def conv_layer(inputs,
kernel_shape,
biases_shape,
stride=1,
padding='SAME',
activation='relu',
norm=None,
dropout=False,
dropout_rate=None,
regularizer=None,
initializer=None,
dimensions=2,
is_training=True):
if initializer is not None:
initializer_obj = get_initializer(initializer)
weights = tf.get_variable('weights',
initializer=initializer_obj(kernel_shape),
regularizer=regularizer)
else:
if activation == 'relu':
initializer = get_initializer('he_uniform')
elif activation == 'sigmoid' or activation == 'tanh':
initializer = get_initializer('glorot_uniform')
# if initializer is None, tensorFlow seems to be using
# a glorot uniform initializer
weights = tf.get_variable('weights',
kernel_shape,
regularizer=regularizer,
initializer=initializer)
logger.debug(' conv_weights: {0}'.format(weights))
biases = tf.get_variable('biases',
biases_shape,
initializer=tf.constant_initializer(0.01))
logger.debug(' conv_biases: {0}'.format(biases))
if dimensions == 1:
return conv_1d(inputs,
weights,
biases,
stride=stride,
padding=padding,
activation=activation,
norm=norm,
dropout=dropout,
dropout_rate=dropout_rate,
is_training=is_training)
elif dimensions == 2:
return conv_2d(inputs,
weights,
biases,
stride=stride,
padding=padding,
activation=activation,
norm=norm,
dropout=dropout,
dropout_rate=dropout_rate,
is_training=is_training)
else:
raise Exception('Unsupported number of dimensions', dimensions)
def fc_layer(inputs,
in_count,
out_count,
activation='relu',
norm=None,
is_training=True,
weights=None,
biases=None,
dropout=False,
dropout_rate=None,
initializer=None,
regularizer=None):
if weights is None:
if initializer is not None:
initializer_obj = get_initializer(initializer)
weights = tf.get_variable('weights',
initializer=initializer_obj(
[in_count, out_count]),
regularizer=regularizer)
else:
if activation == 'relu':
initializer = get_initializer('he_uniform')
elif activation == 'sigmoid' or activation == 'tanh':
initializer = get_initializer('glorot_uniform')
# if initializer is None, tensorFlow seems to be using
# a glorot uniform initializer
weights = tf.compat.v1.get_variable('weights',
[in_count, out_count],
regularizer=regularizer,
initializer=initializer)
logger.debug(' fc_weights: {}'.format(weights))
if biases is None:
biases = tf.compat.v1.get_variable(
'biases', [out_count], initializer=tf.constant_initializer(0.01))
logger.debug(' fc_biases: {}'.format(biases))
hidden = tf.matmul(inputs, weights) + biases
if norm is not None:
if norm == 'batch':
hidden = tf.contrib.layers.batch_norm(hidden,
is_training=is_training)
elif norm == 'layer':
hidden = tf.contrib.layers.layer_norm(hidden)
if activation:
hidden = getattr(tf.nn, activation)(hidden)
if dropout and dropout_rate is not None:
hidden = tf.layers.dropout(hidden,
rate=dropout_rate,
training=is_training)
logger.debug(' fc_dropout: {}'.format(hidden))
return hidden
def embedding_matrix(vocab,
embedding_size,
representation='dense',
embeddings_trainable=True,
pretrained_embeddings=None,
force_embedding_size=False,
initializer=None,
regularizer=None):
vocab_size = len(vocab)
if representation == 'dense':
if pretrained_embeddings is not None and pretrained_embeddings is not False:
embeddings_matrix = load_pretrained_embeddings(
pretrained_embeddings, vocab)
if embeddings_matrix.shape[-1] != embedding_size:
raise ValueError(
'The size of the pretrained embedding size is {}, '
'but the specified embedding_size is {}. '
'Please change the embedding_size accordingly.'.format(
embeddings_matrix.shape[-1], embedding_size))
initializer_obj = tf.constant(embeddings_matrix, dtype=tf.float32)
else:
if vocab_size < embedding_size and not force_embedding_size:
logger.info(
' embedding_size ({}) is greater than vocab_size ({}). '
'Setting embedding size to be equal to vocab_size.'.format(
