class ImagenetResnetModel(model_lib.CNNModel):
+  """Official resnet models."""
+
+  def __init__(self, resnet_size, version=2, params=None):
+    """These are the parameters that work for Imagenet data.
+
+    Args:
+      resnet_size: The number of convolutional layers needed in the model.
+      version: 1 or 2 for v1 or v2, respectively.
+      params: params passed by BenchmarkCNN.
+    """
+    default_batch_sizes = {
+        50: 128,
+        101: 32,
+        152: 32
+    }
+    batch_size = default_batch_sizes.get(resnet_size, 32)
+    default_learning_rate = 0.0125 * batch_size / 32
+    model_name = 'official_resnet_{}_v{}'.format(resnet_size, version)
+    super(ImagenetResnetModel, self).__init__(
+        model_name, 224, batch_size, default_learning_rate, params=params)
+    self.resnet_size = resnet_size
+    self.version = version
+
+  def get_learning_rate(self, global_step, batch_size):
+    num_batches_per_epoch = (
+        float(datasets.IMAGENET_NUM_TRAIN_IMAGES) / batch_size)
+    boundaries = [int(num_batches_per_epoch * x) for x in [30, 60, 80, 90]]
+    values = [1, 0.1, 0.01, 0.001, 0.0001]
+    adjusted_learning_rate = (
+        self.learning_rate / self.default_batch_size * batch_size)
+    values = [v * adjusted_learning_rate for v in values]
+    return tf.train.piecewise_constant(global_step, boundaries, values)
+
+  def build_network(self, images, phase_train=True, nclass=1001,
+                    data_type=tf.float32):
+    # pylint: disable=g-import-not-at-top
+    try:
+      from official.r1.resnet.imagenet_main import ImagenetModel
+    except ImportError:
+      tf.logging.fatal('Please include tensorflow/models to the PYTHONPATH.')
+      raise
+    images = tf.cast(images, data_type)
+    model_class = ImagenetModel(resnet_size=self.resnet_size,
+                                resnet_version=self.version,
+                                # The official model dtype seems to be ignored,
+                                # as the dtype it uses is the dtype of the input
+                                # images. Doesn't hurt to set it though.
+                                dtype=data_type)
+    logits = model_class(images, phase_train)
+    logits = tf.cast(logits, tf.float32)
+    return model_lib.BuildNetworkResult(logits=logits, extra_info=None)

TensorFlow/ComputeVision/Classification/benchmarks-master/scripts/tf_cnn_benchmarks/models/tf1_only/mobilenet.py

+# 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.
+# ==============================================================================
+"""Mobilenet Base Class, branched from slim for fp16 performance study."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import collections
+import contextlib
+import copy
+import os
+
+import tensorflow.compat.v1 as tf
+from tensorflow.contrib import slim as contrib_slim
+
+slim = contrib_slim
+
+
+@slim.add_arg_scope
+def apply_activation(x, name=None, activation_fn=None):
+  return activation_fn(x, name=name) if activation_fn else x
+
+
+def _fixed_padding(inputs, kernel_size, rate=1):
+  """Pads the input along the spatial dimensions independently of input size.
+
+  Pads the input such that if it was used in a convolution with 'VALID' padding,
+  the output would have the same dimensions as if the unpadded input was used
+  in a convolution with 'SAME' padding.
+
+  Args:
+    inputs: A tensor of size [batch, height_in, width_in, channels].
+    kernel_size: The kernel to be used in the conv2d or max_pool2d operation.
+    rate: An integer, rate for atrous convolution.
+
+  Returns:
+    output: A tensor of size [batch, height_out, width_out, channels] with the
+      input, either intact (if kernel_size == 1) or padded (if kernel_size > 1).
+  """
+  kernel_size_effective = [kernel_size[0] + (kernel_size[0] - 1) * (rate - 1),
+                           kernel_size[0] + (kernel_size[0] - 1) * (rate - 1)]
+  pad_total = [kernel_size_effective[0] - 1, kernel_size_effective[1] - 1]
+  pad_beg = [pad_total[0] // 2, pad_total[1] // 2]
+  pad_end = [pad_total[0] - pad_beg[0], pad_total[1] - pad_beg[1]]
+  padded_inputs = tf.pad(inputs, [[0, 0], [pad_beg[0], pad_end[0]],
+                                  [pad_beg[1], pad_end[1]], [0, 0]])
+  return padded_inputs
+
+
+def _make_divisible(v, divisor, min_value=None):
+  if min_value is None:
+    min_value = divisor
+  new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+  # Make sure that round down does not go down by more than 10%.
+  if new_v < 0.9 * v:
+    new_v += divisor
+  return new_v
+
+
+@contextlib.contextmanager
+def _set_arg_scope_defaults(defaults):
+  """Sets arg scope defaults for all items present in defaults.
+
+  Args:
+    defaults: dictionary/list of pairs, containing a mapping from
+    function to a dictionary of default args.
+
+  Yields:
+    context manager where all defaults are set.
+  """
+  if hasattr(defaults, 'items'):
+    items = list(defaults.items())
+  else:
+    items = defaults
+  if not items:
+    yield
+  else:
+    func, default_arg = items[0]
+    with slim.arg_scope(func, **default_arg):
+      with _set_arg_scope_defaults(items[1:]):
+        yield
+
+
+@slim.add_arg_scope
+def depth_multiplier(output_params,
+                     multiplier,
+                     divisible_by=8,
+                     min_depth=8,
+                     **unused_kwargs):
+  if 'num_outputs' not in output_params:
+    return
+  d = output_params['num_outputs']
+  output_params['num_outputs'] = _make_divisible(d * multiplier, divisible_by,
+                                                 min_depth)
+
+
+_Op = collections.namedtuple('Op', ['op', 'params', 'multiplier_func'])
+
+
+def op(opfunc, **params):
+  multiplier = params.pop('multiplier_transorm', depth_multiplier)
+  return _Op(opfunc, params=params, multiplier_func=multiplier)
+
+
+class NoOpScope(object):
+  """No-op context manager."""
+
+  def __enter__(self):
+    return
+
+  def __exit__(self, exc_type, exc_value, traceback):
+    return False
+
+
+def safe_arg_scope(funcs, **kwargs):
+  """Returns `slim.arg_scope` with all None arguments removed.
+
+  Args:
+    funcs: Functions to pass to `arg_scope`.
