def get_model(feed_previous=False):
learning_rate = tf.Variable(float(init_learning_rate), trainable=False, dtype=tf.float32)
learning_rate_decay_op = learning_rate.assign(learning_rate * 0.9)
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in range(input_seq_len):
encoder_inputs.append(tf.placeholder(tf.int32, shape=[None], name="encoder{0}".format(i)))
for i in range(output_seq_len + 1):
decoder_inputs.append(tf.placeholder(tf.int32, shape=[None], name="decoder{0}".format(i)))
for i in range(output_seq_len):
target_weights.append(tf.placeholder(tf.float32, shape=[None], name="weight{0}".format(i)))
targets = [decoder_inputs[i + 1] for i in range(output_seq_len)]
cell = tf.contrib.rnn.BasicLSTMCell(size)
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=feed_previous,
dtype=tf.float32)
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
opt = tf.train.GradientDescentOptimizer(learning_rate)
update = opt.apply_gradients(opt.compute_gradients(loss))
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, update, saver, learning_rate_decay_op, learning_rate
def get_model(feed_previous=False):
"""构造模型
"""
learning_rate = tf.Variable(float(init_learning_rate),
trainable=False,
dtype=tf.float32)
learning_rate_decay_op = learning_rate.assign(learning_rate * 0.9)
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in range(input_seq_len):
encoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="encoder{0}".format(i)))
for i in range(output_seq_len + 1):
decoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="decoder{0}".format(i)))
for i in range(output_seq_len):
target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
# decoder_inputs左移一个时序作为targets
targets = [decoder_inputs[i + 1] for i in range(output_seq_len)]
# cell = tf.contrib.rnn.BasicLSTMCell(size)
dropout = 1
num_layers = 3
cell = tf.contrib.rnn.BasicLSTMCell(size)
cell = tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=dropout)
cell = tf.contrib.rnn.MultiRNNCell([cell] * num_layers) # 纵向上有两个LSTM
# 这里输出的状态我们不需要
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=feed_previous,
dtype=tf.float32)
# 计算加权交叉熵损失
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
# 梯度下降优化器
opt = tf.train.AdamOptimizer(learning_rate).minimize(loss)
# 优化目标：让loss最小化
# update = opt.apply_gradients(opt.compute_gradients(loss))
# 模型持久化
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, opt, saver, learning_rate_decay_op, learning_rate
def get_model(feed_previous=False):
"""
构造模型:seq2seq
feed_previous表示decoder_inputs是我们直接提供训练数据的输入，
还是用前一个RNNCell的输出映射出来的，如果feed_previous为True，
那么就是用前一个RNNCell的输出，并经过Wx+b线性变换成
"""
learning_rate = tf.Variable(float(init_learning_rate),
trainable=False,
dtype=tf.float32)
learning_rate_decay_op = learning_rate.assign(learning_rate * 0.9)
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in range(input_seq_len):
encoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="encoder{0}".format(i)))
for i in range(output_seq_len + 1):
decoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="decoder{0}".format(i)))
for i in range(output_seq_len):
target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
# decoder_inputs左移一个时序作为targets
targets = [decoder_inputs[i + 1] for i in range(output_seq_len)]
cell = tf.contrib.rnn.BasicLSTMCell(size)
# 这里输出的状态我们不需要
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=feed_previous,
dtype=tf.float32)
# 计算加权交叉熵损失
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
# 梯度下降优化器
opt = tf.train.GradientDescentOptimizer(learning_rate)
# 优化目标：让loss最小化
update = opt.apply_gradients(opt.compute_gradients(loss))
# 模型持久化
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, update, saver, learning_rate_decay_op, learning_rate
def EmbeddingAttentionSeq2SeqNoTuple(enc_inp, dec_inp, feed_previous):
cell = core_rnn_cell_impl.BasicLSTMCell(2, state_is_tuple=False)
return seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols,
num_decoder_symbols,
embedding_size=2,
feed_previous=feed_previous)
def GRUSeq2Seq(enc_inp, dec_inp):