embedding_size, vocab_size))
embedding_size = vocab_size
if initializer is not None:
initializer_obj_ref = get_initializer(initializer)
else:
initializer_obj_ref = get_initializer({
'type': 'uniform',
'minval': -1.0,
'maxval': 1.0
})
initializer_obj = initializer_obj_ref([vocab_size, embedding_size])
embeddings = tf.compat.v1.get_variable('embeddings',
initializer=initializer_obj,
trainable=embeddings_trainable,
regularizer=regularizer)
elif representation == 'sparse':
embedding_size = vocab_size
embeddings = tf.compat.v1.get_variable(
'embeddings',
initializer=get_initializer('identity')(
[vocab_size, embedding_size]),
trainable=False)
else:
raise Exception('Embedding representation {} not supported.'.format(
representation))
return embeddings, embedding_size
def _get_predictions(self, hidden, hidden_size, regularizer=None):
if not self.regularize:
regularizer = None
with tf.variable_scope('predictions_{}'.format(self.name)):
initializer_obj = get_initializer(self.initializer)
weights = tf.get_variable('weights',
initializer=initializer_obj(
[hidden_size, self.num_classes]),
regularizer=regularizer)
logging.debug(' class_weights: {0}'.format(weights))
biases = tf.get_variable('biases', [self.num_classes])
logging.debug(' class_biases: {0}'.format(biases))
logits = tf.matmul(hidden, weights) + biases
logging.debug(' logits: {0}'.format(logits))
probabilities = tf.nn.softmax(logits,
name='probabilities_{}'.format(
self.name))
predictions = tf.argmax(logits,
-1,
name='predictions_{}'.format(self.name))
with tf.device('/cpu:0'):
top_k_predictions = tf.nn.top_k(
logits,
k=self.top_k,
sorted=True,
name='top_k_predictions_{}'.format(self.name))
return (predictions, top_k_predictions, probabilities, logits, weights,
biases)
def vector_predictions(
self,
hidden,
hidden_size,
regularizer=None,
):
with tf.variable_scope('predictions_{}'.format(self.name)):
initializer_obj = get_initializer(self.initializer)
weights = tf.compat.v1.get_variable(
'weights',
initializer=initializer_obj([hidden_size, self.vector_size]),
regularizer=regularizer)
logger.debug(' projection_weights: {0}'.format(weights))
biases = tf.compat.v1.get_variable('biases', [self.vector_size])
logger.debug(' projection_biases: {0}'.format(biases))
logits = tf.matmul(hidden, weights) + biases
logger.debug(' logits: {0}'.format(logits))
if self.softmax:
predictions = tf.nn.softmax(logits)
else:
predictions = logits
return logits, self.vector_size, predictions
def _get_predictions(self, hidden, hidden_size, regularizer=None):
if not self.regularize:
regularizer = None
with tf.variable_scope('predictions_{}'.format(self.name)):
initializer_obj = get_initializer(self.initializer)
weights = tf.get_variable('weights',
initializer=initializer_obj(
[hidden_size, 1]),
regularizer=regularizer)
logger.debug(' regression_weights: {0}'.format(weights))
biases = tf.get_variable('biases', [1])
logger.debug(' regression_biases: {0}'.format(biases))
logits = tf.reshape(tf.matmul(hidden, weights) + biases, [-1])
logger.debug(' logits: {0}'.format(logits))
probabilities = tf.nn.sigmoid(logits,
name='probabilities_{}'.format(
self.name))
predictions = tf.greater_equal(probabilities,
self.threshold,
name='predictions_{}'.format(
self.name))
return predictions, probabilities, logits
def _get_predictions(self, hidden, hidden_size, regularizer=None):
if not self.regularize:
regularizer = None
with tf.variable_scope('predictions_{}'.format(self.name)):
initializer_obj = get_initializer(self.initializer)
weights = tf.get_variable('weights',
initializer=initializer_obj(
[hidden_size, 1]),
regularizer=regularizer)
logger.debug(' regression_weights: {0}'.format(weights))
biases = tf.get_variable('biases', [1])
logger.debug(' regression_biases: {0}'.format(biases))
