Python tensorflow.compat.v1.int32() Examples

def testMultilabelMatch3(self):
    predictions = np.random.randint(1, 5, size=(100, 1, 1, 1))
    targets = np.random.randint(1, 5, size=(100, 10, 1, 1))
    weights = np.random.randint(0, 2, size=(100, 1, 1, 1))
    targets *= weights

    predictions_repeat = np.repeat(predictions, 10, axis=1)
    expected = (predictions_repeat == targets).astype(float)
    expected = np.sum(expected, axis=(1, 2, 3))
    expected = np.minimum(expected / 3.0, 1.)
    expected = np.sum(expected * weights[:, 0, 0, 0]) / weights.shape[0]
    with self.test_session() as session:
      scores, weights_ = metrics.multilabel_accuracy_match3(
          tf.one_hot(predictions, depth=5, dtype=tf.float32),
          tf.constant(targets, dtype=tf.int32))
      a, a_op = tf.metrics.mean(scores, weights_)
      session.run(tf.local_variables_initializer())
      session.run(tf.global_variables_initializer())
      _ = session.run(a_op)
      actual = session.run(a)
    self.assertAlmostEqual(actual, expected, places=6) 

def testLSTMSeq2seqBidirectionalEncoder(self):
    vocab_size = 9
    x = np.random.randint(1, high=vocab_size, size=(3, 5, 1, 1))
    y = np.random.randint(1, high=vocab_size, size=(3, 6, 1, 1))
    hparams = lstm.lstm_seq2seq()
    p_hparams = problem_hparams.test_problem_hparams(vocab_size,
                                                     vocab_size,
                                                     hparams)
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(x, dtype=tf.int32),
          "targets": tf.constant(y, dtype=tf.int32),
      }
      model = lstm.LSTMSeq2seqBidirectionalEncoder(
          hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (3, 6, 1, 1, vocab_size)) 

def mask_from_lengths(lengths, max_length=None, dtype=None, name=None):
  """Convert a length scalar to a vector of binary masks.

  This function will convert a vector of lengths to a matrix of binary masks.
  E.g. [2, 4, 3] will become [[1, 1, 0, 0], [1, 1, 1, 1], [1, 1, 1, 0]]

  Args:
    lengths: a d-dimensional vector of integers corresponding to lengths.
    max_length: an optional (default: None) scalar-like or 0-dimensional tensor
      indicating the maximum length of the masks. If not provided, the maximum
      length will be inferred from the lengths vector.
    dtype: the dtype of the returned mask, if specified. If None, the dtype of
      the lengths will be used.
    name: a name for the operation (optional).

  Returns:
    A d x max_length tensor of binary masks (int32).
  """
  with tf.name_scope(name, 'mask_from_lengths'):
    dtype = lengths.dtype if dtype is None else dtype
    max_length = tf.reduce_max(lengths) if max_length is None else max_length
    indexes = tf.range(max_length, dtype=lengths.dtype)
    mask = tf.less(tf.expand_dims(indexes, 0), tf.expand_dims(lengths, 1))
    cast_mask = tf.cast(mask, dtype)
  return tf.stop_gradient(cast_mask) 

def build(self, input_shape):
        with self._sess.graph.as_default():
            self._placeholders["tokens"] = tf.placeholder(
                dtype=tf.int32, shape=[None, None], name="tokens"
            )

            self._ops["output_logits"] = self.compute_logits(
                self._placeholders["tokens"]
            )
            self._ops["output_probs"] = tf.nn.softmax(self._ops["output_logits"], -1)
            result = self.compute_loss_and_acc(
                rnn_output_logits=self._ops["output_logits"],
                target_token_seq=self._placeholders["tokens"],
            )
            self._ops["loss"] = result.token_ce_loss
            self._ops["num_tokens"] = result.num_predictions
            self._ops["num_correct_tokens"] = result.num_correct_token_predictions
            self._ops["train_step"] = self._make_training_step(self._ops["loss"])

            init_op = tf.variables_initializer(
                self._sess.graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
            )
            self._sess.run(init_op) 

