def seq2seq_f(lstm_inputs, decoder_inputs, seq_length, do_decode):
num_hidden = attn_num_layers * attn_num_hidden
lstm_fw_cell = BasicLSTMCell(num_hidden, forget_bias=0.0, state_is_tuple=False)
# Backward direction cell
lstm_bw_cell = BasicLSTMCell(num_hidden, forget_bias=0.0, state_is_tuple=False)
pre_encoder_inputs, output_state_fw, output_state_bw = tf.contrib.rnn.static_bidirectional_rnn(lstm_fw_cell, lstm_bw_cell, lstm_inputs,
initial_state_fw=None, initial_state_bw=None,
dtype=tf.float32, sequence_length=None, scope=None)
encoder_inputs = [e*f for e,f in zip(pre_encoder_inputs,encoder_masks[:seq_length])]
top_states = [array_ops.reshape(e, [-1, 1, num_hidden*2])
for e in encoder_inputs]
attention_states = array_ops.concat(top_states, 1)
initial_state = tf.concat(axis=1, values=[output_state_fw, output_state_bw])
outputs, _, attention_weights_history = embedding_attention_decoder(
decoder_inputs, initial_state, attention_states, cell,
num_symbols=target_vocab_size,
embedding_size=target_embedding_size,
num_heads=1,
output_size=target_vocab_size,
output_projection=None,
feed_previous=do_decode,
initial_state_attention=False,
attn_num_hidden = attn_num_hidden)
return outputs, attention_weights_history
def __init__(self, config):
self._config = config
self._kernel_size = 3
self._train_data, _ = self.get_video_data()
self._nr_training_examples = self._train_data.shape[0]
self._train_data = self.shuffle_train_data(self._train_data)
self._word_to_index, self._index_to_word, self._bias_init_vector, self._caption_matrix, self._longest_caption = \
self.get_caption_dicts(self._train_data)
self._nr_words = len(self._word_to_index)
self._W_emb = tf.get_variable(tf.random_uniform(
[self._nr_words, self._config.dim_hidden], -0.1, 0.1),
name='W_emb')
self._lstm = BasicLSTMCell(self._config.dim_hidden)
self._encode_image_W = tf.get_variable(tf.random_uniform(
[self._config.dim_video, self._config.dim_hidden], -0.1, 0.1),
name='encode_image_W')
self._encode_image_b = tf.get_variable(tf.zeros(
[self._config.dim_hidden]),
name='encode_image_b')
self._embed_word_W = tf.get_variable(tf.random_uniform(
[self._config.dim_hidden, self._nr_words], -0.1, 0.1),
name='embed_word_W')
self._embed_word_b = tf.get_variable(tf.zeros([self._nr_words]),
name='embed_word_b')
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * self._config.learning_rate_decay_factor)
def define_basic_lstm_cell(self):
#num_units = self.config.hidden_size * 2
lstm_cell = BasicLSTMCell(self.config.hidden_size, forget_bias=0.0)
if self.is_training and self.config.keep_prob < 1.0:
lstm_cell = tf.nn.rnn_cell.DropoutWrapper(
lstm_cell, output_keep_prob=self.config.keep_prob)
return tf.nn.rnn_cell.MultiRNNCell([lstm_cell] *
self.config.num_layers)
class RecurrentController(BaseController):
def network_vars(self):
self.lstm_cell = BasicLSTMCell(256)
self.state = self.lstm_cell.zero_state(self.batch_size, tf.float32)
def network_op(self, X, state):
X = tf.convert_to_tensor(X)
return self.lstm_cell(X, state)
def get_state(self):
return self.state
def update_state(self, new_state):
return tf.no_op()
def RNN(x, weights, biases):
# Prepare data shape to match `rnn` function requirements
# Current data input shape: (batch_size, n_steps, n_input)
# Required shape: 'n_steps' tensors list of shape (batch_size, n_input)
# Permuting batch_size and n_steps
x = tf.transpose(x, [1, 0, 2])
# Reshaping to (n_steps*batch_size, n_input)
x = tf.reshape(x, [-1, n_input])
# Split to get a list of 'n_steps' tensors of shape (batch_size, n_input)
# x = tf.split(x, n_steps, 0)
x = tf.split(0, n_steps, x)
# Define a lstm cell with tensorflow
lstm_cell = BasicLSTMCell(n_hidden, forget_bias=1.0)
# Get lstm cell output
outputs, states = tf.nn.rnn(lstm_cell, x, dtype=tf.float32)
# Linear activation, using rnn inner loop last output
return tf.matmul(outputs[-1], weights['out']) + biases['out']
def __init__(self, features, batch_size):
self.cell = BasicLSTMCell(features)
self.h = tf.Variable(tf.zeros([batch_size, features]), trainable=False)
self.c = tf.Variable(tf.zeros([batch_size, features]), trainable=False)
self.batch_size = batch_size
self.features = features
def network_vars(self):
self.lstm_cell = BasicLSTMCell(256)
self.state = self.lstm_cell.zero_state(self.batch_size, tf.float32)
def __init__(self, encoder_masks, encoder_inputs_tensor,
decoder_inputs,
target_weights,
target_vocab_size,
buckets,
target_embedding_size,
attn_num_layers,
attn_num_hidden,
forward_only,
use_gru):
"""Create the model.
Args:
source_vocab_size: size of the source vocabulary.
target_vocab_size: size of the target vocabulary.
buckets: a list of pairs (I, O), where I specifies maximum input length
that will be processed in that bucket, and O specifies maximum output
length. Training instances that have inputs longer than I or outputs
longer than O will be pushed to the next bucket and padded accordingly.
