Python tensorflow.one_hot() Examples

def _build_qnet(self):
    """
    Build q-network
    """
    with tf.variable_scope(self.scope):
      self.state_input = tf.placeholder(tf.float32, [None, self.state_size])
      self.action = tf.placeholder(tf.int32, [None])
      self.target_q = tf.placeholder(tf.float32, [None])

      fc1 = tf_utils.fc(self.state_input, n_output=self.n_hidden_1, activation_fn=tf.nn.relu)
      fc2 = tf_utils.fc(fc1, n_output=self.n_hidden_2, activation_fn=tf.nn.relu)
      self.q_values = tf_utils.fc(fc2, self.action_size, activation_fn=None)

      action_mask = tf.one_hot(self.action, self.action_size, 1.0, 0.0)
      q_value_pred = tf.reduce_sum(self.q_values * action_mask, 1)

      self.loss = tf.reduce_mean(tf.square(tf.subtract(self.target_q, q_value_pred)))
      self.optimizer = tf.train.AdamOptimizer(self.lr)
      self.train_op = self.optimizer.minimize(self.loss, global_step=tf.contrib.framework.get_global_step()) 

def sigmoid_precision_one_hot(logits, labels, weights_fn=None):
  """Calculate precision for a set, given one-hot labels and logits.

  Predictions are converted to one-hot,
  as predictions[example][arg-max(example)] = 1

  Args:
    logits: Tensor of size [batch-size, o=1, p=1, num-classes]
    labels: Tensor of size [batch-size, o=1, p=1, num-classes]
    weights_fn: Function that takes in labels and weighs examples (unused)
  Returns:
    precision (scalar), weights
  """
  with tf.variable_scope("sigmoid_precision_one_hot", values=[logits, labels]):
    del weights_fn
    num_classes = logits.shape[-1]
    predictions = tf.nn.sigmoid(logits)
    predictions = tf.argmax(predictions, -1)
    predictions = tf.one_hot(predictions, num_classes)
    _, precision = tf.metrics.precision(labels=labels, predictions=predictions)
    return precision, tf.constant(1.0) 

def set_precision(predictions, labels,
                  weights_fn=common_layers.weights_nonzero):
  """Precision of set predictions.

  Args:
    predictions : A Tensor of scores of shape [batch, nlabels].
    labels: A Tensor of int32s giving true set elements,
      of shape [batch, seq_length].
    weights_fn: A function to weight the elements.

  Returns:
    hits: A Tensor of shape [batch, nlabels].
    weights: A Tensor of shape [batch, nlabels].
  """
  with tf.variable_scope("set_precision", values=[predictions, labels]):
    labels = tf.squeeze(labels, [2, 3])
    weights = weights_fn(labels)
    labels = tf.one_hot(labels, predictions.shape[-1])
    labels = tf.reduce_max(labels, axis=1)
    labels = tf.cast(labels, tf.bool)
    return tf.to_float(tf.equal(labels, predictions)), weights 

def _testBuildDefaultModel(self):
    images = tf.to_float(np.random.rand(32, 28, 28, 1))
    labels = {}
    labels['classes'] = tf.one_hot(
        tf.to_int32(np.random.randint(0, 9, (32))), 10)

    params = {
        'use_separation': True,
        'layers_to_regularize': 'fc3',
        'weight_decay': 0.0,
        'ps_tasks': 1,
        'domain_separation_startpoint': 1,
        'alpha_weight': 1,
        'beta_weight': 1,
        'gamma_weight': 1,
        'recon_loss_name': 'sum_of_squares',
        'decoder_name': 'small_decoder',
        'encoder_name': 'default_encoder',
    }
    return images, labels, params 