cell = core_rnn_cell_impl.MultiRNNCell(
[core_rnn_cell_impl.GRUCell(24)] * 2, state_is_tuple=True)
return seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=classes,
num_decoder_symbols=classes,
embedding_size=24)
def EmbeddingAttentionSeq2SeqNoTuple(enc_inp, dec_inp, feed_previous):
cell = core_rnn_cell_impl.BasicLSTMCell(2, state_is_tuple=False)
return seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols,
num_decoder_symbols,
embedding_size=2,
feed_previous=feed_previous)
def seq2seq_f(encoder_inputs, decoder_inputs, do_decode):
return seq2seq.embedding_attention_seq2seq(encoder_inputs=encoder_inputs,
decoder_inputs=decoder_inputs,
cell=cell,
num_encoder_symbols=self.vocab_size,
num_decoder_symbols=self.vocab_size,
embedding_size=embedding_size,
output_projection=output_projection,
feed_previous=do_decode,
beam_search=beam_search,
beam_size=beam_size)
def GRUSeq2Seq(enc_inp, dec_inp):
cell = rnn_cell.MultiRNNCell(
[rnn_cell.GRUCell(24) for _ in range(2)], state_is_tuple=True)
return seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=classes,
num_decoder_symbols=classes,
embedding_size=24,
output_projection=(w, b))
def GRUSeq2Seq(enc_inp, dec_inp):
cell = rnn_cell.MultiRNNCell(
[rnn_cell.GRUCell(24) for _ in range(2)], state_is_tuple=True)
return seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=classes,
num_decoder_symbols=classes,
embedding_size=24,
output_projection=(w, b))
def get_model():
# 这个方法需要的参数分别是：inputs_tensor,decoder_tensor,cell,类似与vocab_size的symbols,虽然我不知道encoder_symbolsy有什么用
# 然后是embed_size，应该和cell的size一样，然后是需不需要softmax，decode_inputs是来自前面的RNNcell还是我们自己输入，最后是数
# 据类型
'''
embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs,
cell,
num_encoder_symbols,
num_decoder_symbols,
embedding_size,
num_heads=1,
output_projection=None,
feed_previous=False,
dtype=None,
scope=None,
initial_state_attention=False
)
'''
encoder_inputs = []
decoder_inputs = []
targets_weigh = []
for i in range(input_seq_len):
encoder_inputs.append(tf.placeholder(shape=[None],dtype=tf.int32,name="encoder{0}".format(i)))
for i in range(output_seq_len):
decoder_inputs.append(tf.placeholder(shape=[None],dtype=tf.int32,name="decode{0}".format(i)))
for i in range(output_seq_len):
targets_weigh.append(
tf.placeholder(shape=[None],dtype=tf.float32,name="weight{0}".format(i))
)
targets = [decoder_inputs[i] for i in range(1,output_seq_len)]
targets.append(np.zeros(shape=[2],dtype=np.int32))
cell = tf.nn.rnn_cell.BasicLSTMCell(size)
outputs,_ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs,
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=False,
dtype=tf.float32
)
loss = seq2seq.sequence_loss(
outputs,targets,targets_weigh
)
opt = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
update = opt.apply_gradients(opt.compute_gradients(loss))
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs,decoder_inputs,targets_weigh,outputs,loss,update,saver
pass
def get_model(feed_previous=False):
learning_rate = tf.Variable(float(init_learning_rate),
trainable=False,
dtype=tf.float32)
learning_rate_decay_op = learning_rate.assign(learning_rate * 0.9)
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in range(input_seq_len):
encoder_inputs.append(
tf.compat.v1.placeholder(tf.int32,
shape=[None],
name="encoder{0}".format(i)))
for i in range(output_seq_len + 1):
decoder_inputs.append(
tf.compat.v1.placeholder(tf.int32,
shape=[None],
name="decoder{0}".format(i)))
for i in range(output_seq_len):
target_weights.append(
tf.compat.v1.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
# decoder_inputs左移一个时序作为targets
targets = [decoder_inputs[i + 1] for i in range(output_seq_len)]
cell = tf.contrib.rnn.BasicLSTMCell(size)
# 这里输出的状态我们不需要
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
# 是一个(W, B)结构的tuple，W是shape为[output_size x num_decoder_symbols]的weight矩阵，B是shape为[num_decoder_symbols]的偏置向量
feed_previous=feed_previous,
dtype=tf.float32)
# 计算交叉熵损失
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
# 梯度下降优化器
opt = tf.compat.v1.train.GradientDescentOptimizer(learning_rate)
