def __init__(self, args, is_training=True):
self.args = args
if not is_training:
args.batch_size = 1
args.seq_length = 1
if args.model == 'rnn':
self.cell = rnn_cell.BasicRNNCell(args.rnn_size)
elif args.model == 'gru':
self.cell = rnn_cell.GRUCell(args.rnn_size)
elif args.model == 'lstm':
self.cell = rnn_cell.BasicLSTMCell(args.rnn_size)
else:
raise Exception('model type not supported: {}'.format(args.model))
self.cell = rnn_cell.MultiRNNCell([self.cell] * args.num_layers)
self.input_data = tf.placeholder(tf.int32, [args.batch_size, args.seq_length])
self.targets = tf.placeholder(tf.int32, [args.batch_size, args.seq_length]) # Target replication
self.initial_state = self.cell.zero_state(args.batch_size, tf.float32)
with tf.variable_scope('rnn'):
softmax_w = tf.get_variable('softmax_w', [args.rnn_size, 2])
softmax_b = tf.get_variable('softmax_b', [2])
with tf.device('/cpu:0'):
embedding = tf.get_variable('embedding', [args.vocab_size, args.rnn_size])
inputs = tf.split(1, args.seq_length, tf.nn.embedding_lookup(embedding, self.input_data))
inputs = [tf.squeeze(i, [1]) for i in inputs]
outputs, last_state = seq2seq.rnn_decoder(inputs, self.initial_state,
self.cell, loop_function=None)
output_tf = tf.reshape(tf.concat(1, outputs), [-1, args.rnn_size])
self.logits = tf.nn.xw_plus_b(output_tf, softmax_w, softmax_b)
self.probs = tf.nn.softmax(self.logits)
loss = seq2seq.sequence_loss_by_example(
[self.logits],
[tf.reshape(self.targets, [-1])],
[tf.ones([args.batch_size * args.seq_length])])
self.cost = tf.reduce_sum(loss) / args.batch_size / args.seq_length
self.final_state = last_state
self.lr = tf.Variable(0.0, trainable = False)
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(self.cost, tvars, aggregation_method=2), args.grad_clip)
optimizer = tf.train.AdamOptimizer(self.lr)
self.train_op = optimizer.apply_gradients(zip(grads, tvars))
def __init__(self,
sess,
vocab_size,
cell_size,
embedding_size,
num_layer,
memory_size,
log_dir,
learning_rate=0.001,
momentum=0.9,
learning_rate_decay_factor=0.85,
use_dropout=True,
l2_coef=1e-6):
with tf.name_scope("io"):
self.inputs = tf.placeholder(dtype=tf.int32,
shape=(None, None),
name="prev_words")
self.input_lens = tf.placeholder(dtype=tf.int32,
shape=(None, ),
name="sent_len")
self.labels = tf.placeholder(dtype=tf.int32,
shape=(None, None),
name="next_words")
self.keep_prob = tf.placeholder(dtype=tf.float32, name="keep_prob")
self.learning_rate = tf.Variable(float(learning_rate),
trainable=False)
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * learning_rate_decay_factor)
max_sent_len = array_ops.shape(self.labels)[1]
with variable_scope.variable_scope("word-embedding"):
embedding = tf.get_variable("embedding",
[vocab_size, embedding_size],
dtype=tf.float32)
input_embedding = embedding_ops.embedding_lookup(
embedding,
tf.squeeze(tf.reshape(self.inputs, [-1, 1]), squeeze_dims=[1]))
input_embedding = tf.reshape(input_embedding,
[-1, max_sent_len, embedding_size])
with variable_scope.variable_scope("rnn"):
# cell = tf_helpers.MemoryGRUCell(cell_size, memory_size, attn_size=100)
cell = rnn_cell.BasicLSTMCell(cell_size)
if use_dropout:
cell = rnn_cell.DropoutWrapper(cell,
output_keep_prob=self.keep_prob,
input_keep_prob=self.keep_prob)
if num_layer > 1:
cell = rnn_cell.MultiRNNCell([cell] * num_layer,
state_is_tuple=True)
# add output projection
cell = tf.nn.rnn_cell.OutputProjectionWrapper(cell, vocab_size)
# and enc_last_state will be same as the true last state
self.logits, last_state = tf.nn.dynamic_rnn(
cell,
input_embedding,
dtype=tf.float32,
sequence_length=self.input_lens,
)
self.loss = self.sequence_loss()
tf.scalar_summary("entropy_loss", self.loss)
tf.scalar_summary("perplexity", tf.exp(self.loss))
self.summary_op = tf.merge_all_summaries()
# weight decay
"""
if l2_coef > 0.0:
all_weights = []
vars = tf.trainable_variables()
for v in vars:
if "bias" not in v.name.lower():
all_weights.append(tf.nn.l2_loss(v))
print("adding l2 to %s" %v.name)
