def pointer_decoder(encoder_inputs_emb,
decoder_inputs,
initial_state,
attention_states,
cell,
feed_prev=True,
dtype=dtypes.float32,
scope=None):
#print 'encoder_inputs',encoder_inputs_emb
#print 'decoder_inputs', decoder_inputs
#print 'attention_states', attention_states
encoder_inputs = encoder_inputs_emb
#attn_length = attention_states.get_shape()[1].value
#attn_size = attention_states.get_shape()[2].value
attn_length = shape(attention_states, 1)
attn_size = shape(attention_states, 2)
with tf.name_scope('attention_setup'):
attnw = tf.get_variable("AttnW", [1, attn_size, attn_size])
attention_states = tf.nn.conv1d(attention_states, attnw, 1, 'SAME')
attnv = tf.get_variable("AttnV", [attn_size])
sys.stdout = sys.stderr
def attention_weight(output):
y = _linear(output, attn_size, True)
y = tf.reshape(y, [-1, 1, attn_size])
# Calculate attention weights for every encoder's input by taking an inner product between the weight bector (attnv), and the conbined decoder's state with the encoder's output.
attention_vectors = tf.nn.softmax(
tf.reduce_sum(attnv * tf.tanh(y + attention_states), axis=2))
return attention_vectors
states = [initial_state]
outputs = []
pointed_idxs = []
for i, d in enumerate(tf.unstack(decoder_inputs, axis=1)):
with tf.name_scope('Decode_%d' % i):
if i > 0:
tf.get_variable_scope().reuse_variables()
pointed_idx = d
# in testing, inputs to decoder won't be used except the first one.
if feed_prev and i > 0:
# take argmax, convert the pointed index into one-hot, and get the pointed encoder_inputs by multiplying and reduce_sum.
pointed_idx = tf.argmax(output, axis=1, output_type=tf.int32)
pointed_idxs.append(pointed_idx)
with tf.name_scope('copy_from_encoder_inputs'):
pointed_idx = tf.reshape(
tf.one_hot(pointed_idx, depth=attn_length),
[-1, attn_length, 1])
inp = tf.reduce_sum(encoder_inputs * pointed_idx, axis=1)
inp = tf.stop_gradient(inp)
output, state = cell(inp, states[-1])
with tf.name_scope('attention_weight'):
output = attention_weight(output)
#print 'output', output
states.append(state)
outputs.append(output)
outputs = tf.stack(outputs, axis=1)
states = tf.stack(states, axis=1)
return outputs, states, pointed_idxs
def setup_placeholder(self, config):
'''
Prepare tf.placeholder and their lengthes.
They are kept as instance variables.
'''
self.e_inputs_w_ph = tf.placeholder(tf.int32, [None, None],
name="EncoderInputWords")
self.e_inputs_c_ph = tf.placeholder(tf.int32, [None, None, None],
name="EncoderInputChars")
#self.d_outputs_ph = tf.placeholder(
# tf.int32, [None, None], name="DecoderOutput")
self.d_outputs_ph = self.e_inputs_w_ph
self.is_training = tf.placeholder(tf.bool, [], name='is_training')
with tf.name_scope('keep_prob'):
self.keep_prob = 1.0 - tf.to_float(
self.is_training) * config.dropout_rate
with tf.name_scope('batch_size'):
self.batch_size = batch_size = shape(self.d_outputs_ph, 0)
with tf.name_scope('start_tokens'):
self.start_tokens = tf.tile(tf.constant([BOS_ID], dtype=tf.int32),
[batch_size])
with tf.name_scope('end_tokens'):
self.end_token = PAD_ID
end_tokens = tf.tile(tf.constant([self.end_token], dtype=tf.int32),
[batch_size])
# Count the length of each dialogue, utterance, (word).
with tf.name_scope('utterance_length'):
self.uttr_lengths = tf.count_nonzero(self.e_inputs_w_ph,
axis=1,
dtype=tf.int32)
'''
# Example of the decoder's inputs and outputs.
Against a given input ['how', 'are', 'you', '?'] to the decoder's placeholder,
- decoder's input : ['_BOS', 'how', 'are', 'you', '?']
- decoder's output (target) : ['how', 'are', 'you', '?', '_PAD']
- target_length: 5
- target_weights: [1, 1, 1, 1, 1]
Here, the token _PAD behaves as EOS.