def log_prob_action(self, action, logits,
                      sampling_dim, act_dim, act_type):
    """Calculate log-prob of action sampled from distribution."""
    if self.env_spec.is_discrete(act_type):
      act_log_prob = tf.reduce_sum(
          tf.one_hot(action, act_dim) * tf.nn.log_softmax(logits), -1)
    elif self.env_spec.is_box(act_type):
      means = logits[:, :sampling_dim / 2]
      std = logits[:, sampling_dim / 2:]
      act_log_prob = (- 0.5 * tf.log(2 * np.pi * tf.square(std))
                      - 0.5 * tf.square(action - means) / tf.square(std))
      act_log_prob = tf.reduce_sum(act_log_prob, -1)
    else:
      assert False

    return act_log_prob 

def one_hot_encoding(labels, num_classes=None):
  """One-hot encodes the multiclass labels.

  Example usage:
    labels = tf.constant([1, 4], dtype=tf.int32)
    one_hot = OneHotEncoding(labels, num_classes=5)
    one_hot.eval()    # evaluates to [0, 1, 0, 0, 1]

  Args:
    labels: A tensor of shape [None] corresponding to the labels.
    num_classes: Number of classes in the dataset.
  Returns:
    onehot_labels: a tensor of shape [num_classes] corresponding to the one hot
      encoding of the labels.
  Raises:
    ValueError: if num_classes is not specified.
  """
  with tf.name_scope('OneHotEncoding', values=[labels]):
    if num_classes is None:
      raise ValueError('num_classes must be specified')

    labels = tf.one_hot(labels, num_classes, 1, 0)
    return tf.reduce_max(labels, 0) 

def visit_count_fc(visit_count, last_visit, embed_neurons, wt_decay, fc_dropout):
  with tf.variable_scope('embed_visit_count'):
    visit_count = tf.reshape(visit_count, shape=[-1])
    last_visit = tf.reshape(last_visit, shape=[-1])
    
    visit_count = tf.clip_by_value(visit_count, clip_value_min=-1,
                                   clip_value_max=15)
    last_visit = tf.clip_by_value(last_visit, clip_value_min=-1,
                                   clip_value_max=15)
    visit_count = tf.one_hot(visit_count, depth=16, axis=1, dtype=tf.float32,
                             on_value=10., off_value=0.)
    last_visit = tf.one_hot(last_visit, depth=16, axis=1, dtype=tf.float32,
                             on_value=10., off_value=0.)
    f = tf.concat([visit_count, last_visit], 1)
    x, _ = tf_utils.fc_network(
        f, neurons=embed_neurons, wt_decay=wt_decay, name='visit_count_embed',
        offset=0, batch_norm_param=None, dropout_ratio=fc_dropout,
        is_training=is_training)
  return x 

def label_smoothing_regularization(self, chars_labels, weight=0.1):
    """Applies a label smoothing regularization.

    Uses the same method as in https://arxiv.org/abs/1512.00567.

    Args:
      chars_labels: ground truth ids of charactes,
        shape=[batch_size, seq_length];
      weight: label-smoothing regularization weight.

    Returns:
      A sensor with the same shape as the input.
    """
    one_hot_labels = tf.one_hot(
        chars_labels, depth=self._params.num_char_classes, axis=-1)
    pos_weight = 1.0 - weight
    neg_weight = weight / self._params.num_char_classes
    return one_hot_labels * pos_weight + neg_weight 

def _build_input(self):
        self.tails = tf.placeholder(tf.int32, [None])
        self.heads = tf.placeholder(tf.int32, [None])
        self.targets = tf.one_hot(indices=self.heads, depth=self.num_entity)
            
        if not self.query_is_language:
            self.queries = tf.placeholder(tf.int32, [None, self.num_step])
            self.query_embedding_params = tf.Variable(self._random_uniform_unit(
                                                          self.num_query + 1, # <END> token 
                                                          self.query_embed_size), 
                                                      dtype=tf.float32)
        