loss_l2= tf.add_n(all_weights)
self.reg_loss = self.loss + l2_coef * loss_l2
else:
self.reg_loss = self.loss
"""
# optimization
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(self.loss, tvars), 5.0)
optimizer = tf.train.GradientDescentOptimizer(self.learning_rate)
self.train_ops = optimizer.apply_gradients(zip(grads, tvars))
train_log_dir = os.path.join(log_dir, "train")
valid_log_dir = os.path.join(log_dir, "valid")
print "Save summary to %s" % log_dir
self.train_summary_writer = tf.train.SummaryWriter(
train_log_dir, sess.graph)
self.valid_summary_writer = tf.train.SummaryWriter(
valid_log_dir, sess.graph)
self.saver = tf.train.Saver(tf.all_variables())
def __init__(self, args, is_training=True):
if not is_training:
seq_length = 1
else:
seq_length = args.seq_length
if args.model == 'rnn':
cell_gen = rnn_cell.BasicRNNCell(args.rnn_size)
cell_dis = rnn_cell.BasicRNNCell(args.rnn_size)
elif args.model == 'gru':
cell_gen = rnn_cell.GRUCell(args.rnn_size)
cell_dis = rnn_cell.GRUCell(args.rnn_size)
elif args.model == 'lstm':
cell_gen = rnn_cell.BasicLSTMCell(args.rnn_size)
cell_dis = rnn_cell.BasicLSTMCell(args.rnn_size)
else:
raise Exception('model type not supported: {}'.format(args.model))
# Pass the generated sequences and targets (1)
with tf.name_scope('input'):
with tf.name_scope('data'):
self.input_data = tf.placeholder(tf.int32, [args.batch_size, seq_length])
with tf.name_scope('targets'):
self.targets = tf.placeholder(tf.int32, [args.batch_size, seq_length])
############
# Generator
############
with tf.variable_scope('generator'):
self.cell_gen = rnn_cell.MultiRNNCell([cell_gen] * args.num_layers)
self.initial_state_gen = self.cell_gen.zero_state(args.batch_size, tf.float32)
with tf.variable_scope('rnn'):
softmax_w = tf.get_variable('softmax_w', [args.rnn_size, args.vocab_size])
softmax_b = tf.get_variable('softmax_b', [args.vocab_size])
with tf.device('/cpu:0'):
embedding = tf.get_variable('embedding', [args.vocab_size, args.rnn_size])
inputs_gen = tf.split(1, seq_length, tf.nn.embedding_lookup(
embedding, self.input_data))
inputs_gen = [tf.squeeze(i, [1]) for i in inputs_gen]
outputs_gen, last_state_gen = seq2seq.rnn_decoder(inputs_gen, self.initial_state_gen,
self.cell_gen, loop_function=None)
self.logits_sequence = []
for output_gen in outputs_gen:
logits_gen = tf.nn.xw_plus_b(output_gen, softmax_w, softmax_b)
self.logits_sequence.append(logits_gen)
self.final_state_gen = last_state_gen
################
# Discriminator
################
with tf.variable_scope('discriminator'):
self.cell_dis = rnn_cell.MultiRNNCell([cell_dis] * args.num_layers)
self.initial_state_dis = self.cell_dis.zero_state(args.batch_size, tf.float32)
with tf.variable_scope('rnn'):
softmax_w = tf.get_variable('softmax_w', [args.rnn_size, 2])
softmax_b = tf.get_variable('softmax_b', [2])
inputs_dis = []
embedding = tf.get_variable('embedding', [args.vocab_size, args.rnn_size])
for logit in self.logits_sequence:
inputs_dis.append(tf.matmul(logit, embedding))
# inputs_dis.append(tf.matmul(tf.nn.softmax(logit), embedding))
outputs_dis, last_state_dis = seq2seq.rnn_decoder(inputs_dis,
self.initial_state_dis, self.cell_dis, loop_function=None)
probs, logits = [], []
for output_dis in outputs_dis:
logit = tf.nn.xw_plus_b(output_dis, softmax_w, softmax_b)
prob = tf.nn.softmax(logit)
logits.append(logit)
probs.append(prob)
with tf.name_scope('summary'):
probs = tf.pack(probs)
probs_real = tf.slice(probs, [0,0,1], [args.seq_length, args.batch_size, 1])
variable_summaries(probs_real, 'probability of real')
self.final_state_dis = last_state_dis
#########
# Train
#########
with tf.name_scope('train'):
gen_loss = seq2seq.sequence_loss_by_example(
logits,
tf.unpack(tf.transpose(self.targets)),
tf.unpack(tf.transpose(tf.ones_like(self.targets, dtype=tf.float32))))
self.gen_cost = tf.reduce_sum(gen_loss) / args.batch_size
tf.scalar_summary('training loss', self.gen_cost)
self.lr_gen = tf.Variable(0.0, trainable = False)