We assume that the list is sorted, e.g., [(2, 4), (8, 16)].
size: number of units in each layer of the model.
num_layers: number of layers in the model.
max_gradient_norm: gradients will be clipped to maximally this norm.
learning_rate: learning rate to start with.
learning_rate_decay_factor: decay learning rate by this much when needed.
use_lstm: if true, we use LSTM cells instead of GRU cells.
num_samples: number of samples for sampled softmax.
forward_only: if set, we do not construct the backward pass in the model.
"""
self.encoder_inputs_tensor = encoder_inputs_tensor
self.decoder_inputs = decoder_inputs
self.target_weights = target_weights
self.target_vocab_size = target_vocab_size
self.buckets = buckets
self.encoder_masks = encoder_masks
# Create the internal multi-layer cell for our RNN.
single_cell = BasicLSTMCell(attn_num_hidden, forget_bias=0.0, state_is_tuple=False)
if use_gru:
print("using GRU CELL in decoder")
single_cell = GRUCell(attn_num_hidden)
cell = single_cell
if attn_num_layers > 1:
cell = MultiRNNCell([single_cell] * attn_num_layers, state_is_tuple=False)
# The seq2seq function: we use embedding for the input and attention.
def seq2seq_f(lstm_inputs, decoder_inputs, seq_length, do_decode):
num_hidden = attn_num_layers * attn_num_hidden
lstm_fw_cell = BasicLSTMCell(num_hidden, forget_bias=0.0, state_is_tuple=False)
# Backward direction cell
lstm_bw_cell = BasicLSTMCell(num_hidden, forget_bias=0.0, state_is_tuple=False)
pre_encoder_inputs, output_state_fw, output_state_bw = tf.contrib.rnn.static_bidirectional_rnn(lstm_fw_cell, lstm_bw_cell, lstm_inputs,
initial_state_fw=None, initial_state_bw=None,
dtype=tf.float32, sequence_length=None, scope=None)
encoder_inputs = [e*f for e,f in zip(pre_encoder_inputs,encoder_masks[:seq_length])]
top_states = [array_ops.reshape(e, [-1, 1, num_hidden*2])
for e in encoder_inputs]
attention_states = array_ops.concat(top_states, 1)
initial_state = tf.concat(axis=1, values=[output_state_fw, output_state_bw])
outputs, _, attention_weights_history = embedding_attention_decoder(
decoder_inputs, initial_state, attention_states, cell,
num_symbols=target_vocab_size,
embedding_size=target_embedding_size,
num_heads=1,
output_size=target_vocab_size,
output_projection=None,
feed_previous=do_decode,
initial_state_attention=False,
attn_num_hidden = attn_num_hidden)
return outputs, attention_weights_history
# Our targets are decoder inputs shifted by one.
targets = [decoder_inputs[i + 1]
for i in xrange(len(decoder_inputs) - 1)]
softmax_loss_function = None # default to tf.nn.sparse_softmax_cross_entropy_with_logits
# Training outputs and losses.
if forward_only:
self.outputs, self.losses, self.attention_weights_histories = model_with_buckets(
encoder_inputs_tensor, decoder_inputs, targets,
self.target_weights, buckets, lambda x, y, z: seq2seq_f(x, y, z, True),
softmax_loss_function=softmax_loss_function)
else:
self.outputs, self.losses, self.attention_weights_histories = model_with_buckets(
encoder_inputs_tensor, decoder_inputs, targets,
self.target_weights, buckets, lambda x, y, z: seq2seq_f(x, y, z, False),
softmax_loss_function=softmax_loss_function)
for i in range(len(Questions_Types)):
Questions_Types_embedding.append(
embeddings_index[Questions_Types[i].lower()])
#Question Type Encoder
# Here I have to pass these Question types embeddings e_ti to the fully connected neural network to get the internal vector representation qt_i
Topic_internal_vector = FC(Questions_Types_embedding, 7)
# Number of hidden units.
from tensorflow.python.ops.rnn_cell import BasicLSTMCell
encoder_hidden_units = 600
decoder_hidden_units = 600
# BLSTM Encoder encoding the Question Topic
forward_topics_encoder_cell = BasicLSTMCell(encoder_hidden_units)
backward_topics_encoder_cell = BasicLSTMCell(encoder_hidden_units)
(
(
encoder_topics_fw_outputs, # Contains the outputs of the BLSTM.
encoder_topics_bw_outputs),
(
encoder_topics_fw_final_state, # Contains the last hidden state of the BLSTM.