+    **kwargs: Arguments to pass to `arg_scope`.
+
+  Returns:
+    arg_scope or No-op context manager.
+
+  Note: can be useful if None value should be interpreted as "do not overwrite
+    this parameter value".
+  """
+  filtered_args = {name: value for name, value in kwargs.items()
+                   if value is not None}
+  if filtered_args:
+    return slim.arg_scope(funcs, **filtered_args)
+  else:
+    return NoOpScope()
+
+
+@slim.add_arg_scope
+def mobilenet_base(  # pylint: disable=invalid-name
+    inputs,
+    conv_defs,
+    multiplier=1.0,
+    final_endpoint=None,
+    output_stride=None,
+    use_explicit_padding=False,
+    scope=None,
+    is_training=False):
+  """Mobilenet base network.
+
+  Constructs a network from inputs to the given final endpoint. By default
+  the network is constructed in inference mode. To create network
+  in training mode use:
+
+  with slim.arg_scope(mobilenet.training_scope()):
+     logits, endpoints = mobilenet_base(...)
+
+  Args:
+    inputs: a tensor of shape [batch_size, height, width, channels].
+    conv_defs: A list of op(...) layers specifying the net architecture.
+    multiplier: Float multiplier for the depth (number of channels)
+      for all convolution ops. The value must be greater than zero. Typical
+      usage will be to set this value in (0, 1) to reduce the number of
+      parameters or computation cost of the model.
+    final_endpoint: The name of last layer, for early termination for
+    for V1-based networks: last layer is "layer_14", for V2: "layer_20"
+    output_stride: An integer that specifies the requested ratio of input to
+      output spatial resolution. If not None, then we invoke atrous convolution
+      if necessary to prevent the network from reducing the spatial resolution
+      of the activation maps. Allowed values are 1 or any even number, excluding
+      zero. Typical values are 8 (accurate fully convolutional mode), 16
+      (fast fully convolutional mode), and 32 (classification mode).
+
+      NOTE- output_stride relies on all consequent operators to support dilated
+      operators via "rate" parameter. This might require wrapping non-conv
+      operators to operate properly.
+
+    use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
+      inputs so that the output dimensions are the same as if 'SAME' padding
+      were used.
+    scope: optional variable scope.
+    is_training: How to setup batch_norm and other ops. Note: most of the time
+      this does not need be set directly. Use mobilenet.training_scope() to set
+      up training instead. This parameter is here for backward compatibility
+      only. It is safe to set it to the value matching
+      training_scope(is_training=...). It is also safe to explicitly set
+      it to False, even if there is outer training_scope set to to training.
+      (The network will be built in inference mode). If this is set to None,
+      no arg_scope is added for slim.batch_norm's is_training parameter.
+
+  Returns:
+    tensor_out: output tensor.
+    end_points: a set of activations for external use, for example summaries or
+                losses.
+
+  Raises:
+    ValueError: depth_multiplier <= 0, or the target output_stride is not
+                allowed.
+  """
+  if multiplier <= 0:
+    raise ValueError('multiplier is not greater than zero.')
+
+  # Set conv defs defaults and overrides.
+  conv_defs_defaults = conv_defs.get('defaults', {})
+  conv_defs_overrides = conv_defs.get('overrides', {})
+  if use_explicit_padding:
+    conv_defs_overrides = copy.deepcopy(conv_defs_overrides)
+    conv_defs_overrides[
+        (slim.conv2d, slim.separable_conv2d)] = {'padding': 'VALID'}
+
+  if output_stride is not None:
+    if output_stride == 0 or (output_stride > 1 and output_stride % 2):
+      raise ValueError('Output stride must be None, 1 or a multiple of 2.')
+
+  # a) Set the tensorflow scope
+  # b) set padding to default: note we might consider removing this
+  # since it is also set by mobilenet_scope
+  # c) set all defaults
+  # d) set all extra overrides.
+  with _scope_all(scope, default_scope='Mobilenet'), \
+      safe_arg_scope([slim.batch_norm], is_training=is_training), \
+      _set_arg_scope_defaults(conv_defs_defaults), \
+      _set_arg_scope_defaults(conv_defs_overrides):
+    # The current_stride variable keeps track of the output stride of the
+    # activations, i.e., the running product of convolution strides up to the
+    # current network layer. This allows us to invoke atrous convolution
+    # whenever applying the next convolution would result in the activations
+    # having output stride larger than the target output_stride.
+    current_stride = 1
+
+    # The atrous convolution rate parameter.
+    rate = 1
+
+    net = inputs
+    # Insert default parameters before the base scope which includes
+    # any custom overrides set in mobilenet.
+    end_points = {}
+    scopes = {}
+    for i, opdef in enumerate(conv_defs['spec']):
+      params = dict(opdef.params)
+      opdef.multiplier_func(params, multiplier)
+      stride = params.get('stride', 1)
+      if output_stride is not None and current_stride == output_stride:
+        # If we have reached the target output_stride, then we need to employ
+        # atrous convolution with stride=1 and multiply the atrous rate by the
+        # current unit's stride for use in subsequent layers.
+        layer_stride = 1
+        layer_rate = rate
+        rate *= stride
+      else:
+        layer_stride = stride
+        layer_rate = 1
+        current_stride *= stride
+      # Update params.
+      params['stride'] = layer_stride
+      # Only insert rate to params if rate > 1.
+      if layer_rate > 1:
+        params['rate'] = layer_rate
+      # Set padding
+      if use_explicit_padding:
+        if 'kernel_size' in params:
+          net = _fixed_padding(net, params['kernel_size'], layer_rate)
+        else:
+          params['use_explicit_padding'] = True
+
+      end_point = 'layer_%d' % (i + 1)
+      try:
+        net = opdef.op(net, **params)
+      except Exception:
+        print('Failed to create op %i: %r params: %r' % (i, opdef, params))
+        raise
+      end_points[end_point] = net
+      scope = os.path.dirname(net.name)
+      scopes[scope] = end_point
+      if final_endpoint is not None and end_point == final_endpoint:
+        break
+
+    # Add all tensors that end with 'output' to
+    # endpoints
+    for t in net.graph.get_operations():
+      scope = os.path.dirname(t.name)
+      bn = os.path.basename(t.name)
+      if scope in scopes and t.name.endswith('output'):
+        end_points[scopes[scope] + '/' + bn] = t.outputs[0]
+    return net, end_points
+
+
+@contextlib.contextmanager
+def _scope_all(scope, default_scope=None):
+  with tf.variable_scope(scope, default_name=default_scope) as s,\
+       tf.name_scope(s.original_name_scope):
+    yield s
+
+
+@slim.add_arg_scope
+def mobilenet(inputs,
+              num_classes=1001,
+              prediction_fn=slim.softmax,
+              reuse=None,
+              scope='Mobilenet',
+              base_only=False,
+              **mobilenet_args):
+  """Mobilenet model for classification, supports both V1 and V2.