self.tvars = tf.trainable_variables()
gen_vars = [v for v in self.tvars if not v.name.startswith("discriminator/")]
if is_training:
gen_grads = tf.gradients(self.gen_cost, gen_vars)
self.all_grads = tf.gradients(self.gen_cost, self.tvars)
gen_grads_clipped, _ = tf.clip_by_global_norm(gen_grads, args.grad_clip)
gen_optimizer = tf.train.AdamOptimizer(self.lr_gen)
self.gen_train_op = gen_optimizer.apply_gradients(
zip(gen_grads_clipped, gen_vars))
with tf.name_scope('summary'):
with tf.name_scope('weight_summary'):
for v in self.tvars:
variable_summaries(v, v.op.name)
if is_training:
with tf.name_scope('grad_summary'):
for var, grad in zip(self.tvars, self.all_grads):
variable_summaries(grad, 'grad/' + var.op.name)
self.merged = tf.merge_all_summaries()
def __init__(self, embed_size, lstm_size, vocab_size, \
batch_size, seq_length, learn_rate, \
keep_prob=1.0, num_layers=2, name='char_lstm'):
'''
Initialize a character-level multilayer LSTM language model.
Arguments:
@embed_size: dimensions of embedding space
@vocab_size: number of things in vocabulary (characters!)
@batch_size: sequences per training batch
@seq_length: length of sequences in each training batch
@learn_rate: AdamOptimizer step size
@keep_prob: 1 - dropout probability
@num_layers: number of LSTM cells to stack
'''
# store params
self.embed_size, self.lstm_size = embed_size, lstm_size
self.vocab_size, self.seq_length = vocab_size, seq_length
self.batch_size, self.learn_rate = batch_size, learn_rate
self.kp, self.num_layers = keep_prob, num_layers
self.name = name
# Placeholders for input/output and dropout
self.train_inputs = tf.placeholder(tf.int32,
shape=[batch_size, seq_length])
self.train_targets = tf.placeholder(tf.int32,
shape=[batch_size, seq_length])
self.sample_inputs = tf.placeholder(tf.int32, shape=[1, 1])
# Set up embeddings
E = weight('embedding', [vocab_size, embed_size])
train_embeddings = tf.nn.embedding_lookup(E,
self.train_inputs,
name='train_embeddings')
dropped_train_embeddings = (tf.nn.dropout(train_embeddings, self.kp)
if self.kp < 1.0 else train_embeddings)
sample_embeddings = tf.nn.embedding_lookup(E,
self.sample_inputs,
name='sample_embeddings')
# Set up 2-layer LSTM
# Use dynamic_rnn to run the cells
with tf.variable_scope('lstm') as scope:
single_cell = rnn_cell.BasicLSTMCell(lstm_size)
self.cell = rnn_cell.MultiRNNCell([single_cell] * num_layers,
state_is_tuple=True)
self.train_init_state = self.cell.zero_state(
batch_size, tf.float32)
self.sample_init_state = self.cell.zero_state(1, tf.float32)
train_outputs, self.train_state = dynamic_rnn(
self.cell,
dropped_train_embeddings,
initial_state=self.train_init_state)
scope.reuse_variables()
sample_outputs, self.sample_state = dynamic_rnn(
self.cell,
sample_embeddings,
initial_state=self.sample_init_state)
reshaped_train_outputs = tf.reshape(
train_outputs, (batch_size * seq_length, lstm_size))
reshaped_sample_outputs = tf.reshape(sample_outputs, (1, lstm_size))
# final feedforward layer (model logits)
with tf.variable_scope('ff') as scope:
ff_weights = weight('ff_weights', (lstm_size, vocab_size))
ff_biases = weight('ff_biases', (vocab_size, ))
self.ls = train_logits = tf.add(
ff_biases, tf.matmul(reshaped_train_outputs, ff_weights))
scope.reuse_variables()
sample_logits = tf.add(
ff_biases, tf.matmul(reshaped_sample_outputs, ff_weights))
self.probs = softmax(sample_logits)
# softmax and loss for training
log_perps = tf.nn.seq2seq.sequence_loss_by_example(
[train_logits], [tf.reshape(self.train_targets, [-1])],
[tf.ones([batch_size * seq_length])])
self.loss = tf.reduce_mean(log_perps) / batch_size
# define trainer, saver, inits
self.train_op = tf.train.AdamOptimizer(learn_rate).minimize(self.loss)
self.init_op = tf.initialize_all_variables()
self.saver = tf.train.Saver()
def __init__(self,
sequence_length,
vocab_size,
embedding_size,
hidden_size,
layer_count=1,
**kw):
assert layer_count >= 1, "An LSTM cannot have less than one layer."