            rnn_inputs = tf.nn.embedding_lookup(self.query_embedding_params, 
                                                self.queries)
        else:
            self.queries = tf.placeholder(tf.int32, [None, self.num_step, self.num_word])
            self.vocab_embedding_params = tf.Variable(self._random_uniform_unit(
                                                          self.num_vocab + 1, # <END> token
                                                          self.vocab_embed_size),
                                                      dtype=tf.float32)
            embedded_query = tf.nn.embedding_lookup(self.vocab_embedding_params, 
                                                    self.queries)
            rnn_inputs = tf.reduce_mean(embedded_query, axis=2)

        return rnn_inputs 

def set_recall(predictions, labels, weights_fn=common_layers.weights_nonzero):
  """Recall of set predictions.

  Args:
    predictions : A Tensor of scores of shape [batch, nlabels].
    labels: A Tensor of int32s giving true set elements,
      of shape [batch, seq_length].
    weights_fn: A function to weight the elements.

  Returns:
    hits: A Tensor of shape [batch, nlabels].
    weights: A Tensor of shape [batch, nlabels].
  """
  with tf.variable_scope("set_recall", values=[predictions, labels]):
    labels = tf.squeeze(labels, [2, 3])
    weights = weights_fn(labels)
    labels = tf.one_hot(labels, predictions.shape[-1])
    labels = tf.reduce_max(labels, axis=1)
    labels = tf.cast(labels, tf.bool)
    return tf.to_float(tf.equal(labels, predictions)), weights 

def sigmoid_recall_one_hot(logits, labels, weights_fn=None):
  """Calculate recall for a set, given one-hot labels and logits.

  Predictions are converted to one-hot,
  as predictions[example][arg-max(example)] = 1

  Args:
    logits: Tensor of size [batch-size, o=1, p=1, num-classes]
    labels: Tensor of size [batch-size, o=1, p=1, num-classes]
    weights_fn: Function that takes in labels and weighs examples (unused)
  Returns:
    recall (scalar), weights
  """
  with tf.variable_scope("sigmoid_recall_one_hot", values=[logits, labels]):
    del weights_fn
    num_classes = logits.shape[-1]
    predictions = tf.nn.sigmoid(logits)
    predictions = tf.argmax(predictions, -1)
    predictions = tf.one_hot(predictions, num_classes)
    _, recall = tf.metrics.recall(labels=labels, predictions=predictions)
    return recall, tf.constant(1.0) 

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 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 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 vq_nearest_neighbor(x, hparams):
  """Find the nearest element in means to elements in x."""
  bottleneck_size = 2**hparams.bottleneck_bits
  means = hparams.means
  x_norm_sq = tf.reduce_sum(tf.square(x), axis=-1, keepdims=True)
  means_norm_sq = tf.reduce_sum(tf.square(means), axis=-1, keepdims=True)
  scalar_prod = tf.matmul(x, means, transpose_b=True)
  dist = x_norm_sq + tf.transpose(means_norm_sq) - 2 * scalar_prod
  if hparams.bottleneck_kind == "em":
    x_means_idx = tf.multinomial(-dist, num_samples=hparams.num_samples)
    x_means_hot = tf.one_hot(
        x_means_idx, depth=bottleneck_size)
    x_means_hot = tf.reduce_mean(x_means_hot, axis=1)
  else:
    x_means_idx = tf.argmax(-dist, axis=-1)
    x_means_hot = tf.one_hot(x_means_idx, depth=bottleneck_size)
  x_means = tf.matmul(x_means_hot, means)
  e_loss = tf.reduce_mean(tf.square(x - tf.stop_gradient(x_means)))
  return x_means_hot, e_loss 

def fill_memory_slot(memory, value, index):
  """Fills the memory slot at a particular index with the given value.