+
+  Note: default mode is inference, use mobilenet.training_scope to create
+  training network.
+
+
+  Args:
+    inputs: a tensor of shape [batch_size, height, width, channels].
+    num_classes: number of predicted classes. If 0 or None, the logits layer
+      is omitted and the input features to the logits layer (before dropout)
+      are returned instead.
+    prediction_fn: a function to get predictions out of logits
+      (default softmax).
+    reuse: whether or not the network and its variables should be reused. To be
+      able to reuse 'scope' must be given.
+    scope: Optional variable_scope.
+    base_only: if True will only create the base of the network (no pooling
+    and no logits).
+    **mobilenet_args: passed to mobilenet_base verbatim.
+      - conv_defs: list of conv defs
+      - multiplier: Float multiplier for the depth (number of channels)
+      for all convolution ops. The value must be greater than zero. Typical
+      usage will be to set this value in (0, 1) to reduce the number of
+      parameters or computation cost of the model.
+      - output_stride: will ensure that the last layer has at most total stride.
+      If the architecture calls for more stride than that provided
+      (e.g. output_stride=16, but the architecture has 5 stride=2 operators),
+      it will replace output_stride with fractional convolutions using Atrous
+      Convolutions.
+
+  Returns:
+    logits: the pre-softmax activations, a tensor of size
+      [batch_size, num_classes]
+    end_points: a dictionary from components of the network to the corresponding
+      activation tensor.
+
+  Raises:
+    ValueError: Input rank is invalid.
+  """
+  is_training = mobilenet_args.get('is_training', False)
+  input_shape = inputs.get_shape().as_list()
+  if len(input_shape) != 4:
+    raise ValueError('Expected rank 4 input, was: %d' % len(input_shape))
+
+  with tf.variable_scope(scope, 'Mobilenet', reuse=reuse) as scope:
+    inputs = tf.identity(inputs, 'input')
+    net, end_points = mobilenet_base(inputs, scope=scope, **mobilenet_args)
+    if base_only:
+      return net, end_points
+
+    net = tf.identity(net, name='embedding')
+
+    with tf.variable_scope('Logits'):
+      net = global_pool(net)
+      end_points['global_pool'] = net
+      if not num_classes:
+        return net, end_points
+      net = slim.dropout(net, scope='Dropout', is_training=is_training)
+      # 1 x 1 x num_classes
+      # Note: legacy scope name.
+      logits = slim.conv2d(
+          net,
+          num_classes, [1, 1],
+          activation_fn=None,
+          normalizer_fn=None,
+          biases_initializer=tf.zeros_initializer(),
+          scope='Conv2d_1c_1x1')
+
+      logits = tf.squeeze(logits, [1, 2])
+
+      logits = tf.identity(logits, name='output')
+    end_points['Logits'] = logits
+    if prediction_fn:
+      end_points['Predictions'] = prediction_fn(logits, 'Predictions')
+  return logits, end_points
+
+
+def global_pool(input_tensor, pool_op=tf.nn.avg_pool):
+  """Applies avg pool to produce 1x1 output.
+
+  NOTE: This function is funcitonally equivalenet to reduce_mean, but it has
+  baked in average pool which has better support across hardware.
+
+  Args:
+    input_tensor: input tensor
+    pool_op: pooling op (avg pool is default)
+  Returns:
+    a tensor batch_size x 1 x 1 x depth.
+  """
+  shape = input_tensor.get_shape().as_list()
+  if shape[1] is None or shape[2] is None:
+    kernel_size = tf.convert_to_tensor(
+        [1, tf.shape(input_tensor)[1],
+         tf.shape(input_tensor)[2], 1])
+  else:
+    kernel_size = [1, shape[1], shape[2], 1]
+  output = pool_op(
+      input_tensor, ksize=kernel_size, strides=[1, 1, 1, 1], padding='VALID')
+  # Recover output shape, for unknown shape.
+  output.set_shape([None, 1, 1, None])
+  return output
+
+
+def training_scope(is_training=True,
+                   weight_decay=0.00004,
+                   stddev=0.09,
+                   dropout_keep_prob=0.8,
+                   bn_decay=0.997):
+  """Defines Mobilenet training scope.
+
+  Usage:
+     with tf.contrib.slim.arg_scope(mobilenet.training_scope()):
+       logits, endpoints = mobilenet_v2.mobilenet(input_tensor)
+
+     # the network created will be trainble with dropout/batch norm
+     # initialized appropriately.
+  Args:
+    is_training: if set to False this will ensure that all customizations are
+      set to non-training mode. This might be helpful for code that is reused
+      across both training/evaluation, but most of the time training_scope with
+      value False is not needed. If this is set to None, the parameters is not
+      added to the batch_norm arg_scope.
+
+    weight_decay: The weight decay to use for regularizing the model.
+    stddev: Standard deviation for initialization, if negative uses xavier.
+    dropout_keep_prob: dropout keep probability (not set if equals to None).
+    bn_decay: decay for the batch norm moving averages (not set if equals to
+      None).
+
+  Returns:
+    An argument scope to use via arg_scope.
+  """
+  # Note: do not introduce parameters that would change the inference
+  # model here (for example whether to use bias), modify conv_def instead.
+  batch_norm_params = {
+      'decay': bn_decay,
+      'is_training': is_training
+  }
+  if stddev < 0:
+    weight_intitializer = slim.initializers.xavier_initializer()
+  else:
+    weight_intitializer = tf.truncated_normal_initializer(stddev=stddev)
+
+  # Set weight_decay for weights in Conv and FC layers.