# 优化目标：让loss最小化
update = opt.apply_gradients(opt.compute_gradients(loss))
# 模型持久化,保存所有的变量
saver = tf.compat.v1.train.Saver(tf.compat.v1.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, update, saver, learning_rate_decay_op, learning_rate
def get_model(feed_previous=False):
"""构造模型
"""
learning_rate = tf.Variable(float(init_learning_rate),
trainable=False,
dtype=tf.float32)
learning_rate_decay_op = learning_rate.assign(learning_rate * 0.99)
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in range(input_seq_len):
encoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="encoder{0}".format(i)))
for i in range(output_seq_len + 1):
decoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="decoder{0}".format(i)))
for i in range(output_seq_len):
target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
# decoder_inputs左移一个时序作为targets
targets = [decoder_inputs[i + 1] for i in range(output_seq_len)]
cell = tf.contrib.rnn.LSTMCell(size)
# 这里输出的状态我们不需要
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=feed_previous,
dtype=tf.float32)
# 计算加权交叉熵损失
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
# 梯度下降优化器
update = tf.train.AdamOptimizer(learning_rate).minimize(loss)
# 模型持久化
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, update, saver, learning_rate_decay_op, learning_rate
def get_model(feed_previous=False):
"""构造模型
"""
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in xrange(input_seq_len):
encoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="encoder{0}".format(i)))
for i in xrange(output_seq_len + 1):
decoder_inputs.append(
tf.placeholder(tf.int32, shape=[None],
name="decoder{0}".format(i)))
for i in xrange(output_seq_len):
target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
# decoder_inputs左移一个时序作为targets
targets = [decoder_inputs[i + 1] for i in xrange(output_seq_len)]
cell = tf.contrib.rnn.BasicLSTMCell(size)
# 这里输出的状态我们不需要
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=feed_previous,
dtype=tf.float32)
# 计算加权交叉熵损失
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
# 梯度下降优化器
opt = tf.train.GradientDescentOptimizer(learning_rate)
# 优化目标：让loss最小化
update = opt.apply_gradients(opt.compute_gradients(loss))
# 模型持久化
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, update, saver
def get_model(feed_previous=False):
"""构造模型
"""
learning_rate = tf.Variable(float(init_learning_rate), trainable=False, dtype=tf.float32)
learning_rate_decay_op = learning_rate.assign(learning_rate * 0.9)
encoder_inputs = []
decoder_inputs = []
target_weights = []
for i in xrange(input_seq_len):
encoder_inputs.append(tf.placeholder(tf.int32, shape=[None], name="encoder{0}".format(i)))
for i in xrange(output_seq_len + 1):
decoder_inputs.append(tf.placeholder(tf.int32, shape=[None], name="decoder{0}".format(i)))
for i in xrange(output_seq_len):
target_weights.append(tf.placeholder(tf.float32, shape=[None], name="weight{0}".format(i)))
# decoder_inputs左移一个时序作为targets
targets = [decoder_inputs[i + 1] for i in xrange(output_seq_len)]
cell = tf.contrib.rnn.BasicLSTMCell(size)
# 这里输出的状态我们不需要
outputs, _ = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs[:output_seq_len],
cell,
num_encoder_symbols=num_encoder_symbols,
num_decoder_symbols=num_decoder_symbols,
embedding_size=size,
output_projection=None,
feed_previous=feed_previous,
dtype=tf.float32)
# 计算加权交叉熵损失
loss = seq2seq.sequence_loss(outputs, targets, target_weights)
# 梯度下降优化器
opt = tf.train.GradientDescentOptimizer(learning_rate)
# 优化目标：让loss最小化
update = opt.apply_gradients(opt.compute_gradients(loss))
# 模型持久化
saver = tf.train.Saver(tf.global_variables())
return encoder_inputs, decoder_inputs, target_weights, outputs, loss, update, saver, learning_rate_decay_op, learning_rate
def model(self,
mode="train",
num_layers=1,
cell_size=128,
cell_type="BasicLSTMCell",
embedding_size=20,
learning_rate=0.001,
tensorboard_verbose=0,
checkpoint_path=None):
'''
Build tensor specifying graph of operations for the seq2seq neural network model.