  Args:
    memory: a 4-d tensor [memory_size, batch, length, channel] containing
      the state of all steps
    value: a 3-d tensor [batch, length, channel] as the sate
    index: integer in [0, memory_size)

  Returns:
    filled memory

  """
  mask = tf.to_float(
      tf.one_hot(index,
                 tf.shape(memory)[0])[:, None, None, None])
  fill_memory = (1 - mask) * memory + mask * value[None, ...]
  return fill_memory 

def targets_bottom(self, x, summary_prefix="targets_bottom"):  # pylint: disable=arguments-differ
    inputs = x
    with tf.variable_scope(self.name, reuse=tf.AUTO_REUSE):
      common_layers.summarize_video(inputs, summary_prefix)
      inputs_shape = common_layers.shape_list(inputs)
      # We embed each of 256=self.top_dimensionality possible pixel values.
      embedding_var = tf.get_variable(
          "pixel_embedding",
          [self.top_dimensionality, self.PIXEL_EMBEDDING_SIZE])
      hot_inputs = tf.one_hot(tf.to_int32(inputs), self.top_dimensionality)
      hot_inputs = tf.reshape(hot_inputs, [-1, self.top_dimensionality])
      embedded = tf.matmul(hot_inputs, embedding_var)
      # Let's now merge all channels that were embedded into a single vector.
      merged_size = self.PIXEL_EMBEDDING_SIZE * inputs_shape[4]
      embedded = tf.reshape(embedded, inputs_shape[:4] + [merged_size])
      transposed = common_layers.time_to_channels(embedded)
      return tf.layers.dense(
          transposed,
          self._body_input_depth,
          name="merge_pixel_embedded_frames") 

def one_hot_encoding(labels, num_classes=None):
  """One-hot encodes the multiclass labels.

  Example usage:
    labels = tf.constant([1, 4], dtype=tf.int32)
    one_hot = OneHotEncoding(labels, num_classes=5)
    one_hot.eval()    # evaluates to [0, 1, 0, 0, 1]

  Args:
    labels: A tensor of shape [None] corresponding to the labels.
    num_classes: Number of classes in the dataset.
  Returns:
    onehot_labels: a tensor of shape [num_classes] corresponding to the one hot
      encoding of the labels.
  Raises:
    ValueError: if num_classes is not specified.
  """
  with tf.name_scope('OneHotEncoding', values=[labels]):
    if num_classes is None:
      raise ValueError('num_classes must be specified')

    labels = tf.one_hot(labels, num_classes, 1, 0)
    return tf.reduce_max(labels, 0) 

def coordinate_tensor(shape, axis):
  """Return a tensor with given shape containing coordinate along given axis.

  Args:
    shape: a Tensor representing the shape of the output Tensor
    axis: an integer

  Returns:
    A tensor with shape shape and type tf.int32, where each elements its
    coordinate along the given axis.
  """
  if axis < 0:
    axis = tf.size(shape) + axis  # Convert to positive for the one_hot indice

  r = tf.range(shape[axis])
  r_shape = tf.one_hot(
      axis, tf.size(shape), on_value=-1, off_value=1, dtype=tf.int32)
  return tf.zeros(shape, dtype=tf.int32) + tf.reshape(r, r_shape) 

def matmul_gather_on_zeroth_axis(params, indices, scope=None):
  """Matrix multiplication based implementation of tf.gather on zeroth axis.

  TODO(rathodv, jonathanhuang): enable sparse matmul option.

  Args:
    params: A float32 Tensor. The tensor from which to gather values.
      Must be at least rank 1.
    indices: A Tensor. Must be one of the following types: int32, int64.
      Must be in range [0, params.shape[0])
    scope: A name for the operation (optional).