encoder_topics_bw_final_state)) = (
tf.nn.bidirectional_dynamic_rnn(
cell_fw=forward_topics_encoder_cell,
cell_bw=backward_topics_encoder_cell,
inputs=Topics, # Topics
dtype=tf.float32,
time_major=True))
def _build_forward(self):
config = self.config
N, M, JX, JQ, VW, d, dc, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.hidden_size, \
config.char_emb_size, config.max_word_size
H = config.max_tree_height
x_mask = self.x > 0
q_mask = self.q > 0
tx_mask = self.tx > 0 # [N, M, H, JX]
with tf.variable_scope("char_emb"):
char_emb_mat = tf.get_variable("char_emb_mat", shape=[VC, dc], dtype='float')
Acx = tf.nn.embedding_lookup(char_emb_mat, self.cx) # [N, M, JX, W, dc]
Acq = tf.nn.embedding_lookup(char_emb_mat, self.cq) # [N, JQ, W, dc]
filter = tf.get_variable("filter", shape=[1, config.char_filter_height, dc, d], dtype='float')
bias = tf.get_variable("bias", shape=[d], dtype='float')
strides = [1, 1, 1, 1]
Acx = tf.reshape(Acx, [-1, JX, W, dc])
Acq = tf.reshape(Acq, [-1, JQ, W, dc])
xxc = tf.nn.conv2d(Acx, filter, strides, "VALID") + bias # [N*M, JX, W/filter_stride, d]
qqc = tf.nn.conv2d(Acq, filter, strides, "VALID") + bias # [N, JQ, W/filter_stride, d]
xxc = tf.reshape(tf.reduce_max(tf.nn.relu(xxc), 2), [-1, M, JX, d])
qqc = tf.reshape(tf.reduce_max(tf.nn.relu(qqc), 2), [-1, JQ, d])
with tf.variable_scope("word_emb"):
if config.mode == 'train':
word_emb_mat = tf.get_variable("word_emb_mat", dtype='float', shape=[VW, config.word_emb_size], initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat", shape=[VW, config.word_emb_size], dtype='float')
Ax = tf.nn.embedding_lookup(word_emb_mat, self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat, self.q) # [N, JQ, d]
# Ax = linear([Ax], d, False, scope='Ax_reshape')
# Aq = linear([Aq], d, False, scope='Aq_reshape')
xx = tf.concat(3, [xxc, Ax]) # [N, M, JX, 2d]
qq = tf.concat(2, [qqc, Aq]) # [N, JQ, 2d]
D = d + config.word_emb_size
with tf.variable_scope("pos_emb"):
pos_emb_mat = tf.get_variable("pos_emb_mat", shape=[config.pos_vocab_size, d], dtype='float')
Atx = tf.nn.embedding_lookup(pos_emb_mat, self.tx) # [N, M, H, JX, d]
cell = BasicLSTMCell(D, state_is_tuple=True)
cell = SwitchableDropoutWrapper(cell, self.is_train, input_keep_prob=config.input_keep_prob)
x_len = tf.reduce_sum(tf.cast(x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(q_mask, 'int32'), 1) # [N]
with tf.variable_scope("rnn"):
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(cell, cell, xx, x_len, dtype='float', scope='start') # [N, M, JX, 2d]
tf.get_variable_scope().reuse_variables()
(fw_us, bw_us), (_, (fw_u, bw_u)) = bidirectional_dynamic_rnn(cell, cell, qq, q_len, dtype='float', scope='start') # [N, J, d], [N, d]
u = (fw_u + bw_u) / 2.0
h = (fw_h + bw_h) / 2.0
with tf.variable_scope("h"):
no_op_cell = NoOpCell(D)
tree_rnn_cell = TreeRNNCell(no_op_cell, d, tf.reduce_max)
initial_state = tf.reshape(h, [N*M*JX, D]) # [N*M*JX, D]
inputs = tf.concat(4, [Atx, tf.cast(self.tx_edge_mask, 'float')]) # [N, M, H, JX, d+JX]
inputs = tf.reshape(tf.transpose(inputs, [0, 1, 3, 2, 4]), [N*M*JX, H, d + JX]) # [N*M*JX, H, d+JX]
length = tf.reshape(tf.reduce_sum(tf.cast(tx_mask, 'int32'), 2), [N*M*JX])
# length = tf.reshape(tf.reduce_sum(tf.cast(tf.transpose(tx_mask, [0, 1, 3, 2]), 'float'), 3), [-1])
h, _ = dynamic_rnn(tree_rnn_cell, inputs, length, initial_state=initial_state) # [N*M*JX, H, D]
h = tf.transpose(tf.reshape(h, [N, M, JX, H, D]), [0, 1, 3, 2, 4]) # [N, M, H, JX, D]
u = tf.expand_dims(tf.expand_dims(tf.expand_dims(u, 1), 1), 1) # [N, 1, 1, 1, 4d]
dot = linear(h * u, 1, True, squeeze=True, scope='dot') # [N, M, H, JX]
# self.logits = tf.reshape(dot, [N, M * H * JX])
self.logits = tf.reshape(exp_mask(dot, tx_mask), [N, M * H * JX]) # [N, M, H, JX]
self.yp = tf.reshape(tf.nn.softmax(self.logits), [N, M, H, JX])
def __init__(self, kwd_voc_size, *args, **kwargs):
BasicLSTMCell.__init__(self, *args, **kwargs)
self.key_words_voc_size = kwd_voc_size
in_onehot = tf.one_hot(in_ph, vocab_size, name="input_onehot")
inputs = tf.split(in_onehot, sequence_length, axis=1)
inputs = [tf.squeeze(input_, [1]) for input_ in inputs]
targets = tf.split(targ_ph, sequence_length, axis=1)
# at this point, inputs is a list of length sequence_length
# each element of inputs is [batch_size,vocab_size]
# targets is a list of length sequence_length
# each element of targets is a 1D vector of length batch_size
# ------------------
# YOUR COMPUTATION GRAPH HERE
cell0 = BasicLSTMCell(state_dim)
cell1 = BasicLSTMCell(state_dim)
multi_cell = tf.contrib.rnn.MultiRNNCell([cell0, cell1])
initial_state = multi_cell.zero_state(batch_size, tf.float32)
#######################################################################
#and add an encoder for the decoder...