def testLSTMSeq2SeqAttention(self):
    vocab_size = 9
    x = np.random.randint(1, high=vocab_size, size=(3, 5, 1, 1))
    y = np.random.randint(1, high=vocab_size, size=(3, 6, 1, 1))
    hparams = lstm.lstm_attention()

    p_hparams = problem_hparams.test_problem_hparams(vocab_size,
                                                     vocab_size,
                                                     hparams)
    x = tf.constant(x, dtype=tf.int32)
    x = tf.placeholder_with_default(x, shape=[None, None, 1, 1])

    with self.test_session() as session:
      features = {
          "inputs": x,
          "targets": tf.constant(y, dtype=tf.int32),
      }
      model = lstm.LSTMSeq2seqAttention(
          hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (3, 6, 1, 1, vocab_size)) 

def testLSTMSeq2seqAttentionBidirectionalEncoder(self):
    vocab_size = 9
    x = np.random.randint(1, high=vocab_size, size=(3, 5, 1, 1))
    y = np.random.randint(1, high=vocab_size, size=(3, 6, 1, 1))
    hparams = lstm.lstm_attention()

    p_hparams = problem_hparams.test_problem_hparams(vocab_size, vocab_size)
    x = tf.constant(x, dtype=tf.int32)
    x = tf.placeholder_with_default(x, shape=[None, None, 1, 1])

    with self.test_session() as session:
      features = {
          "inputs": x,
          "targets": tf.constant(y, dtype=tf.int32),
      }
      model = lstm.LSTMSeq2seqAttentionBidirectionalEncoder(
          hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (3, 6, 1, 1, vocab_size)) 

def _test_resnet(self, img_size, output_size):
    vocab_size = 9
    batch_size = 2
    x = np.random.randint(
        256, size=(batch_size, img_size, img_size, 3))
    y = np.random.randint(
        1, high=vocab_size, size=(batch_size, 1, 1, 1))
    hparams = resnet_tiny_cpu()
    p_hparams = problem_hparams.test_problem_hparams(vocab_size,
                                                     vocab_size,
                                                     hparams)
    p_hparams.modality["inputs"] = modalities.ModalityType.IMAGE
    p_hparams.modality["targets"] = modalities.ModalityType.CLASS_LABEL
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(x, dtype=tf.int32),
          "targets": tf.constant(y, dtype=tf.int32),
      }
      model = resnet.Resnet(hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (batch_size,) + output_size + (1, vocab_size)) 

def _create_slots(self, var_list):
    """Create slot variables for Adam with accumulated gradients."""
    first_var = min(var_list, key=lambda x: x.name)
    self._create_non_slot_variable(
        initial_value=self._beta1, name="beta1_power", colocate_with=first_var)
    self._create_non_slot_variable(
        initial_value=self._beta2, name="beta2_power", colocate_with=first_var)
    # if iter is initialized as an int32, this optimizer could not run
    # with tensorflow_hub with a tensorflow-gpu version
    self._create_non_slot_variable(
        initial_value=0.0 if self._n == 1 else 1.0,
        name="iter",
        colocate_with=first_var)
    # Create slots for the first and second moments, as well as grad_acc.
    for v in var_list:
      self._zeros_slot(v, "m", self._name)
      self._zeros_slot(v, "v", self._name)
      self._zeros_slot(v, "grad_acc", self._name) 

def _test_xception(self, img_size):
    vocab_size = 9
    batch_size = 3
    x = np.random.randint(
        256, size=(batch_size, img_size, img_size, 3))
    y = np.random.randint(
        1, high=vocab_size, size=(batch_size, 1, 1, 1))
    hparams = xception.xception_tiny()
    p_hparams = problem_hparams.test_problem_hparams(vocab_size,
                                                     vocab_size,
                                                     hparams)
    p_hparams.modality["inputs"] = modalities.ModalityType.IMAGE
    p_hparams.modality["targets"] = modalities.ModalityType.CLASS_LABEL
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(x, dtype=tf.int32),
          "targets": tf.constant(y, dtype=tf.int32),
      }
      model = xception.Xception(hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (batch_size, 1, 1, 1, vocab_size)) 