  Returns:
    A Tensor. Has the same type as params. Values from params gathered
    from indices given by indices, with shape indices.shape + params.shape[1:].
  """
  with tf.name_scope(scope, 'MatMulGather'):
    params_shape = shape_utils.combined_static_and_dynamic_shape(params)
    indices_shape = shape_utils.combined_static_and_dynamic_shape(indices)
    params2d = tf.reshape(params, [params_shape[0], -1])
    indicator_matrix = tf.one_hot(indices, params_shape[0])
    gathered_result_flattened = tf.matmul(indicator_matrix, params2d)
    return tf.reshape(gathered_result_flattened,
                      tf.stack(indices_shape + params_shape[1:])) 

def iterator(self, is_onehot=True, is_shuffle=False, batch_size=64, buffer_size=512):
        parse_func = lambda example: parse_example(example, shape=self.shape)
        dataset = self.dataset.map(parse_func) # Parse the record into tensors.
        # Only go through the data once with no repeat.
        num_repeats = 1
        if is_shuffle:
            # If training then read a buffer of the given size and randomly shuffle it.
            dataset = dataset.shuffle(buffer_size=buffer_size)
        dataset = dataset.repeat(num_repeats)        # Repeat the input indefinitely.
        if is_onehot:
            onehot_func = lambda feature, label: (feature,
                                                  tf.one_hot(label, self.num_classes))
            dataset = dataset.map(onehot_func)

        dataset = dataset.batch(batch_size)
        iterator = dataset.make_initializable_iterator()
        batch = iterator.get_next()
        return iterator, batch 

def call(self, inputs):
    """Standard Keras call() method."""
    if inputs.dtype not in [tf.uint8, tf.int32, tf.int64]:
      inputs = tf.cast(inputs, dtype=tf.int32)

    if self.default_input_value is not None:
      default_input_value_tensor = tf.constant(
          int(self.default_input_value),
          dtype=inputs.dtype,
          name=DEFAULT_INPUT_VALUE_NAME)
      replacement = tf.zeros_like(inputs) + (self.num_buckets - 1)
      inputs = tf.where(
          tf.equal(inputs, default_input_value_tensor), replacement, inputs)

    # We can't use tf.gather_nd(self.kernel, inputs) as it doesn't support
    # constraints (constraint functions are not supported for IndexedSlices).
    # Instead we use matrix multiplication by one-hot encoding of the index.
    if self.units == 1:
      # This can be slightly faster as it uses matmul.
      return tf.matmul(
          tf.one_hot(tf.squeeze(inputs, axis=[-1]), depth=self.num_buckets),
          self.kernel)
    return tf.reduce_sum(
        tf.one_hot(inputs, axis=1, depth=self.num_buckets) * self.kernel,
        axis=1) 

def __init__(self, sess, output_size, mainNet, targetNet, max_length=1000000):
        self.memory = collections.deque(maxlen=max_length)
        self.mainNet = mainNet
        self.targetNet = targetNet
        self.output_size = output_size
        self.batch_size = 8
        self.sess = sess
        self.gamma = 0.99
        self.lr = 0.00025

        self.target = tf.placeholder(tf.float32, [None])
        self.action = tf.placeholder(tf.int32, [None])

        self.target_vars = self.targetNet.get_trainable_variables()
        self.main_vars = self.mainNet.get_trainable_variables()

        self.update_oldpi_op = [oldp.assign(p) for p, oldp in zip(self.main_vars, self.target_vars)]

        self.action_one_hot = tf.one_hot(self.action, self.output_size)
        self.q_val = self.mainNet.Q
        self.q_val_action = tf.reduce_sum(self.q_val * self.action_one_hot, axis=1)

        self.loss = tf.reduce_mean((self.target - self.q_val_action) ** 2)
        self.train_op = tf.train.AdamOptimizer(learning_rate=self.lr, epsilon=1e-2).minimize(self.loss) 

def _build_policy_net(self):
    """Build policy network"""
    with tf.variable_scope(self.scope):
      self.state_input = tf.placeholder(tf.float32, [None, self.state_size])
      self.action = tf.placeholder(tf.int32, [None])
      self.target = tf.placeholder(tf.float32, [None])

      layer_1 = tf_utils.fc(self.state_input, self.n_hidden_1, tf.nn.relu)
      layer_2 = tf_utils.fc(layer_1, self.n_hidden_2, tf.nn.relu)