mode = string, either "train" or "predict"
cell_type = attribute of rnn_cell specifying which RNN cell type to use
cell_size = size for the hidden layer in the RNN cell
num_layers = number of RNN cell layers to use
Return TFLearn model instance. Use DNN model for this.
'''
assert mode in ["train", "predict"]
checkpoint_path = checkpoint_path or (
"%s%ss2s_checkpoint.tfl" %
(self.data_dir or "", "/" if self.data_dir else ""))
GO_VALUE = self.out_max_int + 1 # unique integer value used to trigger decoder outputs in the seq2seq RNN
tflearn.config.init_graph(seed=None,
log_device=False,
num_cores=int(cpu_count() * 2 / 3),
gpu_memory_fraction=0,
soft_placement=True)
network = tflearn.input_data(
shape=[None, self.in_seq_len + self.out_seq_len],
dtype=tf.int32,
name="XY")
encoder_inputs = tf.slice(network, [0, 0], [-1, self.in_seq_len],
name="enc_in") # get encoder inputs
encoder_inputs = tf.unstack(
encoder_inputs, axis=1
) # transform into list of self.in_seq_len elements, each [-1]
decoder_inputs = tf.slice(network, [0, self.in_seq_len],
[-1, self.out_seq_len],
name="dec_in") # get decoder inputs
decoder_inputs = tf.unstack(
decoder_inputs, axis=1
) # transform into list of self.out_seq_len elements, each [-1]
go_input = tf.multiply(
tf.ones_like(decoder_inputs[0], dtype=tf.int32), GO_VALUE
) # insert "GO" symbol as the first decoder input; drop the last decoder input
decoder_inputs = [
go_input
] + decoder_inputs[:self.out_seq_len -
1] # insert GO as first; drop last decoder input
feed_previous = not (mode == "train")
if self.verbose > 3:
print("feed_previous = %s" % str(feed_previous))
print("encoder inputs: %s" % str(encoder_inputs))
print("decoder inputs: %s" % str(decoder_inputs))
print("len decoder inputs: %s" % len(decoder_inputs))
self.n_input_symbols = self.in_max_int + 1 # default is integers from 0 to 9
self.n_output_symbols = self.out_max_int + 2 # extra "GO" symbol for decoder inputs
single_cell = getattr(tf.nn.rnn_cell, cell_type)(cell_size,
state_is_tuple=True)
if num_layers == 1:
print("rnn net later number:{}".format(num_layers))
cell = single_cell
else:
print("rnn net later number:{}".format(num_layers))
cell = rnn_cell.MultiRNNCell([single_cell] * num_layers)
if self.seq2seq_model == "embedding_rnn":
model_outputs, states = seq2seq.embedding_rnn_seq2seq(
encoder_inputs,
# encoder_inputs: A list of 2D Tensors [batch_size, input_size].
decoder_inputs,
cell,
num_encoder_symbols=self.n_input_symbols,
num_decoder_symbols=self.n_output_symbols,
embedding_size=embedding_size,
feed_previous=feed_previous)
elif self.seq2seq_model == "embedding_attention":
model_outputs, states = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
# encoder_inputs: A list of 2D Tensors [batch_size, input_size].
decoder_inputs,
cell,
num_encoder_symbols=self.n_input_symbols,
num_decoder_symbols=self.n_output_symbols,
embedding_size=embedding_size,
num_heads=1,
initial_state_attention=False,
feed_previous=feed_previous)
else:
raise Exception('[TFLearnSeq2Seq] Unknown seq2seq model %s' %
self.seq2seq_model)
tf.add_to_collection(
tf.GraphKeys.LAYER_VARIABLES + '/' + "seq2seq_model",
model_outputs) # for TFLearn to know what to save and restore
# model_outputs: list of the same length as decoder_inputs of 2D Tensors with shape [batch_size x output_size] containing the generated outputs.