def get_synthetic_inputs(self, input_name, nclass):
    # Synthetic input should be within [0, 255].
    image_shape, label_shape = self.get_input_shapes('train')
    inputs = tf.truncated_normal(
        image_shape,
        dtype=self.data_type,
        mean=127,
        stddev=60,
        name=self.model_name + '_synthetic_inputs')
    inputs = variables_module.VariableV1(
        inputs, trainable=False, collections=[tf.GraphKeys.LOCAL_VARIABLES],
        name=input_name)
    labels = tf.random_uniform(
        label_shape,
        minval=0,
        maxval=nclass - 1,
        dtype=tf.int32,
        name=self.model_name + '_synthetic_labels')
    return (inputs, labels) 

def testSliceNet(self):
    x = np.random.randint(256, size=(3, 5, 5, 3))
    y = np.random.randint(10, size=(3, 5, 1, 1))
    hparams = slicenet.slicenet_params1_tiny()
    hparams.add_hparam("data_dir", "")
    problem = registry.problem("image_cifar10")
    p_hparams = problem.get_hparams(hparams)
    hparams.problem_hparams = p_hparams
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(x, dtype=tf.int32),
          "targets": tf.constant(y, dtype=tf.int32),
          "target_space_id": tf.constant(1, dtype=tf.int32),
      }
      model = slicenet.SliceNet(hparams, tf.estimator.ModeKeys.TRAIN,
                                p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (3, 1, 1, 1, 10)) 

def testSliceNetImageToText(self):
    x = np.random.randint(256, size=(3, 5, 5, 3))
    y = np.random.randint(10, size=(3, 5, 1, 1))
    hparams = slicenet.slicenet_params1_tiny()
    hparams.add_hparam("data_dir", "")
    problem = registry.problem("image_ms_coco_characters")
    p_hparams = problem.get_hparams(hparams)
    hparams.problem_hparams = p_hparams
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(x, dtype=tf.int32),
          "targets": tf.constant(y, dtype=tf.int32),
          "target_space_id": tf.constant(1, dtype=tf.int32),
      }
      model = slicenet.SliceNet(hparams, tf.estimator.ModeKeys.TRAIN,
                                p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (3, 5, 1, 1, 258)) 

def get_synthetic_inputs(self, input_name, nclass):
    """Returns the ops to generate synthetic inputs and labels."""
    def users_init_val():
      return tf.random_uniform((self.batch_size, 1), minval=0,
                               maxval=_NUM_USERS_20M, dtype=tf.int32)
    users = tf.Variable(users_init_val, dtype=tf.int32, trainable=False,
                        collections=[tf.GraphKeys.LOCAL_VARIABLES],
                        name='synthetic_users')
    def items_init_val():
      return tf.random_uniform((self.batch_size, 1), minval=0,
                               maxval=_NUM_ITEMS_20M, dtype=tf.int32)
    items = tf.Variable(items_init_val, dtype=tf.int32, trainable=False,
                        collections=[tf.GraphKeys.LOCAL_VARIABLES],
                        name='synthetic_items')

    def labels_init_val():
      return tf.random_uniform((self.batch_size,), minval=0, maxval=2,
                               dtype=tf.int32)
    labels = tf.Variable(labels_init_val, dtype=tf.int32, trainable=False,
                         collections=[tf.GraphKeys.LOCAL_VARIABLES],
                         name='synthetic_labels')

    return [users, items, labels] 

def compute_logits(self, token_ids: tf.Tensor) -> tf.Tensor:
        """
        Implements a language model, where each output is conditional on the current
        input and inputs processed so far.

        Args:
            token_ids: int32 tensor of shape [B, T], storing integer IDs of tokens.