+  with slim.arg_scope(
+      [slim.conv2d, slim.fully_connected, slim.separable_conv2d],
+      weights_initializer=weight_intitializer,
+      normalizer_fn=slim.batch_norm), \
+      slim.arg_scope([mobilenet_base, mobilenet], is_training=is_training),\
+      safe_arg_scope([slim.batch_norm], **batch_norm_params), \
+      safe_arg_scope([slim.dropout], is_training=is_training,
+                     keep_prob=dropout_keep_prob), \
+      slim.arg_scope([slim.conv2d], \
+                     weights_regularizer=slim.l2_regularizer(weight_decay)), \
+      slim.arg_scope([slim.separable_conv2d], weights_regularizer=None) as s:
+    return s

TensorFlow/ComputeVision/Classification/benchmarks-master/scripts/tf_cnn_benchmarks/models/tf1_only/ssd_model.py

+# 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.
+# ==============================================================================
+
+
+"""SSD300 Model Configuration.
+
+References:
+  Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed,
+  Cheng-Yang Fu, Alexander C. Berg
+  SSD: Single Shot MultiBox Detector
+  arXiv:1512.02325
+
+Ported from MLPerf reference implementation:
+  https://github.com/mlperf/reference/tree/ssd/single_stage_detector/ssd
+
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import multiprocessing
+import os
+import re
+import threading
+import tensorflow.compat.v1 as tf
+
+# pylint: disable=g-direct-tensorflow-import
+import constants
+import mlperf
+import ssd_constants
+from cnn_util import log_fn
+from models import model as model_lib
+from models import resnet_model
+from tensorflow.contrib import layers as contrib_layers
+from tensorflow.python.ops import variables
+
+BACKBONE_MODEL_SCOPE_NAME = 'resnet34_backbone'
+
+
+class SSD300Model(model_lib.CNNModel):
+  """Single Shot Multibox Detection (SSD) model for 300x300 image datasets."""
+
+  def __init__(self, label_num=ssd_constants.NUM_CLASSES, batch_size=32,
+               learning_rate=1e-3, backbone='resnet34', params=None):
+    super(SSD300Model, self).__init__('ssd300', 300, batch_size, learning_rate,
+                                      params=params)
+    # For COCO dataset, 80 categories + 1 background = 81 labels
+    self.label_num = label_num
+
+    # Currently only support ResNet-34 as backbone model
+    if backbone != 'resnet34':
+      raise ValueError('Invalid backbone model %s for SSD.' % backbone)
+    mlperf.logger.log(key=mlperf.tags.BACKBONE, value=backbone)
+
+    # Number of channels and default boxes associated with the following layers:
+    #   ResNet34 layer, Conv7, Conv8_2, Conv9_2, Conv10_2, Conv11_2
+    self.out_chan = [256, 512, 512, 256, 256, 256]
+    mlperf.logger.log(key=mlperf.tags.LOC_CONF_OUT_CHANNELS,
+                      value=self.out_chan)
+
+    # Number of default boxes from layers of different scales
+    #   38x38x4, 19x19x6, 10x10x6, 5x5x6, 3x3x4, 1x1x4
+    self.num_dboxes = [4, 6, 6, 6, 4, 4]
+    mlperf.logger.log(key=mlperf.tags.NUM_DEFAULTS_PER_CELL,
+                      value=self.num_dboxes)
+
+    # TODO(haoyuzhang): in order to correctly restore in replicated mode, need
+    # to create a saver for each tower before graph is finalized. Use variable
+    # manager for better efficiency.
+    self.backbone_savers = []
+
+    # Collected predictions for eval stage. It maps each image id in eval
+    # dataset to a dict containing the following information:
+    #   source_id: raw ID of image
+    #   raw_shape: raw shape of image
+    #   pred_box: encoded box coordinates of prediction
+    #   pred_scores: scores of classes in prediction
+    self.predictions = {}
+
+    # Global step when predictions are collected.
+    self.eval_global_step = 0
+
+    # Average precision. In asynchronous eval mode, this is the latest AP we
+    # get so far and may not be the results at current eval step.
+    self.eval_coco_ap = 0
+
+    # Process, queues, and thread for asynchronous evaluation. When enabled,
+    # create a separate process (async_eval_process) that continuously pull
+    # intermediate results from the predictions queue (a multiprocessing queue),
+    # process them, and push final results into results queue (another
+    # multiprocessing queue). The main thread is responsible to push messages
+    # into predictions queue, and start a separate thread to continuously pull
+    # messages from results queue to update final results.
+    # Message in predictions queue should be a tuple of two elements:
+    #    (evaluation step, predictions)
+    # Message in results queue should be a tuple of two elements:
+    #    (evaluation step, final results)
+    self.async_eval_process = None
+    self.async_eval_predictions_queue = None
+    self.async_eval_results_queue = None
+    self.async_eval_results_getter_thread = None
+
+    # The MLPerf reference uses a starting lr of 1e-3 at bs=32.