if self.verbose > 2:
print("model outputs: %s" % model_outputs)
network = tf.stack(
model_outputs, axis=1
) # shape [-1, n_decoder_inputs (= self.out_seq_len), num_decoder_symbols]
if self.verbose > 2:
print("packed model outputs: %s" % network)
if self.verbose > 3:
all_vars = tf.get_collection(tf.GraphKeys.VARIABLES)
print("all_vars = %s" % all_vars)
with tf.name_scope(
"TargetsData"
): # placeholder for target variable (i.e. trainY input)
targetY = tf.placeholder(shape=[None, self.out_seq_len],
dtype=tf.int32,
name="Y")
network = tflearn.regression(network,
placeholder=targetY,
optimizer='adam',
learning_rate=learning_rate,
loss=self.sequence_loss,
metric=self.accuracy,
name="Y")
model = tflearn.DNN(network,
tensorboard_verbose=tensorboard_verbose,
checkpoint_path=checkpoint_path)
return model
def create_network(self):
self.seq2seq_model = "embedding_attention"
mode = "train"
GO_VALUE = self.out_max_int + 1
self.net = tflearn.input_data(shape=[None, self.in_seq_len],
dtype=tf.int32,
name="XY")
encoder_inputs = tf.slice(self.net, [0, 0], [-1, self.in_seq_len],
name="enc_in") # get encoder inputs
encoder_inputs = tf.unstack(
encoder_inputs,
axis=1) #transform to list of self.in_seq_len elements, each [-1]
decoder_inputs = tf.slice(self.net, [0, 0], [-1, self.out_seq_len],
name="dec_in")
decoder_inputs = tf.unstack(
decoder_inputs,
axis=1) # transform into list of self.out_seq_len elements
go_input = tf.multiply(tf.ones_like(decoder_inputs[0], dtype=tf.int32),
GO_VALUE)
decoder_inputs = [
go_input
] + decoder_inputs[:self.out_seq_len -
1] # insert GO as first; drop last decoder input
feed_previous = not (mode == "train")
self.n_input_symbols = self.in_max_int + 1 # default is integers from 0 to 9
self.n_output_symbols = self.out_max_int + 2 # extra "GO" symbol for decoder inputs
cell = rnn.MultiRNNCell([
rnn.GRUCell(128),
rnn.GRUCell(128),
rnn.GRUCell(128),
rnn.GRUCell(128),
rnn.GRUCell(128),
rnn.GRUCell(128),
rnn.GRUCell(128),
rnn.GRUCell(128)
])
if self.seq2seq_model == "embedding_rnn":
model_outputs, states = seq2seq.embedding_rnn_seq2seq(
encoder_inputs,
# encoder_inputs: A list of 2D Tensors [batch_size, input_size].
decoder_inputs,
cell,
num_encoder_symbols=self.n_input_symbols,
num_decoder_symbols=self.n_output_symbols,
embedding_size=200,
feed_previous=feed_previous)
elif self.seq2seq_model == "embedding_attention":