      self.value = tf_utils.fc(layer_2, 1)

      self.action_values = tf_utils.fc(layer_2, self.action_size)
      action_mask = tf.one_hot(self.action, self.action_size, 1.0, 0.0)
      self.action_value_pred = tf.reduce_sum(tf.nn.softmax(self.action_values) * action_mask, 1)
      
      self.action_probs = tf.nn.softmax(self.action_values)
      self.value_loss = tf.reduce_mean(tf.square(self.target - self.value))
      self.pg_loss = tf.reduce_mean(-tf.log(self.action_value_pred) * (self.target - self.value))
      self.l2_loss = tf.add_n([ tf.nn.l2_loss(v) for v in tf.trainable_variables()  ]) 
      self.loss = self.pg_loss + 5*self.value_loss + 0.002 * self.l2_loss
      
      self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
      self.train_op = self.optimizer.minimize(self.loss, global_step=tf.contrib.framework.get_global_step()) 

def _build_policy_net(self):
    """Build policy network"""
    with tf.variable_scope(self.scope):
      self.state_input = tf.placeholder(tf.float32, [None, self.state_size])
      self.action = tf.placeholder(tf.int32, [None])
      self.target = tf.placeholder(tf.float32, [None])

      layer_1 = tf_utils.fc(self.state_input, self.n_hidden_1, tf.nn.relu)
      layer_2 = tf_utils.fc(layer_1, self.n_hidden_2, tf.nn.relu)

      self.action_values = tf_utils.fc(layer_2, self.action_size)
      action_mask = tf.one_hot(self.action, self.action_size, 1.0, 0.0)
      self.action_prob = tf.nn.softmax(self.action_values)
      self.action_value_pred = tf.reduce_sum(self.action_prob * action_mask, 1)

      # l2 regularization
      self.l2_loss = tf.add_n([ tf.nn.l2_loss(v) for v in tf.trainable_variables()  ]) 
      self.pg_loss = tf.reduce_mean(-tf.log(self.action_value_pred) * self.target)

      self.loss = self.pg_loss + 0.002 * self.l2_loss
      self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
      self.train_op = self.optimizer.minimize(self.loss, global_step=tf.contrib.framework.get_global_step()) 

def one_hot_encoding(labels, num_classes=None):
  """One-hot encodes the multiclass labels.

  Example usage:
    labels = tf.constant([1, 4], dtype=tf.int32)
    one_hot = OneHotEncoding(labels, num_classes=5)
    one_hot.eval()    # evaluates to [0, 1, 0, 0, 1]

  Args:
    labels: A tensor of shape [None] corresponding to the labels.
    num_classes: Number of classes in the dataset.
  Returns:
    onehot_labels: a tensor of shape [num_classes] corresponding to the one hot
      encoding of the labels.
  Raises:
    ValueError: if num_classes is not specified.
  """
  with tf.name_scope('OneHotEncoding', values=[labels]):
    if num_classes is None:
      raise ValueError('num_classes must be specified')

    labels = tf.one_hot(labels, num_classes, 1, 0)
    return tf.reduce_max(labels, 0) 

def neglogp(self, x):
        # return tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits, labels=x)
        # Note: we can't use sparse_softmax_cross_entropy_with_logits because
        #       the implementation does not allow second-order derivatives...
        if x.dtype in {tf.uint8, tf.int32, tf.int64}:
            # one-hot encoding
            x_shape_list = x.shape.as_list()
            logits_shape_list = self.logits.get_shape().as_list()[:-1]
            for xs, ls in zip(x_shape_list, logits_shape_list):
                if xs is not None and ls is not None:
                    assert xs == ls, 'shape mismatch: {} in x vs {} in logits'.format(xs, ls)

            x = tf.one_hot(x, self.logits.get_shape().as_list()[-1])
        else:
            # already encoded
            assert x.shape.as_list() == self.logits.shape.as_list()