n_classes = kw.get('n_classes', 2) # >2 not tested.
self.input_x = tf.placeholder(tf.int32, [None, sequence_length],
name="input_x")
self.input_y = tf.placeholder(tf.float32, [None, n_classes],
name="input_y")
self.dropout_keep_prob = tf.placeholder(tf.float32,
name="dropout_keep_prob")
# Layer 1: Word embeddings
self.embeddings = tf.Variable(tf.random_uniform(
[vocab_size, embedding_size], -0.1, 0.1),
name="embeddings")
embedded_words = tf.nn.embedding_lookup(self.embeddings, self.input_x)
# Funnel the words into the LSTM.
# Current size: (batch_size, n_words, emb_dim)
# Want: [(batch_size, n_hidden) * n_words]
#
# Since otherwise there's no way to feed information into the LSTM cell.
# Yes, it's a bit confusing, because we want a batch of multiple
# sequences, with each step being of 'embedding_size'.
embedded_words = tf.transpose(embedded_words, [1, 0, 2])
embedded_words = tf.reshape(embedded_words, [-1, embedding_size])
# Note: 'tf.split' outputs a **Python** list.
embedded_words = tf.split(0, sequence_length, embedded_words)
# Layer 2: LSTM cell
lstm_use_peepholes = True
# 'state_is_tuple = True' should NOT be used despite the warnings
# (which appear as of TF 0.9), since it doesn't work on the version of
# TF installed on Euler (0.8).
if layer_count > 1:
print("Using deep {0}-layer LSTM with first layer size {1}"
" (embedding size) and hidden layer size {2}.".format(
layer_count, embedding_size, hidden_size))
print("First cell {0}->{1}".format(embedding_size, embedding_size))
first_cell = TextLSTM._cell(embedding_size, embedding_size,
lstm_use_peepholes,
self.dropout_keep_prob)
print("Second cell {0}->{1}".format(embedding_size, hidden_size))
second_cell = TextLSTM._cell(embedding_size, hidden_size,
lstm_use_peepholes,
self.dropout_keep_prob)
print("Third cell+ {0}->{1} (if applicable)".format(
hidden_size, hidden_size))
third_plus = TextLSTM._cell(hidden_size, hidden_size,
lstm_use_peepholes,
self.dropout_keep_prob)
deep_cells = [third_plus] * (layer_count - 2)
lstm_cells = rnn_cell.MultiRNNCell([first_cell, second_cell] +
deep_cells)
else:
print(
"Using simple 1-layer LSTM with hidden layer size {0}.".format(
hidden_size))
lstm_cells = rnn_cell.LSTMCell(num_units=hidden_size,
input_size=embedding_size,
forget_bias=1.0,
use_peepholes=lstm_use_peepholes)