        Returns:
            tf.float32 tensor of shape [B, T, V], storing the distribution over output symbols
            for each timestep for each batch element.
        """
        # TODO 5# 1) Embed tokens
        # TODO 5# 2) Run RNN on embedded tokens
        # TODO 5# 3) Project RNN outputs onto the vocabulary to obtain logits.
        return rnn_output_logits 

def __init__(
      self, batch_size, observation_space, action_space, policy_hparams,
      policy_dir, sampling_temp
  ):
    super(PolicyAgent, self).__init__(
        batch_size, observation_space, action_space
    )
    self._sampling_temp = sampling_temp
    with tf.Graph().as_default():
      self._observations_t = tf.placeholder(
          shape=((batch_size,) + self.observation_space.shape),
          dtype=self.observation_space.dtype
      )
      (logits, self._values_t) = rl.get_policy(
          self._observations_t, policy_hparams, self.action_space
      )
      actions = common_layers.sample_with_temperature(logits, sampling_temp)
      self._probs_t = tf.nn.softmax(logits / sampling_temp)
      self._actions_t = tf.cast(actions, tf.int32)
      model_saver = tf.train.Saver(
          tf.global_variables(policy_hparams.policy_network + "/.*")  # pylint: disable=unexpected-keyword-arg
      )
      self._sess = tf.Session()
      self._sess.run(tf.global_variables_initializer())
      trainer_lib.restore_checkpoint(policy_dir, model_saver, self._sess) 

def _test_model(self, hparams, model_cls):
    batch_size = 3
    target_length = 6
    target_out = 10  # GeneExpressionProblem.num_output_predictions
    input_length = target_length * 128 // 4  # chunk_size=4
    input_vocab_size = 5

    inputs = np.random.randint(
        1, input_vocab_size + 1, size=(batch_size, input_length, 1, 1))
    targets = np.random.random_sample((batch_size, target_length, 1,
                                       target_out))

    features = {
        "inputs": tf.constant(inputs, dtype=tf.int32),
        "targets": tf.constant(targets, dtype=tf.float32),
    }
    p_hparams = hparams.problem_hparams
    logits, _ = model_cls(
        hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)(features)

    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      res = sess.run(logits)

    self.assertEqual(res.shape, (batch_size, target_length, 1, target_out)) 

def __init__(self, *args, **kwargs):
    with tf.Graph().as_default():
      self._batch_env = SimulatedBatchEnv(*args, **kwargs)

      self._actions_t = tf.placeholder(shape=(self.batch_size,), dtype=tf.int32)
      self._rewards_t, self._dones_t = self._batch_env.simulate(self._actions_t)
      with tf.control_dependencies([self._rewards_t]):
        self._obs_t = self._batch_env.observ
      self._indices_t = tf.placeholder(shape=(self.batch_size,), dtype=tf.int32)
      self._reset_op = self._batch_env.reset(
          tf.range(self.batch_size, dtype=tf.int32)
      )

      self._sess = tf.Session()
      self._sess.run(tf.global_variables_initializer())
      self._batch_env.initialize(self._sess) 

def _test_imagetransformer_2d(self, net):
    batch_size = 3
    size = 7
    vocab_size = 256
    hparams = image_transformer_2d.imagetransformer2d_tiny()
    p_hparams = problem_hparams.test_problem_hparams(vocab_size,
                                                     vocab_size,
                                                     hparams)
    inputs = np.random.randint(
        vocab_size, size=(batch_size, 1, 1, 1))
    targets = np.random.randint(
        vocab_size, size=(batch_size, size, size, 3))
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(inputs, dtype=tf.int32),
          "targets": tf.constant(targets, dtype=tf.int32),
          "target_space_id": tf.constant(1, dtype=tf.int32),
      }
      model = net(hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (batch_size, size, size, 3, vocab_size)) 

def _test_img2img_transformer(self, net):
    batch_size = 3
    hparams = image_transformer_2d.img2img_transformer2d_tiny()
    hparams.data_dir = ""
    p_hparams = registry.problem("image_celeba").get_hparams(hparams)
    inputs = np.random.randint(256, size=(batch_size, 4, 4, 3))
    targets = np.random.randint(256, size=(batch_size, 8, 8, 3))
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(inputs, dtype=tf.int32),
          "targets": tf.constant(targets, dtype=tf.int32),
          "target_space_id": tf.constant(1, dtype=tf.int32),
      }
      model = net(hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (batch_size, 8, 8, 3, 256)) 