+    self.base_lr_batch_size = 32
+
+  def skip_final_affine_layer(self):
+    return True
+
+  def gpu_preprocess_nhwc(self, images, phase_train=True):
+    try:
+      import ssd_dataloader  # pylint: disable=g-import-not-at-top
+    except ImportError:
+      raise ImportError('To use the COCO dataset, you must clone the '
+                        'repo https://github.com/tensorflow/models and add '
+                        'tensorflow/models and tensorflow/models/research to '
+                        'the PYTHONPATH, and compile the protobufs by '
+                        'following https://github.com/tensorflow/models/blob/'
+                        'master/research/object_detection/g3doc/installation.md'
+                        '#protobuf-compilation ; To evaluate using COCO'
+                        'metric, download and install Python COCO API from'
+                        'https://github.com/cocodataset/cocoapi')
+
+    if phase_train:
+      images = ssd_dataloader.color_jitter(
+          images, brightness=0.125, contrast=0.5, saturation=0.5, hue=0.05)
+      images = ssd_dataloader.normalize_image(images)
+    return images
+
+  def add_backbone_model(self, cnn):
+    # --------------------------------------------------------------------------
+    # Resnet-34 backbone model -- modified for SSD
+    # --------------------------------------------------------------------------
+
+    # Input 300x300, output 150x150
+    cnn.conv(64, 7, 7, 2, 2, mode='SAME_RESNET', use_batch_norm=True)
+    cnn.mpool(3, 3, 2, 2, mode='SAME')
+
+    resnet34_layers = [3, 4, 6, 3]
+    version = 'v1'
+
+    # ResNet-34 block group 1
+    # Input 150x150, output 75x75
+    for i in range(resnet34_layers[0]):
+      # Last argument forces residual_block to use projection shortcut, even
+      # though the numbers of input and output channels are equal
+      resnet_model.residual_block(cnn, 64, 1, version)
+
+    # ResNet-34 block group 2
+    # Input 75x75, output 38x38
+    for i in range(resnet34_layers[1]):
+      stride = 2 if i == 0 else 1
+      resnet_model.residual_block(cnn, 128, stride, version, i == 0)
+
+    # ResNet-34 block group 3
+    # This block group is modified: first layer uses stride=1 so that the image
+    # size does not change in group of layers
+    # Input 38x38, output 38x38
+    for i in range(resnet34_layers[2]):
+      # The following line is intentionally commented out to differentiate from
+      # the original ResNet-34 model
+      # stride = 2 if i == 0 else 1
+      resnet_model.residual_block(cnn, 256, stride, version, i == 0)
+
+    # ResNet-34 block group 4: removed final block group
+    # The following 3 lines are intentionally commented out to differentiate
+    # from the original ResNet-34 model
+    # for i in range(resnet34_layers[3]):
+    #   stride = 2 if i == 0 else 1
+    #   resnet_model.residual_block(cnn, 512, stride, version, i == 0)
+
+  def add_inference(self, cnn):
+    cnn.use_batch_norm = True
+    cnn.batch_norm_config = {'decay': ssd_constants.BATCH_NORM_DECAY,
+                             'epsilon': ssd_constants.BATCH_NORM_EPSILON,
+                             'scale': True}
+
+    with tf.variable_scope(BACKBONE_MODEL_SCOPE_NAME):
+      self.add_backbone_model(cnn)
+
+    # --------------------------------------------------------------------------
+    # SSD additional layers
+    # --------------------------------------------------------------------------
+
+    def add_ssd_layer(cnn, depth, k_size, stride, mode):
+      return cnn.conv(
+          depth,
+          k_size,
+          k_size,
+          stride,
+          stride,
+          mode=mode,
+          use_batch_norm=False,
+          kernel_initializer=contrib_layers.xavier_initializer())
+
+    # Activations for feature maps of different layers
+    self.activations = [cnn.top_layer]
+    # Conv7_1, Conv7_2
+    # Input 38x38, output 19x19
+    add_ssd_layer(cnn, 256, 1, 1, 'valid')
+    self.activations.append(add_ssd_layer(cnn, 512, 3, 2, 'same'))
+
+    # Conv8_1, Conv8_2
+    # Input 19x19, output 10x10
+    add_ssd_layer(cnn, 256, 1, 1, 'valid')
+    self.activations.append(add_ssd_layer(cnn, 512, 3, 2, 'same'))
+
+    # Conv9_1, Conv9_2
+    # Input 10x10, output 5x5
+    add_ssd_layer(cnn, 128, 1, 1, 'valid')
+    self.activations.append(add_ssd_layer(cnn, 256, 3, 2, 'same'))
+
+    # Conv10_1, Conv10_2
+    # Input 5x5, output 3x3
+    add_ssd_layer(cnn, 128, 1, 1, 'valid')
+    self.activations.append(add_ssd_layer(cnn, 256, 3, 1, 'valid'))
+
+    # Conv11_1, Conv11_2
+    # Input 3x3, output 1x1
+    add_ssd_layer(cnn, 128, 1, 1, 'valid')
+    self.activations.append(add_ssd_layer(cnn, 256, 3, 1, 'valid'))
+
+    self.loc = []
+    self.conf = []
+
+    for nd, ac, oc in zip(self.num_dboxes, self.activations, self.out_chan):
+      l = cnn.conv(
+          nd * 4,
+          3,
+          3,
+          1,
+          1,
+          input_layer=ac,
+          num_channels_in=oc,
+          activation=None,
+          use_batch_norm=False,
+          kernel_initializer=contrib_layers.xavier_initializer())
+      scale = l.get_shape()[-1]
+      # shape = [batch_size, nd * 4, scale, scale]
+      l = tf.reshape(l, [self.batch_size, nd, 4, scale, scale])
+      # shape = [batch_size, nd, 4, scale, scale]
+      l = tf.transpose(l, [0, 1, 3, 4, 2])
+      # shape = [batch_size, nd, scale, scale, 4]
+      self.loc.append(tf.reshape(l, [self.batch_size, -1, 4]))
+      # shape = [batch_size, nd * scale * scale, 4]
+
+      c = cnn.conv(
+          nd * self.label_num,
+          3,
+          3,
+          1,
+          1,
+          input_layer=ac,
+          num_channels_in=oc,
+          activation=None,
+          use_batch_norm=False,
+          kernel_initializer=contrib_layers.xavier_initializer())
+      # shape = [batch_size, nd * label_num, scale, scale]
+      c = tf.reshape(c, [self.batch_size, nd, self.label_num, scale, scale])
+      # shape = [batch_size, nd, label_num, scale, scale]
+      c = tf.transpose(c, [0, 1, 3, 4, 2])
+      # shape = [batch_size, nd, scale, scale, label_num]
+      self.conf.append(tf.reshape(c, [self.batch_size, -1, self.label_num]))
+      # shape = [batch_size, nd * scale * scale, label_num]
+
+    # Shape of locs: [batch_size, NUM_SSD_BOXES, 4]
+    # Shape of confs: [batch_size, NUM_SSD_BOXES, label_num]
+    locs, confs = tf.concat(self.loc, 1), tf.concat(self.conf, 1)
+
+    # Pack location and confidence outputs into a single output layer
+    # Shape of logits: [batch_size, NUM_SSD_BOXES, 4+label_num]
+    logits = tf.concat([locs, confs], 2)
+
+    cnn.top_layer = logits
+    cnn.top_size = 4 + self.label_num
+
+    return cnn.top_layer
+
+  def get_learning_rate(self, global_step, batch_size):
+    rescaled_lr = self.get_scaled_base_learning_rate(batch_size)
+    # Defined in MLPerf reference model
+    boundaries = [160000, 200000]
+    boundaries = [b * self.base_lr_batch_size // batch_size for b in boundaries]
+    decays = [1, 0.1, 0.01]
+    learning_rates = [rescaled_lr * d for d in decays]
+    lr = tf.train.piecewise_constant(global_step, boundaries, learning_rates)
+    warmup_steps = int(118287 / batch_size * 5)
+    warmup_lr = (
+        rescaled_lr * tf.cast(global_step, tf.float32) / tf.cast(
+            warmup_steps, tf.float32))
+    return tf.cond(global_step < warmup_steps, lambda: warmup_lr, lambda: lr)
+
+  def get_scaled_base_learning_rate(self, batch_size):
+    """Calculates base learning rate for creating lr schedule.