model_outputs, states = seq2seq.embedding_attention_seq2seq(
encoder_inputs,
# encoder_inputs: A list of 2D Tensors [batch_size, input_size].
decoder_inputs,
cell,
num_encoder_symbols=self.n_input_symbols,
num_decoder_symbols=self.n_output_symbols,
embedding_size=200,
num_heads=1,
initial_state_attention=False,
feed_previous=feed_previous)
else:
raise Exception('[TFLearnSeq2Seq] Unknown seq2seq model %s' %
self.seq2seq_model)
tf.add_to_collection(
tf.GraphKeys.LAYER_VARIABLES + '/' + "seq2seq_model",
model_outputs) # for TFLearn to know what to save and restore
self.net = tf.stack(
model_outputs, axis=1
) # shape [-1, n_decoder_inputs (= self.out_seq_len), num_decoder_symbols]
with tf.name_scope(
"TargetsData"
): # placeholder for target variable (i.e. trainY input)
targetY = tf.placeholder(shape=[None, self.out_seq_len],
dtype=tf.int32,
name="Y")
self.net = tflearn.regression(self.net,
placeholder=targetY,
optimizer='adam',
learning_rate=0.00005,
loss=self.sequence_loss,
metric=self.accuracy,
name="Y")
self.model = tflearn.DNN(self.net)
def model(self, mode="train", num_layers=1, cell_size=128, cell_type="BasicLSTMCell", embedding_size=20,
learning_rate=0.001,
tensorboard_verbose=0, checkpoint_path=None):
'''
Build tensor specifying graph of operations for the seq2seq neural network model.
mode = string, either "train" or "predict"
cell_type = attribute of rnn_cell specifying which RNN cell type to use
cell_size = size for the hidden layer in the RNN cell
num_layers = number of RNN cell layers to use
Return TFLearn model instance. Use DNN model for this.
'''
assert mode in ["train", "predict"]
checkpoint_path = checkpoint_path or (
"%s%ss2s_checkpoint.tfl" % (self.data_dir or "", "/" if self.data_dir else ""))
GO_VALUE = self.out_max_int + 1 # unique integer value used to trigger decoder outputs in the seq2seq RNN
tflearn.config.init_graph(seed=None, log_device=False, num_cores=int(cpu_count() * 2 / 3),
gpu_memory_fraction=0, soft_placement=True)
network = tflearn.input_data(shape=[None, self.in_seq_len + self.out_seq_len], dtype=tf.int32, name="XY")
encoder_inputs = tf.slice(network, [0, 0], [-1, self.in_seq_len], name="enc_in") # get encoder inputs
encoder_inputs = tf.unstack(encoder_inputs,
axis=1) # transform into list of self.in_seq_len elements, each [-1]
decoder_inputs = tf.slice(network, [0, self.in_seq_len], [-1, self.out_seq_len],
name="dec_in") # get decoder inputs
decoder_inputs = tf.unstack(decoder_inputs,
axis=1) # transform into list of self.out_seq_len elements, each [-1]
go_input = tf.multiply(tf.ones_like(decoder_inputs[0], dtype=tf.int32),
GO_VALUE) # insert "GO" symbol as the first decoder input; drop the last decoder input
decoder_inputs = [go_input] + decoder_inputs[
: self.out_seq_len - 1] # insert GO as first; drop last decoder input
feed_previous = not (mode == "train")
if self.verbose > 3:
print("feed_previous = %s" % str(feed_previous))
print("encoder inputs: %s" % str(encoder_inputs))
print("decoder inputs: %s" % str(decoder_inputs))
print("len decoder inputs: %s" % len(decoder_inputs))
self.n_input_symbols = self.in_max_int + 1 # default is integers from 0 to 9
self.n_output_symbols = self.out_max_int + 2 # extra "GO" symbol for decoder inputs
single_cell = getattr(tf.nn.rnn_cell, cell_type)(cell_size, state_is_tuple=True)
if num_layers == 1:
print("rnn net later number:{}".format(num_layers))
cell = single_cell
else:
print("rnn net later number:{}".format(num_layers))
cell = rnn_cell.MultiRNNCell([single_cell] * num_layers)
if self.seq2seq_model == "embedding_rnn":
model_outputs, states = seq2seq.embedding_rnn_seq2seq(encoder_inputs,
# encoder_inputs: A list of 2D Tensors [batch_size, input_size].
decoder_inputs,
cell,
num_encoder_symbols=self.n_input_symbols,
num_decoder_symbols=self.n_output_symbols,
embedding_size=embedding_size,
feed_previous=feed_previous)
elif self.seq2seq_model == "embedding_attention":