def testByteNet(self):
    vocab_size = 9
    x = np.random.randint(1, high=vocab_size, size=(3, 5, 1, 1))
    y = np.random.randint(1, high=vocab_size, size=(3, 6, 1, 1))
    hparams = bytenet.bytenet_base()
    p_hparams = problem_hparams.test_problem_hparams(vocab_size,
                                                     vocab_size,
                                                     hparams)
    with self.test_session() as session:
      features = {
          "inputs": tf.constant(x, dtype=tf.int32),
          "targets": tf.constant(y, dtype=tf.int32),
      }
      model = bytenet.ByteNet(
          hparams, tf.estimator.ModeKeys.TRAIN, p_hparams)
      logits, _ = model(features)
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    self.assertEqual(res.shape, (3, 50, 1, 1, vocab_size)) 

def shuffle_layer(inputs, shuffle_fn=rol):
  """Shuffles the elements according to bitwise left or right rotation.

  Args:
    inputs: Tensor input from previous layer
    shuffle_fn: Shift function rol or ror

  Returns:
    tf.Tensor: Inputs shifted according to shuffle_fn
  """

  length = tf.shape(inputs)[1]
  n_bits = tf.log(tf.cast(length - 1, tf.float32)) / tf.log(2.0)
  n_bits = tf.cast(n_bits, tf.int32) + 1

  indices = tf.range(0, length)
  rev_indices = shuffle_fn(indices, n_bits)
  return tf.gather(inputs, rev_indices, axis=1) 

def testAccuracyTopKMetric(self):
    predictions = np.random.randint(1, 5, size=(12, 12, 12, 1))
    targets = np.random.randint(1, 5, size=(12, 12, 12, 1))
    expected = np.mean((predictions == targets).astype(float))
    with self.test_session() as session:
      predicted = tf.one_hot(predictions, depth=5, dtype=tf.float32)
      scores1, _ = metrics.padded_accuracy_topk(
          predicted, tf.constant(targets, dtype=tf.int32), k=1)
      scores2, _ = metrics.padded_accuracy_topk(
          predicted, tf.constant(targets, dtype=tf.int32), k=7)
      a1 = tf.reduce_mean(scores1)
      a2 = tf.reduce_mean(scores2)
      session.run(tf.global_variables_initializer())
      actual1, actual2 = session.run([a1, a2])
    self.assertAlmostEqual(actual1, expected)
    self.assertAlmostEqual(actual2, 1.0) 

def pad_batch(features, batch_multiple):
  """Pad batch dim of features to nearest multiple of batch_multiple."""
  feature = list(features.items())[0][1]
  batch_size = tf.shape(feature)[0]
  mod = batch_size % batch_multiple
  has_mod = tf.cast(tf.cast(mod, tf.bool), tf.int32)
  batch_padding = batch_multiple * has_mod - mod

  padded_features = {}
  for k, feature in features.items():
    rank = len(feature.shape)
    paddings = [[0, 0] for _ in range(rank)]
    paddings[0][1] = batch_padding
    padded_feature = tf.pad(feature, paddings)
    padded_features[k] = padded_feature
  return padded_features


# TODO(lukaszkaiser): refactor the API to not be just a list of self params
#   but make sense for other uses too. 

def get_mnist_random_output(self, model_name, hparams_set=None,
                              mode=tf.estimator.ModeKeys.TRAIN):
    hparams_set = hparams_set or model_name
    x = np.random.randint(256, size=(1, 28, 28, 1))
    y = np.random.randint(10, size=(1, 1))
    features = {
        "targets": tf.constant(x, dtype=tf.int32),
        "inputs": tf.constant(y, dtype=tf.int32),
    }
    hparams = trainer_lib.create_hparams(
        hparams_set, problem_name="image_mnist_rev", data_dir=".")
    model = registry.model(model_name)(hparams, mode)
    tf.train.create_global_step()
    logits, _ = model(features)
    with self.test_session() as session:
      session.run(tf.global_variables_initializer())
      res = session.run(logits)
    return res 

def residual_shuffle_network(inputs, hparams):
  """Residual Shuffle-Exchange network with weight sharing.