'''
with tf.name_scope('decoder_inputs'):
self.decoder_inputs = tf.concat(
[tf.expand_dims(self.start_tokens, 1), self.d_outputs_ph],
axis=1)
# the length of decoder's inputs/outputs is increased by 1 because of BOS or EOS.
with tf.name_scope('target_lengths'):
self.target_length = tf.count_nonzero(
self.d_outputs_ph, axis=1, dtype=tf.int32) + 1
with tf.name_scope('target_weights'):
self.target_weights = tf.sequence_mask(self.target_length,
dtype=tf.float32)
with tf.name_scope('targets'):
self.targets = tf.concat(
[self.d_outputs_ph,
tf.expand_dims(end_tokens, 1)],
axis=1)[:, :shape(self.target_weights, 1)]
def encode(self, inputs, sequence_length):
with tf.variable_scope(self.shared_scope or "CNNEncoder"):
target_rank = 3 # [*, max_sequence_length, hidden_size]
flattened_inputs, prev_shape = flatten(inputs, target_rank)
flattened_aggregated_outputs = cnn(flattened_outputs,
activation=self.activation)
target_shape = prev_shape[:-2] + [
shape(flattened_aggregated_outputs, -1)
]
outputs = tf.reshape(flattened_aggregated_outputs, target_shape)
outputs = tf.nn.dropout(outputs, self.keep_prob)
return outputs, outputs
def setup_decoder_states(self,
config,
encoder_outputs,
encoder_state,
scope=None):
attention_states = encoder_outputs
response_emb = tf.nn.embedding_lookup(self.w_embeddings,
self.d_outputs_ph)
response_lengths = tf.count_nonzero(self.d_outputs_ph,
axis=1,
dtype=tf.int32)
print 'encoder_state', encoder_state
print 'encoder_outputs', encoder_outputs
_, h_future = self.uttr_encoder.encode(response_emb, response_lengths)
print 'h_future', h_future
def _get_distribution(state, output_size):
h = state
num_layers = 1
for i in range(num_layers):
with tf.variable_scope('linear%d' % i) as scope:
h = linear(h, output_size, scope=scope)
with tf.variable_scope('Mean'):
mean = linear(h, output_size, activation=None)
with tf.variable_scope('Var'):
var = linear(h, output_size, activation=tf.nn.softplus)
return tfd.MultivariateNormalDiag(mean, var)
output_size = shape(encoder_state, -1)
with tf.variable_scope('Prior'):
self.prior = _get_distribution(encoder_state, output_size)
with tf.variable_scope('Posterior'):
self.posterior = _get_distribution(
tf.concat([encoder_state, h_future], axis=-1), output_size)
train_decoder_state = tf.concat(
[encoder_state, self.posterior.sample()], axis=-1)
test_decoder_state = tf.concat(
[encoder_state, self.prior.sample()], axis=-1)
#train_decoder_state = encoder_state + self.posterior.sample()
#test_decoder_state = encoder_state + self.prior.sample()
print train_decoder_state
print test_decoder_state
return train_decoder_state, test_decoder_state, attention_states
def __init__(self, sess, conf, vocab):
ModelBase.__init__(self, sess, conf)
self.vocab = vocab
input_max_len, output_max_len = None, conf.output_max_len
self.is_training = tf.placeholder(tf.bool, [], name='is_training')
with tf.name_scope('keep_prob'):
self.keep_prob = 1.0 - tf.to_float(
self.is_training) * conf.dropout_rate
with tf.name_scope('EncoderInput'):
self.e_inputs_ph = tf.placeholder(tf.int32, [None, input_max_len],
name="EncoderInput")
with tf.name_scope('batch_size'):
batch_size = shape(self.e_inputs_ph, 0)
with tf.variable_scope('Embeddings') as scope:
self.w_embeddings = self.initialize_embeddings(
'Word',
vocab.embeddings.shape,
initializer=tf.constant_initializer(vocab.embeddings),
trainable=conf.train_embedding)
def __init__(self, sess, conf, vocab):
ModelBase.__init__(self, sess, conf)
self.vocab = vocab
input_max_len, output_max_len = None, conf.output_max_len
self.is_training = tf.placeholder(tf.bool, [], name='is_training')
with tf.name_scope('keep_prob'):
self.keep_prob = 1.0 - tf.to_float(