        return tf.nn.softmax_cross_entropy_with_logits_v2(
            logits=self.logits,
            labels=x) 

def _build_net(self):
        with tf.variable_scope("placeholder"):
            self.input_x = tf.placeholder(tf.float32, shape=(None, self.vocab_size))
            self.input_y = tf.placeholder(tf.int32, shape=(None,))
            Y_one_hot = tf.one_hot(self.input_y, self.n_class)
        
        with tf.variable_scope("output", reuse=tf.AUTO_REUSE):
            W = tf.get_variable('W', dtype=tf.float32,
                               initializer=tf.truncated_normal((self.vocab_size, self.n_class)))
            b = tf.get_variable('b', dtype=tf.float32,
                               initializer=tf.constant(0.1, shape=(self.n_class,)))
            logits = tf.nn.xw_plus_b(self.input_x, W, b)
            self.prob = tf.reduce_max(tf.nn.softmax(logits), axis=1)
            self.prediction = tf.cast(tf.argmax(logits, axis=1), tf.int32)
            
            
        with tf.variable_scope("loss"):
            self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y_one_hot))
        
        with tf.variable_scope("train", reuse=tf.AUTO_REUSE):
            optimizer = tf.train.AdamOptimizer(self.lr)
            self.train_op = optimizer.minimize(self.loss)
            
        with tf.variable_scope("accuracy"):
            correct = tf.equal(self.prediction, self.input_y)
            self.accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
        
        self.sess.run(tf.global_variables_initializer()) 

def _torso(self, input_):
    """Processing of all the visual and language inputs to the LSTM core."""

    # Extract the inputs
    last_action, env_output = input_
    last_reward, _, _, observation = env_output
    frame = observation[self._idx_frame]
    goal = observation[self._idx_goal]
    goal = tf.to_float(goal)

    # Convert to image to floats and normalise.
    frame = tf.to_float(frame)
    frame = snt.FlattenTrailingDimensions(dim_from=3)(frame)
    frame /= 255.0

    # Feed image through convnet.
    with tf.variable_scope('convnet'):
      # Convolutional layers.
      conv_out = self._convnet(frame)
      # Fully connected layer.
      conv_out = snt.BatchFlatten()(conv_out)
      conv_out = snt.Linear(256)(conv_out)
      conv_out = tf.nn.relu(conv_out)

    # Concatenate outputs of the visual and instruction pathways.
    if self._feed_action_and_reward:
      # Append clipped last reward and one hot last action.
      tf.logging.info('Append last reward clipped to: %f', self._max_reward)
      clipped_last_reward = tf.expand_dims(
          tf.clip_by_value(last_reward, -self._max_reward, self._max_reward),
          -1)
      tf.logging.info('Append last action (one-hot of %d)', self._num_actions)
      one_hot_last_action = tf.one_hot(last_action, self._num_actions)
      tf.logging.info('Append goal:')
      tf.logging.info(goal)
      action_and_reward = tf.concat([clipped_last_reward, one_hot_last_action],
                                    axis=1)
    else:
      action_and_reward = tf.constant([0], dtype=tf.float32)
    return conv_out, action_and_reward, goal 

def margin_logit_loss(model_logits, label, num_classes=10):
    """Computes difference between logit for `label` and next highest logit.

    The loss is high when `label` is unlikely (targeted by default).
    This follows the same interface as `loss_fn` for UnrolledOptimizer and
    pgd_attack, i.e. it returns a batch of loss values.
    """
    logit_mask = tf.one_hot(label, depth=num_classes, axis=-1)
    label_logits = reduce_sum(logit_mask * model_logits, axis=-1)
    logits_with_target_label_neg_inf = model_logits - logit_mask * 99999
    highest_nonlabel_logits = reduce_max(
        logits_with_target_label_neg_inf, axis=-1)
    loss = highest_nonlabel_logits - label_logits
    return loss