+
+    In replicated mode, gradients are summed rather than averaged which, with
+    the sgd and momentum optimizers, increases the effective learning rate by
+    lr * num_gpus. Dividing the base lr by num_gpus negates the increase.
+
+    Args:
+      batch_size: Total batch-size.
+
+    Returns:
+      Base learning rate to use to create lr schedule.
+    """
+    base_lr = self.learning_rate
+    if self.params.variable_update == 'replicated':
+      base_lr = self.learning_rate / self.params.num_gpus
+    scaled_lr = base_lr * (batch_size / self.base_lr_batch_size)
+    return scaled_lr
+
+  def _collect_backbone_vars(self):
+    backbone_vars = tf.get_collection(
+        tf.GraphKeys.GLOBAL_VARIABLES, scope='.*'+ BACKBONE_MODEL_SCOPE_NAME)
+    var_list = {}
+
+    # Assume variables in the checkpoint are following the naming convention of
+    # a model checkpoint trained with TF official model
+    # TODO(haoyuzhang): the following variable name parsing is hacky and easy
+    # to break if there is change in naming convention of either benchmarks or
+    # official models.
+    for v in backbone_vars:
+      # conv2d variable example (model <-- checkpoint):
+      #   v/cg/conv24/conv2d/kernel:0 <-- conv2d_24/kernel
+      if 'conv2d' in v.name:
+        re_match = re.search(r'conv(\d+)/conv2d/(.+):', v.name)
+        if re_match:
+          layer_id = int(re_match.group(1))
+          param_name = re_match.group(2)
+          vname_in_ckpt = self._var_name_in_official_model_ckpt(
+              'conv2d', layer_id, param_name)
+          var_list[vname_in_ckpt] = v
+
+      # batchnorm varariable example:
+      #   v/cg/conv24/batchnorm25/gamma:0 <-- batch_normalization_25/gamma
+      elif 'batchnorm' in v.name:
+        re_match = re.search(r'batchnorm(\d+)/(.+):', v.name)
+        if re_match:
+          layer_id = int(re_match.group(1))
+          param_name = re_match.group(2)
+          vname_in_ckpt = self._var_name_in_official_model_ckpt(
+              'batch_normalization', layer_id, param_name)
+          var_list[vname_in_ckpt] = v
+
+    return var_list
+
+  def _var_name_in_official_model_ckpt(self, layer_name, layer_id, param_name):
+    """Return variable names according to convention in TF official models."""
+    vname_in_ckpt = layer_name
+    if layer_id > 0:
+      vname_in_ckpt += '_' + str(layer_id)
+    vname_in_ckpt += '/' + param_name
+    return vname_in_ckpt
+
+  def loss_function(self, inputs, build_network_result):
+    logits = build_network_result.logits
+
+    # Unpack model output back to locations and confidence scores of predictions
+    # Shape of pred_loc: [batch_size, NUM_SSD_BOXES, 4]
+    # Shape of pred_label: [batch_size, NUM_SSD_BOXES, label_num]
+    pred_loc, pred_label = tf.split(logits, [4, self.label_num], 2)
+
+    # Shape of gt_loc: [batch_size, NUM_SSD_BOXES, 4]
+    # Shape of gt_label: [batch_size, NUM_SSD_BOXES, 1]
+    # Shape of num_gt: [batch_size]
+    _, gt_loc, gt_label, num_gt = inputs
+    gt_label = tf.cast(gt_label, tf.int32)
+
+    box_loss = self._localization_loss(pred_loc, gt_loc, gt_label, num_gt)
+    class_loss = self._classification_loss(pred_label, gt_label, num_gt)
+
+    tf.summary.scalar('box_loss', tf.reduce_mean(box_loss))
+    tf.summary.scalar('class_loss', tf.reduce_mean(class_loss))
+    return class_loss + box_loss
+
+  def _localization_loss(self, pred_loc, gt_loc, gt_label, num_matched_boxes):
+    """Computes the localization loss.
+
+    Computes the localization loss using smooth l1 loss.
+    Args:
+      pred_loc: a flatten tensor that includes all predicted locations. The
+        shape is [batch_size, num_anchors, 4].
+      gt_loc: a tensor representing box regression targets in
+        [batch_size, num_anchors, 4].
+      gt_label: a tensor that represents the classification groundtruth targets.
+        The shape is [batch_size, num_anchors, 1].
+      num_matched_boxes: the number of anchors that are matched to a groundtruth
+        targets, used as the loss normalizater. The shape is [batch_size].
+    Returns:
+      box_loss: a float32 representing total box regression loss.
+    """
+    mask = tf.greater(tf.squeeze(gt_label), 0)
+    float_mask = tf.cast(mask, tf.float32)
+
+    smooth_l1 = tf.reduce_sum(tf.losses.huber_loss(
+        gt_loc, pred_loc,
+        reduction=tf.losses.Reduction.NONE
+    ), axis=2)
+    smooth_l1 = tf.multiply(smooth_l1, float_mask)
+    box_loss = tf.reduce_sum(smooth_l1, axis=1)
+
+    return tf.reduce_mean(box_loss / num_matched_boxes)
+
+  def _classification_loss(self, pred_label, gt_label, num_matched_boxes):
+    """Computes the classification loss.
+
+    Computes the classification loss with hard negative mining.
+    Args:
+      pred_label: a flatten tensor that includes all predicted class. The shape
+        is [batch_size, num_anchors, num_classes].
+      gt_label: a tensor that represents the classification groundtruth targets.
+        The shape is [batch_size, num_anchors, 1].
+      num_matched_boxes: the number of anchors that are matched to a groundtruth
+        targets. This is used as the loss normalizater.
+
+    Returns:
+      box_loss: a float32 representing total box regression loss.