self.is_training) * conf.dropout_rate
# <Sample input>
# e_inputs: [1, 40, 44, 0, 0], d_outputs: [2, 0, 0] (target=44)
with tf.name_scope('EncoderInput'):
self.e_inputs_ph = tf.placeholder(tf.int32, [None, input_max_len],
name="EncoderInput")
with tf.name_scope('batch_size'):
batch_size = shape(self.e_inputs_ph, 0)
with tf.variable_scope('Embeddings') as scope:
w_embeddings = self.initialize_embeddings(
'Word',
vocab.embeddings.shape,
initializer=tf.constant_initializer(vocab.embeddings),
trainable=conf.train_embedding)
with tf.variable_scope('WordEncoder') as scope:
word_encoder = WordEncoder(conf,
w_embeddings,
self.keep_prob,
shared_scope=scope)
e_inputs_emb = word_encoder.encode([self.e_inputs_ph])
with tf.variable_scope('SentEncoder') as scope:
sent_encoder = SentenceEncoder(conf,
self.keep_prob,
shared_scope=scope)
e_inputs_length = tf.count_nonzero(self.e_inputs_ph, axis=1)
e_outputs, e_state = sent_encoder.encode(e_inputs_emb,
e_inputs_length)
attention_states = e_outputs
self.d_outputs_ph = []
self.losses = []
self.greedy_predictions = []
self.copied_inputs = []
for i, col_name in enumerate(conf.target_columns):
with tf.name_scope('DecoderOutput%d' % i):
d_outputs_ph = tf.placeholder(tf.int32, [None, output_max_len],
name="DecoderOutput")
ds_name = 'Decoder' if conf.share_decoder else 'Decoder%d' % i
with tf.variable_scope(ds_name) as scope:
d_cell = setup_cell(conf.cell_type,
conf.rnn_size,
conf.num_layers,
keep_prob=self.keep_prob)
teacher_forcing = conf.teacher_forcing if 'teacher_forcing' in conf else False
d_outputs, predictions, copied_inputs = setup_decoder(
d_outputs_ph,
e_inputs_emb,
e_state,
attention_states,
d_cell,
batch_size,
output_max_len,
scope=scope,
teacher_forcing=teacher_forcing)
self.copied_inputs.append(copied_inputs)
d_outputs_length = tf.count_nonzero(d_outputs_ph,
axis=1,
name='outputs_length')
with tf.name_scope('add_eos'):
targets = tf.concat([
d_outputs_ph,
tf.zeros([batch_size, 1], dtype=tf.int32)
],
axis=1)
# the length of outputs should be also added by 1 because of EOS.
with tf.name_scope('output_weights'):
d_outputs_weights = tf.sequence_mask(
d_outputs_length + 1,
maxlen=shape(d_outputs_ph, 1) + 1,
dtype=tf.float32)
with tf.name_scope('loss%d' % i):
loss = tf.contrib.seq2seq.sequence_loss(
d_outputs, targets, d_outputs_weights)
self.d_outputs_ph.append(d_outputs_ph)
self.losses.append(loss)
self.greedy_predictions.append(predictions)
with tf.name_scope('Loss'):
self.loss = tf.reduce_mean(self.losses)
self.updates = self.get_updates(self.loss)
def setup_decoder(self,
config,
train_decoder_state,
test_decoder_state,
embeddings,
encoder_input_lengths=None,
attention_states=None,
projection_layer=None,
scope=None):
batch_size = self.batch_size
decoder_inputs_emb = tf.nn.embedding_lookup(embeddings,
self.decoder_inputs)
# TODO: 多言語対応にする時はbias, trainableをfalseにしてembeddingをconstantにしたい
decoder_cell = setup_cell(config.decoder.cell_type,
shape(train_decoder_state, -1),
config.decoder.num_layers,
keep_prob=self.keep_prob)
if projection_layer is None:
with tf.variable_scope('projection') as scope:
kernel = tf.transpose(embeddings, perm=[1, 0])
projection_layer = SharedKernelDense(shape(embeddings, 0),
use_bias=False,
trainable=False,
shared_kernel=kernel)
with tf.name_scope('Training'):
train_decoder_cell = decoder_cell
decoder_initial_state = train_decoder_state
helper = tf.contrib.seq2seq.TrainingHelper(
decoder_inputs_emb,
sequence_length=self.target_length,
time_major=False)
decoder = tf.contrib.seq2seq.BasicDecoder(
train_decoder_cell,
helper,
decoder_initial_state,
output_layer=projection_layer)
train_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(
decoder,