predictions = tf.reshape(tf.matmul(hidden, weights) + biases, [-1])
logger.debug(' predictions: {0}'.format(predictions))
if self.clip is not None:
if isinstance(self.clip,
(list, tuple)) and len(self.clip) == 2:
predictions = tf.clip_by_value(predictions, self.clip[0],
self.clip[1])
logger.debug(
' clipped_predictions: {0}'.format(predictions))
else:
raise ValueError(
'The clip parameter of {} is {}. '
'It must be a list or a tuple of length 2.'.format(
self.name, self.clip))
return predictions
def _get_predictions(
self,
hidden,
hidden_size,
regularizer=None
):
if not self.regularize:
regularizer = None
with tf.variable_scope('predictions_{}'.format(self.name)):
initializer_obj = get_initializer(self.initializer)
weights = tf.get_variable(
'weights',
initializer=initializer_obj([hidden_size, 1]),
regularizer=regularizer
)
logging.debug(' regression_weights: {0}'.format(weights))
biases = tf.get_variable('biases', [1])
logging.debug(' regression_biases: {0}'.format(biases))
predictions = tf.reshape(
tf.matmul(hidden, weights) + biases,
[-1]
)
logging.debug(' predictions: {0}'.format(predictions))
return predictions
def __call__(self,
output_feature,
targets,
hidden,
hidden_size,
regularizer,
is_timeseries=False):
logging.info(' hidden shape: {0}'.format(hidden.shape))
if len(hidden.shape) != 3:
raise ValueError(
'Decoder inputs rank is {}, but should be 3 [batch x sequence x hidden] '
'when using a tagger sequential decoder. '
'Consider setting reduce_output to null / None if a sequential encoder / combiner is used.'
.format(len(hidden.shape)))
if is_timeseries:
output_feature['num_classes'] = 1
if not self.regularize:
regularizer = None
sequence_length = tf.shape(hidden)[1]
if self.attention:
hidden, hidden_size = feed_forward_memory_attention(
hidden, hidden, hidden_size)
targets_sequence_length = sequence_length_2D(targets)
initializer_obj = get_initializer(self.initializer)
class_weights = tf.get_variable('weights',
initializer=initializer_obj([
hidden_size,
output_feature['num_classes']
]),
regularizer=regularizer)
logging.debug(' weights: {0}'.format(class_weights))
class_biases = tf.get_variable('biases',
[output_feature['num_classes']])
logging.debug(' biases: {0}'.format(class_biases))
hidden_reshape = tf.reshape(hidden, [-1, hidden_size])
logits_to_reshape = tf.matmul(hidden_reshape,
class_weights) + class_biases
logits = tf.reshape(
logits_to_reshape,
[-1, sequence_length, output_feature['num_classes']])
logging.debug(' logits: {0}'.format(logits))
if is_timeseries:
probabilities_sequence = tf.zeros_like(logits)
predictions_sequence = tf.reshape(logits, [-1, sequence_length])
else:
probabilities_sequence = tf.nn.softmax(
logits, name='probabilities_{}'.format(output_feature['name']))
predictions_sequence = tf.argmax(logits,
-1,
name='predictions_{}'.format(
output_feature['name']),
output_type=tf.int32)
predictions_sequence_length = sequence_length_3D(hidden)
return predictions_sequence, probabilities_sequence, \
predictions_sequence_length, \
probabilities_sequence, targets_sequence_length, \
logits, hidden, class_weights, class_biases
def recurrent_decoder(encoder_outputs,
targets,
max_sequence_length,
vocab_size,
cell_type='rnn',
state_size=256,
embedding_size=50,
num_layers=1,
attention_mechanism=None,
beam_width=1,
projection=True,
tied_target_embeddings=True,
embeddings=None,
initializer=None,
regularizer=None,
is_timeseries=False):
with tf.variable_scope('rnn_decoder',
reuse=tf.AUTO_REUSE,
regularizer=regularizer):
# ================ Setup ================
if beam_width > 1 and is_timeseries:
raise ValueError('Invalid beam_width: {}'.format(beam_width))
GO_SYMBOL = vocab_size
END_SYMBOL = 0
batch_size = tf.shape(encoder_outputs)[0]