+    """
+    cross_entropy = tf.losses.sparse_softmax_cross_entropy(
+        gt_label, pred_label, reduction=tf.losses.Reduction.NONE)
+
+    mask = tf.greater(tf.squeeze(gt_label), 0)
+    float_mask = tf.cast(mask, tf.float32)
+
+    # Hard example mining
+    neg_masked_cross_entropy = cross_entropy * (1 - float_mask)
+    relative_position = tf.argsort(
+        tf.argsort(
+            neg_masked_cross_entropy, direction='DESCENDING'))
+    num_neg_boxes = tf.minimum(
+        tf.to_int32(num_matched_boxes) * ssd_constants.NEGS_PER_POSITIVE,
+        ssd_constants.NUM_SSD_BOXES)
+    top_k_neg_mask = tf.cast(tf.less(
+        relative_position,
+        tf.tile(num_neg_boxes[:, tf.newaxis], (1, ssd_constants.NUM_SSD_BOXES))
+    ), tf.float32)
+
+    class_loss = tf.reduce_sum(
+        tf.multiply(cross_entropy, float_mask + top_k_neg_mask), axis=1)
+
+    return tf.reduce_mean(class_loss / num_matched_boxes)
+
+  def add_backbone_saver(self):
+    # Create saver with mapping from variable names in checkpoint of backbone
+    # model to variables in SSD model
+    backbone_var_list = self._collect_backbone_vars()
+    self.backbone_savers.append(tf.train.Saver(backbone_var_list))
+
+  def load_backbone_model(self, sess, backbone_model_path):
+    for saver in self.backbone_savers:
+      saver.restore(sess, backbone_model_path)
+
+  def get_input_data_types(self, subset):
+    if subset == 'validation':
+      return [self.data_type, tf.float32, tf.float32, tf.float32, tf.int32]
+    return [self.data_type, tf.float32, tf.float32, tf.float32]
+
+  def get_input_shapes(self, subset):
+    """Return encoded tensor shapes for train and eval data respectively."""
+    if subset == 'validation':
+      # Validation data shapes:
+      # 1. images
+      # 2. ground truth locations of boxes
+      # 3. ground truth classes of objects in boxes
+      # 4. source image IDs
+      # 5. raw image shapes
+      return [
+          [self.batch_size, self.image_size, self.image_size, self.depth],
+          [self.batch_size, ssd_constants.MAX_NUM_EVAL_BOXES, 4],
+          [self.batch_size, ssd_constants.MAX_NUM_EVAL_BOXES, 1],
+          [self.batch_size],
+          [self.batch_size, 3],
+      ]
+
+    # Training data shapes:
+    # 1. images
+    # 2. ground truth locations of boxes
+    # 3. ground truth classes of objects in boxes
+    # 4. numbers of objects in images
+    return [
+        [self.batch_size, self.image_size, self.image_size, self.depth],
+        [self.batch_size, ssd_constants.NUM_SSD_BOXES, 4],
+        [self.batch_size, ssd_constants.NUM_SSD_BOXES, 1],
+        [self.batch_size]
+    ]
+
+  def accuracy_function(self, inputs, logits):
+    """Returns the ops to measure the mean precision of the model."""
+    try:
+      import ssd_dataloader  # pylint: disable=g-import-not-at-top
+      from object_detection.box_coders import faster_rcnn_box_coder  # pylint: disable=g-import-not-at-top
+      from object_detection.core import box_coder  # pylint: disable=g-import-not-at-top
+      from object_detection.core import box_list  # pylint: disable=g-import-not-at-top
+    except ImportError:
+      raise ImportError('To use the COCO dataset, you must clone the '
+                        'repo https://github.com/tensorflow/models and add '
+                        'tensorflow/models and tensorflow/models/research to '
+                        'the PYTHONPATH, and compile the protobufs by '
+                        'following https://github.com/tensorflow/models/blob/'
+                        'master/research/object_detection/g3doc/installation.md'
+                        '#protobuf-compilation ; To evaluate using COCO'
+                        'metric, download and install Python COCO API from'
+                        'https://github.com/cocodataset/cocoapi')
+
+    # Unpack model output back to locations and confidence scores of predictions
+    # pred_locs: relative locations (coordinates) of objects in all SSD boxes
+    # shape: [batch_size, NUM_SSD_BOXES, 4]
+    # pred_labels: confidence scores of objects being of all categories
+    # shape: [batch_size, NUM_SSD_BOXES, label_num]
+    pred_locs, pred_labels = tf.split(logits, [4, self.label_num], 2)
+
+    ssd_box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
+        scale_factors=ssd_constants.BOX_CODER_SCALES)
+    anchors = box_list.BoxList(
+        tf.convert_to_tensor(ssd_dataloader.DefaultBoxes()('ltrb')))
+    pred_boxes = box_coder.batch_decode(
+        encoded_boxes=pred_locs, box_coder=ssd_box_coder, anchors=anchors)
+
+    pred_scores = tf.nn.softmax(pred_labels, axis=2)
+
+    # TODO(haoyuzhang): maybe use `gt_boxes` and `gt_classes` for visualization.
+    _, gt_boxes, gt_classes, source_id, raw_shape = inputs  # pylint: disable=unused-variable
+
+    return {
+        (constants.UNREDUCED_ACCURACY_OP_PREFIX +
+         ssd_constants.PRED_BOXES): pred_boxes,
+        (constants.UNREDUCED_ACCURACY_OP_PREFIX +
+         ssd_constants.PRED_SCORES): pred_scores,
+        # TODO(haoyuzhang): maybe use these values for visualization.
+        # constants.UNREDUCED_ACCURACY_OP_PREFIX+'gt_boxes': gt_boxes,
+        # constants.UNREDUCED_ACCURACY_OP_PREFIX+'gt_classes': gt_classes,
+        (constants.UNREDUCED_ACCURACY_OP_PREFIX +
+         ssd_constants.SOURCE_ID): source_id,
+        (constants.UNREDUCED_ACCURACY_OP_PREFIX +
+         ssd_constants.RAW_SHAPE): raw_shape
+    }
+
+  def postprocess(self, results):
+    """Postprocess results returned from model."""
+    try:
+      import coco_metric  # pylint: disable=g-import-not-at-top
+    except ImportError:
+      raise ImportError('To use the COCO dataset, you must clone the '
+                        'repo https://github.com/tensorflow/models and add '
+                        'tensorflow/models and tensorflow/models/research to '
+                        'the PYTHONPATH, and compile the protobufs by '
+                        'following https://github.com/tensorflow/models/blob/'
+                        'master/research/object_detection/g3doc/installation.md'
+                        '#protobuf-compilation ; To evaluate using COCO'
+                        'metric, download and install Python COCO API from'
+                        'https://github.com/cocodataset/cocoapi')
+
+    pred_boxes = results[ssd_constants.PRED_BOXES]
+    pred_scores = results[ssd_constants.PRED_SCORES]
+    # TODO(haoyuzhang): maybe use these values for visualization.