impute_finished=True,
maximum_iterations=tf.reduce_max(self.target_length),
scope=scope)
logits = train_decoder_outputs.rnn_output
with tf.name_scope('Test'):
beam_width = config.beam_width
test_decoder_cell = decoder_cell
decoder_initial_state = tf.contrib.seq2seq.tile_batch(
test_decoder_state, multiplier=beam_width)
decoder = tf.contrib.seq2seq.BeamSearchDecoder(
test_decoder_cell,
embeddings,
self.start_tokens,
self.end_token,
decoder_initial_state,
beam_width,
output_layer=projection_layer,
length_penalty_weight=config.length_penalty_weight)
test_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(
decoder,
impute_finished=False,
maximum_iterations=config.utterance_max_len,
scope=scope)
predictions = test_decoder_outputs.predicted_ids
#self.beam_scores = test_decoder_outputs.beam_search_decoder_output.scores
# memo: 出力結果はbeam_scoresの低い順にならんでいて (負の値を取る)，概ねそれはちゃんと正確さと一致してそう？
return logits, predictions
def pointer_decoder(encoder_inputs_emb, decoder_inputs, initial_state,
attention_states, cell,
feed_prev=True, dtype=dtypes.float32, scope=None):
encoder_inputs = encoder_inputs_emb
attn_length = shape(attention_states, 1)
attn_size = shape(attention_states, 2)
with tf.name_scope('attention_setup'):
# Prepare the weights for attention calculation. We assume here the sizes of attention_states (encoder's outputs), encoder's state, decoder's output are same.
attnw = tf.get_variable("AttnW1", [1, attn_size, attn_size])
attnw2 = tf.get_variable("AttnW2", [attn_size, attn_size])
attnv = tf.get_variable("AttnV", [attn_size])
# Calculate W1 * attention_states in advance since each output and state of encoder is unchanged while decoding.
attention_states = tf.nn.conv1d(attention_states, attnw, 1, 'SAME')
sys.stdout = sys.stderr
def attention_weight(output):
"""
Calculate attention weights for every encoder's input by taking an inner product the weight bector (attnv) with the conbined and transformed the encoder's output and decoder's state.
output_probabilities[i] = V・tanh(W1・attention_state[i] + W2・decoder's output[t])
- i: the index of an input word
- t: current time-step in decoding
- V: a tensor with the shape [attention_size]
- W1: a tensor with the shape [attention_size, encoder's rnn_size]
- W2: a tensor with the shape [attention_size, decoder's rnn_size]
"""
y = tf.matmul(output, attnw2)
y = tf.reshape(y, [-1, 1, attn_size])
attention_vectors = tf.nn.softmax(tf.reduce_sum(attnv * tf.tanh(attention_states + y), axis=2))
return attention_vectors
states = [initial_state]
outputs = []
pointed_idxs = []
with tf.name_scope('Decode_Timestep'):
for i, d in enumerate(tf.unstack(decoder_inputs, axis=1)):
with tf.name_scope('Decode_%d' % i):
if i > 0:
tf.get_variable_scope().reuse_variables()
# The first input to the decoder is something like _START (or just a _PAD) token we prepared to start decoding.
pointed_idx = d
# If feed_prev == True, inputs to decoder won't be used except the first one. The model makes decisions of which should be the next inputs by itself.
if feed_prev and i > 0:
# Take argmax to decide which indices of input should be most possible.
pointed_idx = tf.argmax(output, axis=1, output_type=tf.int32)
pointed_idxs.append(pointed_idx)
with tf.name_scope('copy_from_encoder_inputs'):
# Convert the pointed index into one-hot, and get the pointed encoder_inputs by multiplying and reduce_sum.
pointed_idx = tf.reshape(tf.one_hot(pointed_idx, depth=attn_length), [-1, attn_length, 1])
inp = tf.reduce_sum(encoder_inputs * pointed_idx, axis=1)
# In their original paper, the gradients shouldn't be propagated to input embeddings through these copying. The embeddings should be updated only from the encoder.