# ================ Projection ================
# Project the encoder outputs to the size of the decoder state
encoder_outputs_size = encoder_outputs.shape[-1]
if projection and encoder_outputs_size != state_size:
with tf.variable_scope('projection'):
encoder_output_rank = len(encoder_outputs.shape)
if encoder_output_rank > 2:
sequence_length = tf.shape(encoder_outputs)[1]
encoder_outputs = tf.reshape(encoder_outputs,
[-1, encoder_outputs_size])
encoder_outputs = fc_layer(encoder_outputs,
encoder_outputs.shape[-1],
state_size,
activation=None,
initializer=initializer)
encoder_outputs = tf.reshape(
encoder_outputs, [-1, sequence_length, state_size])
else:
encoder_outputs = fc_layer(encoder_outputs,
encoder_outputs.shape[-1],
state_size,
activation=None,
initializer=initializer)
# ================ Targets sequence ================
# Calculate the length of inputs and the batch size
with tf.variable_scope('sequence'):
targets_sequence_length = sequence_length_2D(targets)
start_tokens = tf.tile([GO_SYMBOL], [batch_size])
end_tokens = tf.tile([END_SYMBOL], [batch_size])
if is_timeseries:
start_tokens = tf.cast(start_tokens, tf.float32)
end_tokens = tf.cast(end_tokens, tf.float32)
targets_with_go_and_eos = tf.concat([
tf.expand_dims(start_tokens, 1), targets,
tf.expand_dims(end_tokens, 1)
], 1)
logging.debug(
' targets_with_go: {0}'.format(targets_with_go_and_eos))
targets_sequence_length_with_eos = targets_sequence_length + 1 # the EOS symbol is 0 so it's not increasing the real length of the sequence
# ================ Embeddings ================
if is_timeseries:
targets_embedded = tf.expand_dims(targets_with_go_and_eos, -1)
targets_embeddings = None
else:
with tf.variable_scope('embedding'):
if embeddings is not None:
embedding_size = embeddings.shape.as_list()[-1]
if tied_target_embeddings:
state_size = embedding_size
elif tied_target_embeddings:
embedding_size = state_size
if embeddings is not None:
embedding_go = tf.get_variable(
'embedding_GO',
initializer=tf.random_uniform([1, embedding_size],
-1.0, 1.0))
targets_embeddings = tf.concat([embeddings, embedding_go],
axis=0)
else:
initializer_obj = get_initializer(initializer)
targets_embeddings = tf.get_variable(
'embeddings',
initializer=initializer_obj(
[vocab_size + 1, embedding_size]),
regularizer=regularizer)
logging.debug(
' targets_embeddings: {0}'.format(targets_embeddings))
targets_embedded = tf.nn.embedding_lookup(
targets_embeddings,
targets_with_go_and_eos,
name='decoder_input_embeddings')
logging.debug(' targets_embedded: {0}'.format(targets_embedded))
# ================ Class prediction ================
if tied_target_embeddings:
class_weights = tf.transpose(targets_embeddings)
else:
initializer_obj = get_initializer(initializer)
class_weights = tf.get_variable('class_weights',
initializer=initializer_obj(
[state_size, vocab_size + 1]),
regularizer=regularizer)
logging.debug(' class_weights: {0}'.format(class_weights))
class_biases = tf.get_variable('class_biases', [vocab_size + 1])
logging.debug(' class_biases: {0}'.format(class_biases))
projection_layer = Projection(class_weights, class_biases)
# ================ RNN ================
initial_state = encoder_outputs
with tf.variable_scope('rnn_cells') as vs:
# Cell
cell_fun = get_cell_fun(cell_type)
if num_layers == 1:
cell = cell_fun(state_size)
if cell_type.startswith('lstm'):
initial_state = LSTMStateTuple(c=initial_state,
h=initial_state)
elif num_layers > 1:
cell = MultiRNNCell(
[cell_fun(state_size) for _ in range(num_layers)],
state_is_tuple=True)
if cell_type.startswith('lstm'):
initial_state = LSTMStateTuple(c=initial_state,
h=initial_state)
initial_state = tuple([initial_state] * num_layers)
else:
raise ValueError(
'num_layers in recurrent decoser: {}. '
'Number of layers in a recurrenct decoder cannot be <= 0'.