seq_out = tf.contrib.legacy_seq2seq.rnn_decoder(inputs, initial_state,
multi_cell)
seq_out = tf.reshape(seq_out, (50, 50, 128, 1))
initializer = tf.contrib.layers.variance_scaling_initializer()
logits = tf.contrib.layers.fully_connected(seq_out,
vocab_size,
activation_fn=None,
weights_initializer=initializer,
biases_initializer=initializer)
#logits =
def _build_forward(self):
config = self.config
N, M, JX, JQ, VW, VC, d, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.char_vocab_size, config.hidden_size, \
config.max_word_size
JX = tf.shape(self.x)[2]
JQ = tf.shape(self.q)[1]
M = tf.shape(self.x)[1]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
with tf.variable_scope("emb"):
if config.use_word_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
if config.mode == 'train':
word_emb_mat = tf.get_variable(
"word_emb_mat",
dtype='float',
shape=[VW, dw],
initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat",
shape=[VW, dw],
dtype='float')
if config.use_glove_for_unk:
word_emb_mat = tf.concat(
0, [word_emb_mat, self.new_emb_mat])
print(word_emb_mat.get_shape().as_list())
with tf.name_scope("word"):
Ax = tf.nn.embedding_lookup(word_emb_mat,
self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat,
self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
xx = Ax
qq = Aq
# highway network
if config.highway:
with tf.variable_scope("highway"):
xx = highway_network(xx,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
tf.get_variable_scope().reuse_variables()
qq = highway_network(qq,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
cell = BasicLSTMCell(d, state_is_tuple=True)
d_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N]
with tf.variable_scope("prepro"):
(fw_u, bw_u), ((_, fw_u_f), (_,
bw_u_f)) = bidirectional_dynamic_rnn(
d_cell,
d_cell,
qq,
q_len,
dtype='float',
scope='u1') # [N, J, d], [N, d]
u = tf.concat(2, [fw_u, bw_u])
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='u1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='h1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
self.tensor_dict['u'] = u
self.tensor_dict['h'] = h
with tf.variable_scope("main"):
if config.dynamic_att:
p0 = h
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]),
[N * M, JQ, 2 * d])
q_mask = tf.reshape(
tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]),
[N * M, JQ])
first_cell = AttentionCell(
cell,
u,
mask=q_mask,
mapper='sim',
input_keep_prob=self.config.input_keep_prob,
is_train=self.is_train)
else:
p0, p1 = attention_layer(config,
self.is_train,
h,
u,
h_mask=self.x_mask,
u_mask=self.q_mask,
scope="p0",
tensor_dict=self.tensor_dict)
first_cell = d_cell
# with tf.variable_scope("activate"):
# p0 = tf.nn.relu(_linear(tf.reshape(p0,[-1,1200]),300,bias=0.01,bias_start=0.0,scope='relu'))
# if config.share_lstm_weights:
# tf.get_variable_scope().reuse_variables()
# p1 = tf.nn.relu(_linear(tf.reshape(p1,[-1,1200]),300,bias=0.01,bias_start=0.0,scope='relu'))
with tf.variable_scope('two_lstm'):
p0 = tf.reshape(p0, [N, 1, -1, 300])
p1 = tf.reshape(p1, [N, 1, -1, 300])
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, p0, x_len, dtype='float',
scope='g0') # [N, M, JX, 2d]
g0 = tf.concat(3, [fw_g0, bw_g0])
q_len_new = tf.tile(tf.expand_dims(q_len, 1), [1, M])
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(
first_cell,
first_cell,
p1,
q_len_new,
dtype='float',
scope='g0') # [N, M, JX, 2d]
g1 = tf.concat(3, [fw_g1, bw_g1])
# with tf.variable_scope('two_lstm_1'):
# (fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(first_cell, first_cell, g0, x_len, dtype='float', scope='g0') # [N, M, JX, 2d]
# g2 = tf.concat(3, [fw_g2, bw_g2])
# q_len_new = tf.tile(tf.expand_dims(q_len,1),[1,M])
# if config.share_lstm_weights:
# tf.get_variable_scope().reuse_variables()
# (fw_g3, bw_g3), _ = bidirectional_dynamic_rnn(first_cell, first_cell, g1, q_len_new, dtype='float', scope='g0') # [N, M, JX, 2d]
# g3 = tf.concat(3, [fw_g3, bw_g3])
g0 = tf.reduce_sum(tf.reduce_max(g0, 2), 1)
g1 = tf.reduce_sum(tf.reduce_max(g1, 2), 1)
logits = _linear([g0, g1, tf.abs(tf.subtract(g0, g1)), g0 * g1],
2,
bias=0.01,
bias_start=0.0,
scope='logits1')
flat_logits2 = tf.reshape(logits, [N, 2])
yp = tf.nn.softmax(flat_logits2) # [-1, M*JX]
self.tensor_dict['g0'] = g0
self.tensor_dict['g1'] = g1
self.logits = flat_logits2
self.yp = yp
def _build_forward(self):
config = self.config
N, M, JX, JQ, VW, VC, d, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.char_vocab_size, config.hidden_size, \
config.max_word_size
JX = tf.shape(self.x)[2]
JQ = tf.shape(self.q)[1]
M = tf.shape(self.x)[1]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
with tf.variable_scope("emb"):
if config.use_char_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
char_emb_mat = tf.get_variable("char_emb_mat",
shape=[VC, dc],
dtype='float')
with tf.variable_scope("char"):
Acx = tf.nn.embedding_lookup(char_emb_mat,
self.cx) # [N, M, JX, W, dc]
Acq = tf.nn.embedding_lookup(char_emb_mat,
self.cq) # [N, JQ, W, dc]
Acx = tf.reshape(Acx, [-1, JX, W, dc])
Acq = tf.reshape(Acq, [-1, JQ, W, dc])
filter_sizes = list(
map(int, config.out_channel_dims.split(',')))
heights = list(map(int, config.filter_heights.split(',')))