  Args:
    inputs: inputs to the Shuffle-Exchange network. Should be in length of power
      of 2.
    hparams: Model configuration

  Returns:
    tf.Tensor: Outputs of the Shuffle-Exchange last layer
  """
  input_shape = tf.shape(inputs)
  n_bits = tf.log(tf.cast(input_shape[1] - 1, tf.float32)) / tf.log(2.0)
  n_bits = tf.cast(n_bits, tf.int32) + 1

  block_out = inputs

  for k in range(hparams.num_hidden_layers):
    with tf.variable_scope("benes_block_" + str(k), reuse=tf.AUTO_REUSE):
      forward_output = forward_part(block_out, hparams, n_bits)
      block_out = reverse_part(forward_output, hparams, n_bits)

  return RSU("last_layer", hparams.dropout, hparams.mode)(block_out) 

def testSequenceEditDistanceMetric(self):
    predictions = np.array([[3, 4, 5, 1, 0, 0],
                            [2, 1, 3, 4, 0, 0],
                            [2, 1, 3, 4, 0, 0]])
    # Targets are just a bit different:
    #  - first sequence has a different prediction
    #  - second sequence has a different prediction and one extra step
    #  - third sequence is identical
    targets = np.array([[5, 4, 5, 1, 0, 0],
                        [2, 5, 3, 4, 1, 0],
                        [2, 1, 3, 4, 0, 0]])
    # Reshape to match expected input format by metric fns.
    predictions = np.reshape(predictions, [3, 6, 1, 1])
    targets = np.reshape(targets, [3, 6, 1, 1])
    with self.test_session() as session:
      scores, weight = metrics.sequence_edit_distance(
          tf.one_hot(predictions, depth=6, dtype=tf.float32),
          tf.constant(targets, dtype=tf.int32))
      session.run(tf.global_variables_initializer())
      actual_scores, actual_weight = session.run([scores, weight])
    self.assertAlmostEqual(actual_scores, 3.0 / 13)
    self.assertEqual(actual_weight, 13) 

def testRougeLMetricE2E(self):
    vocab_size = 4
    batch_size = 12
    seq_length = 12
    predictions = tf.one_hot(
        np.random.randint(vocab_size, size=(batch_size, seq_length, 1, 1)),
        depth=4,
        dtype=tf.float32)
    targets = np.random.randint(4, size=(12, 12, 1, 1))
    with self.test_session() as session:
      scores, _ = rouge.rouge_l_fscore(
          predictions,
          tf.constant(targets, dtype=tf.int32))
      a = tf.reduce_mean(scores)
      session.run(tf.global_variables_initializer())
      session.run(a) 

def testNegativeLogPerplexityMaskedAssert(self):
    predictions = np.random.randint(4, size=(12, 12, 12, 1))
    targets = np.random.randint(4, size=(12, 12, 12, 1))
    features = {}

    with self.assertRaisesRegexp(
        ValueError,
        'masked_neg_log_perplexity requires targets_mask feature'):
      with self.test_session() as session:
        scores, _ = metrics.padded_neg_log_perplexity_with_masking(
            tf.one_hot(predictions, depth=4, dtype=tf.float32),
            tf.constant(targets, dtype=tf.int32),
            features)
        a = tf.reduce_mean(scores)
        session.run(tf.global_variables_initializer())
        _ = session.run(a) 

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 max_pad_length(self, features):
    """Finds max padding length.

    If target length not specified use fixed padding
    length from hparams.max_length.

    Args:
      features: Dictionary with input and target tensors

    Returns:
      tf.Tensor:  Length of input and output sequence. Length is power of 2.
    """

    if self.hparams.force_max_length or features.get("targets") is None:
      assert math.log(self.hparams.max_length, 2).is_integer(), \
        "hparams.max_length should be power of w"

      return self.hparams.max_length

    length = tf.shape(features["inputs"])[1]
    targets_length = tf.shape(features["targets"])[1]
    length = tf.maximum(length, targets_length)

    p = tf.log(tf.cast(length, tf.float32)) / tf.log(2.0)
    p = tf.cast(tf.ceil(p), tf.int32)
    return tf.pow(2, p)