model_outputs, states = seq2seq.embedding_attention_seq2seq(encoder_inputs,
# encoder_inputs: A list of 2D Tensors [batch_size, input_size].
decoder_inputs,
cell,
num_encoder_symbols=self.n_input_symbols,
num_decoder_symbols=self.n_output_symbols,
embedding_size=embedding_size,
num_heads=1,
initial_state_attention=False,
feed_previous=feed_previous)
else:
raise Exception('[TFLearnSeq2Seq] Unknown seq2seq model %s' % self.seq2seq_model)
tf.add_to_collection(tf.GraphKeys.LAYER_VARIABLES + '/' + "seq2seq_model",
model_outputs) # for TFLearn to know what to save and restore
# model_outputs: list of the same length as decoder_inputs of 2D Tensors with shape [batch_size x output_size] containing the generated outputs.
if self.verbose > 2:
print("model outputs: %s" % model_outputs)
network = tf.stack(model_outputs, axis=1) # shape [-1, n_decoder_inputs (= self.out_seq_len), num_decoder_symbols]
if self.verbose > 2:
print("packed model outputs: %s" % network)
if self.verbose > 3:
all_vars = tf.get_collection(tf.GraphKeys.VARIABLES)
print("all_vars = %s" % all_vars)
with tf.name_scope("TargetsData"): # placeholder for target variable (i.e. trainY input)
targetY = tf.placeholder(shape=[None, self.out_seq_len], dtype=tf.int32, name="Y")
network = tflearn.regression(network,
placeholder=targetY,
optimizer='adam',
learning_rate=learning_rate,
loss=self.sequence_loss,
metric=self.accuracy,
name="Y")
model = tflearn.DNN(network, tensorboard_verbose=tensorboard_verbose, checkpoint_path=checkpoint_path)
return model
def testEmbeddingAttentionSeq2Seq(self):
with self.test_session() as sess:
with variable_scope.variable_scope(
"root", initializer=init_ops.constant_initializer(0.5)):
enc_inp = [
constant_op.constant(
1, dtypes.int32, shape=[2]) for i in range(2)
]
dec_inp = [
constant_op.constant(
i, dtypes.int32, shape=[2]) for i in range(3)
]
cell_fn = lambda: core_rnn_cell_impl.BasicLSTMCell(2)
cell = cell_fn()
dec, mem = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2)
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 5), res[0].shape)
res = sess.run([mem])
self.assertEqual((2, 2), res[0].c.shape)
self.assertEqual((2, 2), res[0].h.shape)
# Test with state_is_tuple=False.
with variable_scope.variable_scope("no_tuple"):
cell_fn = functools.partial(
core_rnn_cell_impl.BasicLSTMCell,
2, state_is_tuple=False)
cell_nt = cell_fn()
dec, mem = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell_nt,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2)
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 5), res[0].shape)
res = sess.run([mem])
self.assertEqual((2, 4), res[0].shape)
# Test externally provided output projection.
w = variable_scope.get_variable("proj_w", [2, 5])
b = variable_scope.get_variable("proj_b", [5])
with variable_scope.variable_scope("proj_seq2seq"):
dec, _ = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell_fn(),
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2,
output_projection=(w, b))
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 2), res[0].shape)
def testEmbeddingAttentionSeq2Seq(self):
with self.test_session() as sess:
with variable_scope.variable_scope(
"root", initializer=init_ops.constant_initializer(0.5)):
enc_inp = [
constant_op.constant(
1, dtypes.int32, shape=[2]) for i in range(2)
]
dec_inp = [
constant_op.constant(
i, dtypes.int32, shape=[2]) for i in range(3)
]
cell_fn = lambda: core_rnn_cell_impl.BasicLSTMCell(2)
cell = cell_fn()
dec, mem = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2)
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 5), res[0].shape)
res = sess.run([mem])
self.assertEqual((2, 2), res[0].c.shape)
self.assertEqual((2, 2), res[0].h.shape)