assert sum(filter_sizes) == dco
with tf.variable_scope("conv"):
xx = multi_conv1d(Acx,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
if config.share_cnn_weights:
tf.get_variable_scope().reuse_variables()
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
else:
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="qq")
xx = tf.reshape(xx, [-1, M, JX, dco])
qq = tf.reshape(qq, [-1, JQ, dco])
if config.use_word_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
if config.mode == 'train':
word_emb_mat = tf.get_variable(
"word_emb_mat",
dtype='float',
shape=[VW, dw],
initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat",
shape=[VW, dw],
dtype='float')
if config.use_glove_for_unk:
word_emb_mat = tf.concat(
0, [word_emb_mat, self.new_emb_mat])
with tf.name_scope("word"):
Ax = tf.nn.embedding_lookup(word_emb_mat,
self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat,
self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
if config.use_char_emb:
xx = tf.concat(3, [xx, Ax]) # [N, M, JX, di]
qq = tf.concat(2, [qq, Aq]) # [N, JQ, di]
else:
xx = Ax
qq = Aq
# highway network
if config.highway:
with tf.variable_scope("highway"):
xx = highway_network(xx,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
tf.get_variable_scope().reuse_variables()
qq = highway_network(qq,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
self.tensor_dict['xx'] = xx
self.tensor_dict['qq'] = qq
cell = BasicLSTMCell(d, state_is_tuple=True)
d_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N]
with tf.variable_scope("prepro"):
(fw_u, bw_u), ((_, fw_u_f), (_,
bw_u_f)) = bidirectional_dynamic_rnn(
d_cell,
d_cell,
qq,
q_len,
dtype='float',
scope='u1') # [N, J, d], [N, d]
u = tf.concat(2, [fw_u, bw_u])
if config.two_prepro_layers:
(fw_u, bw_u), ((_, fw_u_f),
(_, bw_u_f)) = bidirectional_dynamic_rnn(
d_cell,
d_cell,
u,
q_len,
dtype='float',
scope='u2') # [N, J, d], [N, d]
u = tf.concat(2, [fw_u, bw_u])
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='u1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
if config.two_prepro_layers:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, h, x_len, dtype='float',
scope='u2') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='h1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
if config.two_prepro_layers:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, h, x_len, dtype='float',
scope='h2') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
self.tensor_dict['u'] = u
self.tensor_dict['h'] = h
with tf.variable_scope("main"):
if config.dynamic_att:
p0 = h
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]),
[N * M, JQ, 2 * d])
q_mask = tf.reshape(
tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]),
[N * M, JQ])
first_cell = AttentionCell(
cell,
u,
mask=q_mask,
mapper='sim',
input_keep_prob=self.config.input_keep_prob,
is_train=self.is_train)
else:
p0 = attention_layer(config,
self.is_train,
h,
u,
h_mask=self.x_mask,
u_mask=self.q_mask,
scope="p0",
tensor_dict=self.tensor_dict)
first_cell = d_cell
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, p0, x_len, dtype='float',
scope='g0') # [N, M, JX, 2d]
g0 = tf.concat(3, [fw_g0, bw_g0])
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, g0, x_len, dtype='float',
scope='g1') # [N, M, JX, 2d]
g1 = tf.concat(3, [fw_g1, bw_g1])
if config.late:
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(
d_cell,
d_cell,
tf.concat(3, [g1, p0]),
x_len,
dtype='float',
scope='g2') # [N, M, JX, 2d]
g2 = tf.concat(3, [fw_g2, bw_g2])
# logits2 = u_logits(config, self.is_train, tf.concat(3, [g1, a1i]), u, h_mask=self.x_mask, u_mask=self.q_mask, scope="logits2")
logits = get_logits([g1, g2, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits1')
if config.feed_gt:
logy = tf.log(tf.cast(self.y, 'float') + VERY_SMALL_NUMBER)
logits = tf.cond(self.is_train, lambda: logy,
lambda: logits)
if config.feed_hard:
hard_yp = tf.argmax(tf.reshape(logits, [N, M * JX]), 1)
hard_logits = tf.reshape(tf.one_hot(hard_yp, M * JX),
[N, M, JX]) # [N, M, JX]
logits = tf.cond(self.is_train, lambda: logits,
lambda: hard_logits)
flat_logits = tf.reshape(logits, [-1, M * JX])
flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
yp = tf.reshape(flat_yp, [-1, M, JX])
logits2 = get_logits([g1, g2, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits2')
flat_logits2 = tf.reshape(logits2, [-1, M * JX])
flat_yp2 = tf.nn.softmax(flat_logits2)
yp2 = tf.reshape(flat_yp2, [-1, M, JX])
else:
logits = get_logits([g1, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits1')
a1i = softsel(tf.reshape(g1, [N, M * JX, 2 * d]),
tf.reshape(logits, [N, M * JX]))
if config.feed_gt:
logy = tf.log(tf.cast(self.y, 'float') + VERY_SMALL_NUMBER)
logits = tf.cond(self.is_train, lambda: logy,
lambda: logits)
if config.feed_hard:
hard_yp = tf.argmax(tf.reshape(logits, [N, M * JX]), 1)
hard_logits = tf.reshape(tf.one_hot(hard_yp, M * JX),
[N, M, JX]) # [N, M, JX]
logits = tf.cond(self.is_train, lambda: logits,
lambda: hard_logits)
flat_logits = tf.reshape(logits, [-1, M * JX])
flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
yp = tf.reshape(flat_yp, [-1, M, JX])
a1i = tf.tile(tf.expand_dims(tf.expand_dims(a1i, 1), 1),
[1, M, JX, 1])