# Q: Can't batches end up containing both positive and negative labels?
# Can the LSTM batch training deal with this?
#
# A: Yes. Each batch feeds each sentence into the LSTM, incurs the loss,
# and backpropagates the error separately. Each example in a bath
# is independent. Note that as opposed to language models, for
# instance, where we incur a loss for all outputs, in this case we
# only care about the final output of the RNN, since it doesn't make
# sense to classify incomplete tweets.
outputs, _states = rnn(lstm_cells,
inputs=embedded_words,
dtype=tf.float32)
# Layer 3: Final Softmax
out_weight = tf.Variable(tf.random_normal([hidden_size, n_classes]))
out_bias = tf.Variable(tf.random_normal([n_classes]))
with tf.name_scope("output"):
lstm_final_output = outputs[-1]
self.scores = tf.nn.xw_plus_b(lstm_final_output,
out_weight,
out_bias,
name="scores")
self.predictions = tf.nn.softmax(self.scores, name="predictions")
with tf.name_scope("loss"):
self.losses = tf.nn.softmax_cross_entropy_with_logits(
self.scores, self.input_y)
self.loss = tf.reduce_mean(self.losses, name="loss")
with tf.name_scope("accuracy"):
self.correct_pred = tf.equal(tf.argmax(self.predictions, 1),
tf.argmax(self.input_y, 1))
self.accuracy = tf.reduce_mean(tf.cast(self.correct_pred, "float"),
name="accuracy")
def __init__(self,
sess,
vocab_size,
cell_size,
embedding_size,
num_layer,
memory_size,
log_dir,
learning_rate=0.001,
momentum=0.9,
use_dropout=True,
l2_coef=1e-6):
with tf.name_scope("io"):
self.inputs = tf.placeholder(dtype=tf.int32,
shape=(None, None),
name="prev_words")
self.input_lens = tf.placeholder(dtype=tf.int32,
shape=(None, ),
name="sent_len")
self.labels = tf.placeholder(dtype=tf.int32,
shape=(None, ),
name="next_word")
self.keep_prob = tf.placeholder(dtype=tf.float32, name="keep_prob")
with variable_scope.variable_scope("word-embedding"):
max_sent_len = array_ops.shape(self.inputs)[1]
embedding = tf_helpers.weight_and_bias(vocab_size,
embedding_size,
"embedding_w",
include_bias=False)
input_embedding = embedding_ops.embedding_lookup(
embedding,
tf.squeeze(tf.reshape(self.inputs, [-1, 1]), squeeze_dims=[1]))
input_embedding = tf.reshape(input_embedding,
[-1, max_sent_len, embedding_size])
with variable_scope.variable_scope("rnn"):
# cell = rnn_cell.LSTMCell(cell_size, use_peepholes=True, state_is_tuple=True)
cell = tf_helpers.MemoryLSTMCell(cell_size,
memory_size,
use_peepholes=True)
if use_dropout:
cell = rnn_cell.DropoutWrapper(cell,
output_keep_prob=self.keep_prob)
if num_layer > 1:
cell = rnn_cell.MultiRNNCell([cell] * num_layer,
state_is_tuple=True)
# and enc_last_state will be same as the true last state
outputs, _ = tf.nn.dynamic_rnn(
cell,
input_embedding,
dtype=tf.float32,
sequence_length=self.input_lens,
)
# get the TRUE last outputs
last_outputs = tf.reduce_sum(
tf.mul(
outputs,
tf.expand_dims(
tf.one_hot(self.input_lens - 1, max_sent_len), -1)), 1)
proj_w, proj_b = tf_helpers.weight_and_bias(cell_size,
vocab_size,
"output_project",
include_bias=True)
self.logits = tf.matmul(last_outputs, proj_w) + proj_b
self.loss = tf.reduce_mean(
nn_ops.sparse_softmax_cross_entropy_with_logits(
self.logits, self.labels))
tf.scalar_summary("entropy_loss", self.loss)
tf.scalar_summary("perplexity", tf.exp(self.loss))
self.summary_op = tf.merge_all_summaries()
# weight decay
vars = tf.trainable_variables()
all_weights = []
for v in vars:
if "bias" not in v.name.lower():
all_weights.append(tf.nn.l2_loss(v))
print("adding l2 to %s" % v.name)
loss_l2 = tf.add_n(all_weights)
self.reg_loss = self.loss + l2_coef * loss_l2
# optimization
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
self.train_ops = optimizer.minimize(self.reg_loss)