# Test with state_is_tuple=False.
with variable_scope.variable_scope("no_tuple"):
cell_fn = functools.partial(
core_rnn_cell_impl.BasicLSTMCell,
2, state_is_tuple=False)
cell_nt = cell_fn()
dec, mem = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell_nt,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2)
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 5), res[0].shape)
res = sess.run([mem])
self.assertEqual((2, 4), res[0].shape)
# Test externally provided output projection.
w = variable_scope.get_variable("proj_w", [2, 5])
b = variable_scope.get_variable("proj_b", [5])
with variable_scope.variable_scope("proj_seq2seq"):
dec, _ = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell_fn(),
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2,
output_projection=(w, b))
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 2), res[0].shape)
# add one more target
targets.append(tf.placeholder(dtype = tf.int32, shape = [None], name = 'last_target'))
target_weights = [tf.placeholder(dtype = tf.float32, shape = [None], name = 'target_w{}'.format(i)) for i in range(output_seq_len)]
# output projection
size = 512
w_t = tf.get_variable('proj_w', [en_vocab_size, size], tf.float32)
b = tf.get_variable('proj_b', [en_vocab_size], tf.float32)
w = tf.transpose(w_t)
output_projection = (w, b)
outputs, states = seq2seq_lib.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs,
BasicLSTMCell(size),
num_encoder_symbols = zh_vocab_size,
num_decoder_symbols = en_vocab_size,
embedding_size = 100,
feed_previous = False,
output_projection = output_projection,
dtype = tf.float32)
# define our loss function
def sampled_loss(labels, logits):
return tf.nn.sampled_softmax_loss(
weights = w_t,
biases = b,
labels = tf.reshape(labels, [-1, 1]),
inputs = logits,
num_sampled = 512,
num_classes = en_vocab_size)
def testEmbeddingAttentionSeq2Seq(self):
with self.test_session() as sess:
with variable_scope.variable_scope(
"root", initializer=init_ops.constant_initializer(0.5)):
enc_inp = [
constant_op.constant(
1, dtypes.int32, shape=[2]) for i in range(2)
]
dec_inp = [
constant_op.constant(
i, dtypes.int32, shape=[2]) for i in range(3)
]
cell = core_rnn_cell_impl.BasicLSTMCell(2, state_is_tuple=True)
dec, mem = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2)
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 5), res[0].shape)
res = sess.run([mem])
self.assertEqual((2, 2), res[0].c.shape)
self.assertEqual((2, 2), res[0].h.shape)
# Test with state_is_tuple=False.
with variable_scope.variable_scope("no_tuple"):
cell = core_rnn_cell_impl.BasicLSTMCell(2, state_is_tuple=False)
dec, mem = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2)
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 5), res[0].shape)
res = sess.run([mem])
self.assertEqual((2, 4), res[0].shape)
# Test externally provided output projection.
w = variable_scope.get_variable("proj_w", [2, 5])
b = variable_scope.get_variable("proj_b", [5])
with variable_scope.variable_scope("proj_seq2seq"):
dec, _ = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2,
output_projection=(w, b))
sess.run([variables.global_variables_initializer()])
res = sess.run(dec)
self.assertEqual(3, len(res))
self.assertEqual((2, 2), res[0].shape)
# Test that previous-feeding model ignores inputs after the first.
dec_inp2 = [
constant_op.constant(
0, dtypes.int32, shape=[2]) for _ in range(3)
]
with variable_scope.variable_scope("other"):
d3, _ = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp2,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2,
feed_previous=constant_op.constant(True))
sess.run([variables.global_variables_initializer()])
variable_scope.get_variable_scope().reuse_variables()
d1, _ = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2,
feed_previous=True)
d2, _ = seq2seq_lib.embedding_attention_seq2seq(
enc_inp,
dec_inp2,
cell,
num_encoder_symbols=2,
num_decoder_symbols=5,
embedding_size=2,
feed_previous=True)
res1 = sess.run(d1)
res2 = sess.run(d2)
res3 = sess.run(d3)
self.assertAllClose(res1, res2)
self.assertAllClose(res1, res3)