yp_aug = tf.expand_dims(yp, -1)
g1yp = g1 * yp_aug
if config.prev_mode == 'a':
prev = a1i
elif config.prev_mode == 'y':
prev = yp_aug
elif config.prev_mode == 'gy':
prev = g1yp
else:
raise Exception()
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(
d_cell,
d_cell,
tf.concat(3, [p0, g1, prev, g1 * prev]),
x_len,
dtype='float',
scope='g2') # [N, M, JX, 2d]
g2 = tf.concat(3, [fw_g2, bw_g2])
# logits2 = u_logits(config, self.is_train, tf.concat(3, [g1, a1i]), u, h_mask=self.x_mask, u_mask=self.q_mask, scope="logits2")
logits2 = get_logits([g2, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits2')
flat_logits2 = tf.reshape(logits2, [-1, M * JX])
flat_yp2 = tf.nn.softmax(flat_logits2)
yp2 = tf.reshape(flat_yp2, [-1, M, JX])
self.tensor_dict['g1'] = g1
self.tensor_dict['g2'] = g2
self.logits = flat_logits
self.logits2 = flat_logits2
self.yp = yp
self.yp2 = yp2
inputs = [tf.squeeze(input_, [1]) for input_ in inputs]
targets = tf.split(1, sequence_length, targ_ph)
# at this point, inputs is a list of length sequence_length
# each element of inputs is [batch_size,vocab_size]
# targets is a list of length sequence_length
# each element of targets is a 1D vector of length batch_size
# ------------------
# YOUR COMPUTATION GRAPH HERE
with tf.variable_scope("Graph_") as scope:
# create a BasicLSTMCell
cell = BasicLSTMCell(state_dim, state_is_tuple=True)
# use it to create a MultiRNNCell
#tf.nn.rnn_cell.MultiRNNCell.__init__(cells, state_is_tuple=False)
stacked_lstm = tf.nn.rnn_cell.MultiRNNCell([cell] * 2, state_is_tuple=True)
#stacked_lstm = tf.nn.rnn_cell.MultiRNNCell([cell]*2)
# use it to create an initial_state
initial_state = stacked_lstm.zero_state(batch_size, tf.float32)
#note that initial_state will be a *list* of tensors!
# call seq2seq.rnn_decoder
# rnn_decoder(decoder_inputs, initial_state, cell, loop_function=None, scope=None):
#outputs, state = tf.nn.seq2seq.rnn_decoder(inputs, initial_state, stacked_lstm, loop_function=None, scope=None)
outputs, state = tf.nn.seq2seq.rnn_decoder(inputs,
initial_state,
stacked_lstm,
loop_function=None,
def __init__(self,
source_vocab_size,
target_vocab_size,
buckets,
size,
num_layers,
max_gradient_norm,
batch_size,
learning_rate,
learning_rate_decay_factor,
use_lstm=True,
num_samples=512,
forward_only=False):
"""Create the model.
Args:
source_vocab_size: size of the source vocabulary.
target_vocab_size: size of the target vocabulary.
buckets: a list of pairs (I, O), where I specifies maximum input length
that will be processed in that bucket, and O specifies maximum output
length. Training instances that have inputs longer than I or outputs
longer than O will be pushed to the next bucket and padded accordingly.
We assume that the list is sorted, e.g., [(2, 4), (8, 16)].
size: number of units in each layer of the model.
num_layers: number of layers in the model.
max_gradient_norm: gradients will be clipped to maximally this norm.
batch_size: the size of the batches used during training;
the model construction is independent of batch_size, so it can be
changed after initialization if this is convenient, e.g., for decoding.
learning_rate: learning rate to start with.
learning_rate_decay_factor: decay learning rate by this much when needed.
use_lstm: if true, we use LSTM cells instead of GRU cells.
num_samples: number of samples for sampled softmax.
forward_only: if set, we do not construct the backward pass in the model.
"""
self.source_vocab_size = source_vocab_size
self.target_vocab_size = target_vocab_size
self.buckets = buckets
self.batch_size = batch_size
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)
self.global_step = tf.Variable(0, trainable=False)
# If we use sampled softmax, we need an output projection.
output_projection = None
softmax_loss_function = None
# Sampled softmax only makes sense if we sample less than vocabulary size.
if num_samples > 0 and num_samples < self.target_vocab_size:
w = tf.get_variable("proj_w", [size, self.target_vocab_size])
w_t = tf.transpose(w)
b = tf.get_variable("proj_b", [self.target_vocab_size])
output_projection = (w, b)
def sampled_loss(inputs, labels):
labels = tf.reshape(labels, [-1, 1])
return tf.nn.sampled_softmax_loss(w_t, b, inputs, labels,
num_samples,
self.target_vocab_size)
softmax_loss_function = sampled_loss
# Create the internal multi-layer cell for our RNN.
output_keep_prob = tf.constant(0.8)
single_cell = GRUCell(size)
# Add dropout layer for regularization.
single_cell = DropoutWrapper(single_cell,
output_keep_prob=output_keep_prob)
if use_lstm:
single_cell = BasicLSTMCell(size)
single_cell = DropoutWrapper(single_cell,
output_keep_prob=output_keep_prob)
cell = single_cell
if num_layers > 1:
cell = MultiRNNCell([single_cell] * num_layers)
def seq2seq_f(encoder_inputs, decoder_inputs, do_decode):
return tf.nn.seq2seq.embedding_attention_seq2seq(
encoder_inputs,
decoder_inputs,
cell,
num_encoder_symbols=source_vocab_size,
num_decoder_symbols=target_vocab_size,
embedding_size=size,
output_projection=output_projection,
feed_previous=do_decode)
self.encoder_inputs = []
self.decoder_inputs = []
self.target_weights = []
for i in xrange(buckets[-1][0]): # Last bucket is the biggest one.