train_log_dir = os.path.join(log_dir, "train")
valid_log_dir = os.path.join(log_dir, "valid")
print "Save summary to %s" % log_dir
self.train_summary_writer = tf.train.SummaryWriter(
train_log_dir, sess.graph)
self.valid_summary_writer = tf.train.SummaryWriter(
valid_log_dir, sess.graph)
self.saver = tf.train.Saver(tf.all_variables())
def __init__(self, sess, config, data_feed, log_dir):
vocab_size = len(data_feed.vocab)
self.data_feed = data_feed
with tf.name_scope("io"):
self.inputs = tf.placeholder(dtype=tf.int32,
shape=(None, None),
name="input_seq")
self.input_lens = tf.placeholder(dtype=tf.int32,
shape=(None, ),
name="seq_len")
self.da_labels = tf.placeholder(dtype=tf.int32,
shape=(None, ),
name="dialog_acts")
self.senti_labels = tf.placeholder(
dtype=tf.float32,
shape=(None, data_feed.feature_size[data_feed.SENTI_ID]),
name="sentiments")
self.learning_rate = tf.Variable(float(config.init_lr),
trainable=False)
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * config.lr_decay)
max_sent_len = array_ops.shape(self.inputs)[1]
batch_size = array_ops.shape(self.inputs)[0]
with variable_scope.variable_scope("word-embedding"):
embedding = tf.get_variable("embedding",
[vocab_size, config.embed_size],
dtype=tf.float32)
input_embedding = embedding_ops.embedding_lookup(
embedding,
tf.squeeze(tf.reshape(self.inputs, [-1, 1]), squeeze_dims=[1]))
input_embedding = tf.reshape(input_embedding,
[-1, max_sent_len, config.embed_size])
with variable_scope.variable_scope("rnn"):
if config.cell_type == "gru":
cell = rnn_cell.GRUCell(config.cell_size)
elif config.cell_type == "lstm":
cell = rnn_cell.LSTMCell(config.cell_size,
use_peepholes=False,
forget_bias=1.0)
elif config.cell_type == "rnn":
cell = rnn_cell.BasicRNNCell(config.cell_size)
else:
raise ValueError("unknown RNN type")
if config.keep_prob < 1.0:
cell = rnn_cell.DropoutWrapper(
cell,
output_keep_prob=config.keep_prob,
input_keep_prob=config.keep_prob)
if config.num_layer > 1:
cell = rnn_cell.MultiRNNCell([cell] * config.num_layer,
state_is_tuple=True)
# and enc_last_state will be same as the true last state
outputs, _ = tf.nn.dynamic_rnn(
cell,
input_embedding,
dtype=tf.float32,
sequence_length=self.input_lens,
)
# get the TRUE last outputs
last_outputs = tf.reduce_sum(
tf.mul(
outputs,
tf.expand_dims(
tf.one_hot(self.input_lens - 1, max_sent_len), -1)), 1)
self.dialog_acts = self.fnn(
last_outputs, data_feed.feature_size[data_feed.DA_ID], [100],
"dialog_act_fnn")
self.sentiments = self.fnn(
last_outputs, data_feed.feature_size[data_feed.SENTI_ID],
[100], "setiment_fnn")
self.loss = tf.reduce_sum(nn_ops.sparse_softmax_cross_entropy_with_logits(self.dialog_acts, self.da_labels)) \
+ tf.reduce_sum(nn_ops.softmax_cross_entropy_with_logits(self.sentiments, self.senti_labels))
self.loss /= tf.to_float(batch_size)
tf.scalar_summary("entropy_loss", self.loss)
self.summary_op = tf.merge_all_summaries()
# weight decay
tvars = tf.trainable_variables()
for v in tvars:
print("Trainable %s" % v.name)
# optimization
if config.op == "adam":
print("Use Adam")
optimizer = tf.train.AdamOptimizer(self.learning_rate)
elif config.op == "rmsprop":
print("Use RMSProp")
optimizer = tf.train.RMSPropOptimizer(self.learning_rate)
else:
print("Use SGD")
optimizer = tf.train.GradientDescentOptimizer(self.learning_rate)
grads, _ = tf.clip_by_global_norm(tf.gradients(self.loss, tvars),
config.grad_clip)
self.train_ops = optimizer.apply_gradients(zip(grads, tvars))
self.saver = tf.train.Saver(tf.all_variables(),
write_version=tf.train.SaverDef.V2)
if log_dir is not None:
train_log_dir = os.path.join(log_dir, "train")
print("Save summary to %s" % log_dir)
self.train_summary_writer = tf.train.SummaryWriter(
train_log_dir, sess.graph)