format(num_layers))
# Attention
if attention_mechanism is not None:
if attention_mechanism == 'bahdanau':
attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(
num_units=state_size,
memory=encoder_outputs,
memory_sequence_length=sequence_length_3D(
encoder_outputs))
elif attention_mechanism == 'luong':
attention_mechanism = tf.contrib.seq2seq.LuongAttention(
num_units=state_size,
memory=encoder_outputs,
memory_sequence_length=sequence_length_3D(
encoder_outputs))
else:
raise ValueError(
'Attention mechanism {} not supported'.format(
attention_mechanism))
cell = tf.contrib.seq2seq.AttentionWrapper(
cell, attention_mechanism, attention_layer_size=state_size)
initial_state = cell.zero_state(dtype=tf.float32,
batch_size=batch_size)
initial_state = initial_state.clone(
cell_state=reduce_sequence(encoder_outputs, 'last'))
for v in tf.global_variables():
if v.name.startswith(vs.name):
logging.debug(' {}: {}'.format(v.name, v))
# ================ Decoding ================
def decode(initial_state,
cell,
helper,
beam_width=1,
projection_layer=None):
# The decoder itself
if beam_width > 1:
# Tile inputs for beam search decoder
beam_initial_state = tf.contrib.seq2seq.tile_batch(
initial_state, beam_width)
decoder = tf.contrib.seq2seq.BeamSearchDecoder(
cell=cell,
embedding=targets_embeddings,
start_tokens=start_tokens,
end_token=END_SYMBOL,
initial_state=beam_initial_state,
beam_width=beam_width,
output_layer=projection_layer)
else:
decoder = BasicDecoder(cell=cell,
helper=helper,
initial_state=initial_state,
output_layer=projection_layer)
# The decoding operation
outputs = tf.contrib.seq2seq.dynamic_decode(
decoder=decoder,
output_time_major=False,
impute_finished=False if beam_width > 1 else True,
maximum_iterations=max_sequence_length)
return outputs
# ================ Decoding helpers ================
if is_timeseries:
train_helper = TimeseriesTrainingHelper(
inputs=targets_embedded,
sequence_length=targets_sequence_length_with_eos)
final_outputs_pred, final_state_pred, final_sequence_lengths_pred = decode(
initial_state,
cell,
train_helper,
projection_layer=projection_layer)
eval_logits = final_outputs_pred.rnn_output
train_logits = final_outputs_pred.projection_input
predictions_sequence = tf.reshape(eval_logits, [batch_size, -1])
predictions_sequence_length_with_eos = final_sequence_lengths_pred
else:
train_helper = tf.contrib.seq2seq.TrainingHelper(
inputs=targets_embedded,
sequence_length=targets_sequence_length_with_eos)
final_outputs_train, final_state_train, final_sequence_lengths_train = decode(
initial_state,
cell,
train_helper,
projection_layer=projection_layer)
eval_logits = final_outputs_train.rnn_output
train_logits = final_outputs_train.projection_input
# train_predictions = final_outputs_train.sample_id
pred_helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(
embedding=targets_embeddings,
start_tokens=start_tokens,
end_token=END_SYMBOL)
final_outputs_pred, final_state_pred, final_sequence_lengths_pred = decode(
initial_state,
cell,
pred_helper,
beam_width,
projection_layer=projection_layer)
if beam_width > 1:
predictions_sequence = final_outputs_pred.beam_search_decoder_output.predicted_ids[:, :,
0]
# final_outputs_pred..predicted_ids[:,:,0] would work too, but it contains -1s for padding
predictions_sequence_scores = final_outputs_pred.beam_search_decoder_output.scores[:, :,
0]
predictions_sequence_length_with_eos = final_sequence_lengths_pred[:,
0]
else:
predictions_sequence = final_outputs_pred.sample_id
predictions_sequence_scores = final_outputs_pred.rnn_output
predictions_sequence_length_with_eos = final_sequence_lengths_pred
logging.debug(' train_logits: {0}'.format(train_logits))
logging.debug(' eval_logits: {0}'.format(eval_logits))
logging.debug(' predictions_sequence: {0}'.format(predictions_sequence))
logging.debug(' predictions_sequence_scores: {0}'.format(
predictions_sequence_scores))
return predictions_sequence, predictions_sequence_scores, predictions_sequence_length_with_eos, \
targets_sequence_length_with_eos, eval_logits, train_logits, class_weights, class_biases