self.encoder_inputs.append(
tf.placeholder(tf.int32,
shape=[None],
name="encoder{0}".format(i)))
for i in xrange(buckets[-1][1] + 1):
self.decoder_inputs.append(
tf.placeholder(tf.int32,
shape=[None],
name="decoder{0}".format(i)))
self.target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
targets = [
self.decoder_inputs[i + 1]
for i in xrange(len(self.decoder_inputs) - 1)
]
if forward_only:
self.outputs, self.losses = tf.nn.seq2seq.model_with_buckets(
self.encoder_inputs,
self.decoder_inputs,
targets,
self.target_weights,
buckets,
lambda x, y: seq2seq_f(x, y, True),
softmax_loss_function=softmax_loss_function)
if output_projection is not None:
for b in xrange(len(buckets)):
self.outputs[b] = [
tf.matmul(output, output_projection[0]) +
output_projection[1] for output in self.outputs[b]
]
else:
self.outputs, self.losses = tf.nn.seq2seq.model_with_buckets(
self.encoder_inputs,
self.decoder_inputs,
targets,
self.target_weights,
buckets,
lambda x, y: seq2seq_f(x, y, False),
softmax_loss_function=softmax_loss_function)
params = tf.trainable_variables()
if not forward_only:
self.gradient_norms = []
self.updates = []
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
for b in xrange(len(buckets)):
gradients = tf.gradients(self.losses[b], params)
clipped_gradients, norm = tf.clip_by_global_norm(
gradients, max_gradient_norm)
self.gradient_norms.append(norm)
self.updates.append(
opt.apply_gradients(zip(clipped_gradients, params),
global_step=self.global_step))
self.saver = tf.train.Saver(tf.all_variables())
def _build_forward(self):
config = self.config
N = config.batch_size
M = config.max_num_sents
JX = config.max_sent_size
JQ = config.max_ques_size
VW = config.word_vocab_size
VC = config.char_vocab_size
W = config.max_word_size
d = config.hidden_size
JX = tf.shape(self.x)[2] # JX max sentence size, length,
JQ = tf.shape(self.q)[1] # JQ max questions size, length, is the
M = tf.shape(self.x)[1] # m is the max number of sentences
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
# dc = 8, each char will be map to 8-number vector, "char-level word embedding size [100]"
with tf.variable_scope("emb"):
if config.use_char_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
char_emb_mat = tf.get_variable("char_emb_mat",
shape=[VC, dc],
dtype='float')
# 330,8 a matrix for each char to its 8-number vector
with tf.variable_scope("char"):
Acx = tf.nn.embedding_lookup(char_emb_mat, self.cx)
# [N, M, JX, W, dc] 60,None,None,16,8, batch-size,
# N is the number of batch_size
# M the max number of sentences
# JX is the max sentence length
# W is the max length of a word
# dc is the vector for each char
# map each char to a vector
Acq = tf.nn.embedding_lookup(char_emb_mat,
self.cq) # [N, JQ, W, dc]
# JQ the max length of question
# W the max length of words
# mao each char in questiosn to vectors
Acx = tf.reshape(Acx, [-1, JX, W, dc])
Acq = tf.reshape(Acq, [-1, JQ, W, dc])
# max questions size, length, max_word_size(16), char_emb_size(8)
filter_sizes = list(
map(int, config.out_channel_dims.split(',')))
heights = list(map(int, config.filter_heights.split(',')))
# so here, there are 100 filters and the size of each filter is 5
# different heights and there are different number of these filter, but here just 100 5-long filters
assert sum(filter_sizes) == dco, (filter_sizes, dco)
with tf.variable_scope("conv"):
xx = multi_conv1d(Acx,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
if config.share_cnn_weights:
tf.get_variable_scope().reuse_variables()
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
else:
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="qq")
xx = tf.reshape(xx, [-1, M, JX, dco])
qq = tf.reshape(
qq, [-1, JQ, dco
]) # here, xx and qq are the output of cnn,
if config.use_word_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
if config.mode == 'train':
word_emb_mat = tf.get_variable(
"word_emb_mat",
dtype='float',
shape=[VW, dw],
initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat",
shape=[VW, dw],
dtype='float')
if config.use_glove_for_unk: # create a new word embedding or use the glove?