+    # gt_boxes = results['gt_boxes']
+    # gt_classes = results['gt_classes']
+    source_id = results[ssd_constants.SOURCE_ID]
+    raw_shape = results[ssd_constants.RAW_SHAPE]
+
+    # COCO evaluation requires processing COCO_NUM_VAL_IMAGES exactly once. Due
+    # to rounding errors (i.e., COCO_NUM_VAL_IMAGES % batch_size != 0), setting
+    # `num_eval_epochs` to 1 is not enough and will often miss some images. We
+    # expect user to set `num_eval_epochs` to >1, which will leave some unused
+    # images from previous steps in `predictions`. Here we check if we are doing
+    # eval at a new global step.
+    if results['global_step'] > self.eval_global_step:
+      self.eval_global_step = results['global_step']
+      self.predictions.clear()
+
+    for i, sid in enumerate(source_id):
+      self.predictions[int(sid)] = {
+          ssd_constants.PRED_BOXES: pred_boxes[i],
+          ssd_constants.PRED_SCORES: pred_scores[i],
+          ssd_constants.SOURCE_ID: source_id[i],
+          ssd_constants.RAW_SHAPE: raw_shape[i]
+      }
+
+    # COCO metric calculates mAP only after a full epoch of evaluation. Return
+    # dummy results for top_N_accuracy to be compatible with benchmar_cnn.py.
+    if len(self.predictions) >= ssd_constants.COCO_NUM_VAL_IMAGES:
+      log_fn('Got results for all {:d} eval examples. Calculate mAP...'.format(
+          ssd_constants.COCO_NUM_VAL_IMAGES))
+
+      annotation_file = os.path.join(self.params.data_dir,
+                                     ssd_constants.ANNOTATION_FILE)
+      # Size of predictions before decoding about 15--30GB, while size after
+      # decoding is 100--200MB. When using async eval mode, decoding takes
+      # 20--30 seconds of main thread time but is necessary to avoid OOM during
+      # inter-process communication.
+      decoded_preds = coco_metric.decode_predictions(self.predictions.values())
+      self.predictions.clear()
+
+      if self.params.collect_eval_results_async:
+        def _eval_results_getter():
+          """Iteratively get eval results from async eval process."""
+          while True:
+            step, eval_results = self.async_eval_results_queue.get()
+            self.eval_coco_ap = eval_results['COCO/AP']
+            mlperf.logger.log_eval_accuracy(
+                self.eval_coco_ap, step, self.batch_size * self.params.num_gpus,
+                ssd_constants.COCO_NUM_TRAIN_IMAGES)
+            if self.reached_target():
+              # Reached target, clear all pending messages in predictions queue
+              # and insert poison pill to stop the async eval process.
+              while not self.async_eval_predictions_queue.empty():
+                self.async_eval_predictions_queue.get()
+              self.async_eval_predictions_queue.put('STOP')
+              break
+
+        if not self.async_eval_process:
+          # Limiting the number of messages in predictions queue to prevent OOM.
+          # Each message (predictions data) can potentially consume a lot of
+          # memory, and normally there should only be few messages in the queue.
+          # If often blocked on this, consider reducing eval frequency.
+          self.async_eval_predictions_queue = multiprocessing.Queue(2)
+          self.async_eval_results_queue = multiprocessing.Queue()
+
+          # Reason to use a Process as opposed to Thread is mainly the
+          # computationally intensive eval runner. Python multithreading is not
+          # truly running in parallel, a runner thread would get significantly
+          # delayed (or alternatively delay the main thread).
+          self.async_eval_process = multiprocessing.Process(
+              target=coco_metric.async_eval_runner,
+              args=(self.async_eval_predictions_queue,
+                    self.async_eval_results_queue,
+                    annotation_file))
+          self.async_eval_process.daemon = True
+          self.async_eval_process.start()
+
+          self.async_eval_results_getter_thread = threading.Thread(
+              target=_eval_results_getter, args=())
+          self.async_eval_results_getter_thread.daemon = True
+          self.async_eval_results_getter_thread.start()
+
+        self.async_eval_predictions_queue.put(
+            (self.eval_global_step, decoded_preds))
+        return {'top_1_accuracy': 0, 'top_5_accuracy': 0.}
+
+      eval_results = coco_metric.compute_map(decoded_preds, annotation_file)
+      self.eval_coco_ap = eval_results['COCO/AP']
+      ret = {'top_1_accuracy': self.eval_coco_ap, 'top_5_accuracy': 0.}
+      for metric_key, metric_value in eval_results.items():
+        ret[constants.SIMPLE_VALUE_RESULT_PREFIX + metric_key] = metric_value
+      mlperf.logger.log_eval_accuracy(self.eval_coco_ap, self.eval_global_step,
+                                      self.batch_size * self.params.num_gpus,
+                                      ssd_constants.COCO_NUM_TRAIN_IMAGES)
+      return ret
+    log_fn('Got {:d} out of {:d} eval examples.'
+           ' Waiting for the remaining to calculate mAP...'.format(
+               len(self.predictions), ssd_constants.COCO_NUM_VAL_IMAGES))
+    return {'top_1_accuracy': self.eval_coco_ap, 'top_5_accuracy': 0.}
+
+  def get_synthetic_inputs(self, input_name, nclass):
+    """Generating synthetic data matching real data shape and type."""
+    inputs = tf.random_uniform(
+        self.get_input_shapes('train')[0], dtype=self.data_type)
+    inputs = variables.VariableV1(inputs, trainable=False,
+                                  collections=[tf.GraphKeys.LOCAL_VARIABLES],
+                                  name=input_name)
+    boxes = tf.random_uniform(
+        [self.batch_size, ssd_constants.NUM_SSD_BOXES, 4], dtype=tf.float32)
+    classes = tf.random_uniform(
+        [self.batch_size, ssd_constants.NUM_SSD_BOXES, 1], dtype=tf.float32)
+    nboxes = tf.random_uniform(
+        [self.batch_size], minval=1, maxval=10, dtype=tf.float32)
+    return (inputs, boxes, classes, nboxes)
+
+  def reached_target(self):
+    return (self.params.stop_at_top_1_accuracy and
+            self.eval_coco_ap >= self.params.stop_at_top_1_accuracy)