inp = tf.stop_gradient(inp)
output, state = cell(inp, states[-1])
# Calculate the output (and the next input) distribution
with tf.name_scope('attention_weight'):
output = attention_weight(output)
states.append(state)
outputs.append(output)
with tf.name_scope('outputs'):
outputs = tf.stack(outputs, axis=1)
with tf.name_scope('states'):
states = tf.stack(states, axis=1)
with tf.name_scope('pointed_idx'):
pointed_idxs = tf.stack(pointed_idxs, axis=1)
return outputs, states, pointed_idxs
def setup_decoder(self,
config,
train_decoder_state,
test_decoder_state,
encoder_input_lengths=None,
attention_states=None,
projection_layer=None,
scope=None):
batch_size = self.batch_size
decoder_inputs_emb = tf.nn.embedding_lookup(self.w_embeddings,
self.decoder_inputs)
# TODO: 多言語対応にする時はbias, trainableをfalseにしてembeddingをconstantにしたい
decoder_cell = setup_cell(config.decoder.cell_type,
shape(train_decoder_state, -1),
config.decoder.num_layers,
keep_prob=self.keep_prob)
if projection_layer is None:
with tf.variable_scope('projection') as scope:
projection_layer = tf.layers.Dense(config.w_vocab_size,
use_bias=True,
trainable=True)
with tf.name_scope('Training'):
if config.attention_type:
assert attention_states is not None
num_units = shape(attention_states, -1)
attention = tf.contrib.seq2seq.LuongAttention(
num_units,
attention_states,
memory_sequence_length=encoder_input_length)
train_decoder_cell = tf.contrib.seq2seq.AttentionWrapper(
decoder_cell, attention)
decoder_initial_state = train_decoder_cell.zero_state(
batch_size,
tf.float32).clone(cell_state=train_decoder_state)
else:
train_decoder_cell = decoder_cell
decoder_initial_state = train_decoder_state
# encoder_state can't be directly copied into decoder_cell when using the attention mechanisms, initial_state must be an instance of AttentionWrapperState. (https://github.com/tensorflow/nmt/issues/205)
helper = tf.contrib.seq2seq.TrainingHelper(
decoder_inputs_emb,
sequence_length=self.target_length,
time_major=False)
decoder = tf.contrib.seq2seq.BasicDecoder(
train_decoder_cell,
helper,
decoder_initial_state,
output_layer=projection_layer)
train_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(
decoder,
impute_finished=True,
maximum_iterations=tf.reduce_max(self.target_length),
scope=scope)
logits = train_decoder_outputs.rnn_output
with tf.name_scope('Test'):
beam_width = config.beam_width
if config.attention_type:
num_units = shape(attention_states, -1)
attention = tf.contrib.seq2seq.LuongAttention(
num_units,
tf.contrib.seq2seq.tile_batch(attention_states,
multiplier=beam_width),
memory_sequence_length=tf.contrib.seq2seq.tile_batch(
encoder_input_length, multiplier=beam_width))
test_decoder_cell = tf.contrib.seq2seq.AttentionWrapper(
decoder_cell, attention)
decoder_initial_state = test_decoder_cell.zero_state(
batch_size * beam_width,
tf.float32).clone(cell_state=tf.contrib.seq2seq.tile_batch(
test_decoder_state, multiplier=beam_width))
else:
test_decoder_cell = decoder_cell
decoder_initial_state = tf.contrib.seq2seq.tile_batch(
test_decoder_state, multiplier=beam_width)
decoder = tf.contrib.seq2seq.BeamSearchDecoder(
test_decoder_cell,
self.w_embeddings,
self.start_tokens,
self.end_token,
decoder_initial_state,
beam_width,
output_layer=projection_layer,
length_penalty_weight=config.length_penalty_weight)
test_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(
decoder,
impute_finished=False,
maximum_iterations=config.utterance_max_len,
scope=scope)
predictions = test_decoder_outputs.predicted_ids
return logits, predictions
def __init__(self, sess, config, vocab, encoder=None, is_training=None):
PointerNetworkBase.__init__(self,
sess,
config,
vocab,
is_training=is_training)
input_max_len, output_max_len = None, config.output_max_len
# <Sample input>
# e_inputs: [1, 40, 44, 0, 0], d_outputs: [2, 0, 0] (target=44)
with tf.name_scope('EncoderInput'):
self.e_inputs_ph = tf.placeholder(tf.int32, [None, input_max_len],
name="EncoderInput")
self.pos_inputs_ph = tf.placeholder(tf.int32,
[None, input_max_len],
name="EncoderInputPOS")
self.wtype_inputs_ph = tf.placeholder(tf.int32,
[None, input_max_len],
name="EncoderInputWordType")
with tf.name_scope('batch_size'):
batch_size = shape(self.e_inputs_ph, 0)
with tf.variable_scope('Embeddings') as scope:
e_inputs_emb = []
w_embeddings = self.initialize_embeddings(
'Word',
vocab.word.embeddings.shape,
initializer=tf.constant_initializer(vocab.word.embeddings),
trainable=config.train_embedding)
e_inputs_emb.append(
tf.nn.embedding_lookup(w_embeddings, self.e_inputs_ph))
if self.use_pos:
pos_embeddings = self.initialize_embeddings(
'POS', [vocab.pos.size, config.feature_size],
trainable=True)
e_inputs_emb.append(
tf.nn.embedding_lookup(pos_embeddings, self.pos_inputs_ph))
if self.use_wtype:
wtype_embeddings = self.initialize_embeddings(
'Wtype', [vocab.wtype.size, config.feature_size],
trainable=True)
e_inputs_emb.append(
tf.nn.embedding_lookup(wtype_embeddings,
self.wtype_inputs_ph))
e_inputs_emb = tf.concat(e_inputs_emb, axis=-1)
e_inputs_emb = tf.nn.dropout(e_inputs_emb, self.keep_prob)
with tf.variable_scope('SentEncoder') as scope:
# If an encoder is not given, prepare a new one.
if encoder is None:
encoder_type = getattr(encoder_class, config.encoder_type)
sent_encoder = encoder_type(config,
self.keep_prob,
shared_scope=scope)
else:
sent_encoder = encoder
e_inputs_length = tf.count_nonzero(self.e_inputs_ph, axis=1)
e_outputs, e_state = sent_encoder.encode(e_inputs_emb,
e_inputs_length)
attention_states = e_outputs
self.d_outputs_ph = []
self.losses = []
self.greedy_predictions = []
self.copied_inputs = []
for i, col_name in enumerate(self.target_columns):
with tf.name_scope('DecoderOutput%d' % i):
d_outputs_ph = tf.placeholder(tf.int32, [None, output_max_len],
name="DecoderOutput")
ds_name = 'Decoder' if config.share_decoder else 'Decoder%d' % i
with tf.variable_scope(ds_name) as scope:
d_cell = setup_cell(config.cell_type,
config.rnn_size,
config.num_layers,
keep_prob=self.keep_prob)
teacher_forcing = config.teacher_forcing if 'teacher_forcing' in config else False
d_outputs, predictions, copied_inputs = setup_decoder(
d_outputs_ph,
e_inputs_emb,
e_state,
attention_states,
d_cell,
batch_size,
output_max_len,
scope=scope,
teacher_forcing=teacher_forcing)
self.copied_inputs.append(copied_inputs)
d_outputs_length = tf.count_nonzero(d_outputs_ph,
axis=1,
name='outputs_length')
with tf.name_scope('add_eos'):
targets = tf.concat([
d_outputs_ph,
tf.zeros([batch_size, 1], dtype=tf.int32)
],
axis=1)
# the length of outputs should be also added by 1 because of EOS.
with tf.name_scope('output_weights'):
d_outputs_weights = tf.sequence_mask(
d_outputs_length + 1,
maxlen=shape(d_outputs_ph, 1) + 1,
dtype=tf.float32)
with tf.name_scope('loss%d' % i):
loss = tf.contrib.seq2seq.sequence_loss(
d_outputs, targets, d_outputs_weights)
self.d_outputs_ph.append(d_outputs_ph)
self.losses.append(loss)
self.greedy_predictions.append(predictions)
with tf.name_scope('Loss'):
self.loss = tf.reduce_mean(self.losses)
self.updates = self.get_updates(self.loss)