word_emb_mat = tf.concat(
[word_emb_mat, self.new_emb_mat], 0)
with tf.name_scope("word"):
Ax = tf.nn.embedding_lookup(word_emb_mat,
self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat,
self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
if config.use_char_emb:
xx = tf.concat([xx, Ax], 3) # [N, M, JX, di]
qq = tf.concat([qq, Aq], 2) # [N, JQ, di]
else:
xx = Ax
qq = Aq # here we used cnn and word embedding represented each word with a 200-unit vector
# so for, xx, (batch_size, sentence#, word#, embedding), qq (batch_size, word#, embedding)
# highway network
if config.highway:
with tf.variable_scope("highway"):
xx = highway_network(xx,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
tf.get_variable_scope().reuse_variables()
qq = highway_network(qq,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
self.tensor_dict['xx'] = xx
self.tensor_dict['qq'] = qq
# same shape with line 173
cell = BasicLSTMCell(
d, state_is_tuple=True) # d = 100, hidden state number
d_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'),
2) # [N, M], [60,?]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N] [60]
# masks are true and false, here, he sums up those truths,
with tf.variable_scope("prepro"):
(fw_u, bw_u), ((_, fw_u_f), (_,
bw_u_f)) = bidirectional_dynamic_rnn(
d_cell,
d_cell,
qq,
q_len,
dtype='float',
scope='u1') # [N, J, d], [N, d]
u = tf.concat(
[fw_u, bw_u],
2) # (60, ?, 200) | 200 becahse combined 2 100 hidden states
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='u1') # [N, M, JX, 2d]
h = tf.concat([fw_h, bw_h], 3) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='h1') # [N, M, JX, 2d]
h = tf.concat([fw_h, bw_h], 3) # [N, M, JX, 2d]
self.tensor_dict['u'] = u # [60, ?, 200] for question
self.tensor_dict['h'] = h # [60, ?, ?, 200] for article
with tf.variable_scope("main"):
if config.dynamic_att: # todo what is this dynamic attention.
p0 = h
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]),
[N * M, JQ, 2 * d])
q_mask = tf.reshape(
tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]),
[N * M, JQ])
first_cell = AttentionCell(
cell,
u,
mask=q_mask,
mapper='sim',
input_keep_prob=self.config.input_keep_prob,
is_train=self.is_train)
else:
p0 = attention_layer(config,
self.is_train,
h,
u,
h_mask=self.x_mask,
u_mask=self.q_mask,
scope="p0",
tensor_dict=self.tensor_dict)
cell2 = BasicLSTMCell(
d, state_is_tuple=True) # d = 100, hidden state number
first_cell = SwitchableDropoutWrapper(
cell2,
self.is_train,
input_keep_prob=config.input_keep_prob)
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(
first_cell,
first_cell,
inputs=p0,
sequence_length=x_len,
dtype='float',
scope='g0') # [N, M, JX, 2d]
g0 = tf.concat([fw_g0, bw_g0], 3)
cell3 = BasicLSTMCell(
d, state_is_tuple=True) # d = 100, hidden state number
first_cell3 = SwitchableDropoutWrapper(
cell3, self.is_train, input_keep_prob=config.input_keep_prob)
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(
first_cell3, first_cell3, g0, x_len, dtype='float',
scope='g1') # [N, M, JX, 2d]
g1 = tf.concat([fw_g1, bw_g1], 3)
logits = get_logits([g1, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits1')
a1i = softsel(tf.reshape(g1, [N, M * JX, 2 * d]),
tf.reshape(logits, [N, M * JX]))
a1i = tf.tile(tf.expand_dims(tf.expand_dims(a1i, 1), 1),
[1, M, JX, 1])
cell4 = BasicLSTMCell(
d, state_is_tuple=True) # d = 100, hidden state number
first_cell4 = SwitchableDropoutWrapper(
cell4, self.is_train, input_keep_prob=config.input_keep_prob)
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(
first_cell4,
first_cell4,
tf.concat([p0, g1, a1i, g1 * a1i], 3),
x_len,
dtype='float',
scope='g2') # [N, M, JX, 2d]
g2 = tf.concat([fw_g2, bw_g2], 3)
logits2 = get_logits([g2, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits2')
flat_logits = tf.reshape(logits, [-1, M * JX])
flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
yp = tf.reshape(flat_yp, [-1, M, JX])
flat_logits2 = tf.reshape(logits2, [-1, M * JX])
flat_yp2 = tf.nn.softmax(flat_logits2)
yp2 = tf.reshape(flat_yp2, [-1, M, JX])
self.tensor_dict['g1'] = g1
self.tensor_dict['g2'] = g2
self.logits = flat_logits
self.logits2 = flat_logits2
self.yp = yp
self.yp2 = yp2
def decode(self, knowledge_rep, masks, is_train, hparams):
"""
takes in a knowledge representation
and output a probability estimation over
all paragraph tokens on which token should be
the start of the answer span, and which should be
the end of the answer span.
:param knowledge_rep: it is a representation of the paragraph and question,
decided by how you choose to implement the encoder
:return:
"""
p0 = knowledge_rep
p_mask, q_mask = masks
batch_size = hparams.batch_size
input_keep_prob = hparams.input_keep_prob
p_len = tf.reduce_sum(tf.cast(p_mask, 'int32'), 1) # [N]
q_len = tf.reduce_sum(tf.cast(q_mask, 'int32'), 1) # [N]
JX = tf.shape(p_mask)[1]
with tf.variable_scope("main"):
cell = BasicLSTMCell(self.state_size, state_is_tuple=True)
first_cell = SwitchableDropoutWrapper(cell, is_train, input_keep_prob=input_keep_prob)
# [N, JX, 2d]
(fw_g0, bw_g0), _ = _bidirectional_dynamic_rnn(first_cell, first_cell, p0,
p_len, dtype='float', scope='g0')
g0 = tf.concat([fw_g0, bw_g0], 2)
cell = BasicLSTMCell(self.state_size, state_is_tuple=True)
first_cell = SwitchableDropoutWrapper(cell, is_train, input_keep_prob=input_keep_prob)
# [N, JX, 2d]
(fw_g1, bw_g1), _ = _bidirectional_dynamic_rnn(first_cell, first_cell, g0,
p_len, dtype='float', scope='g1')
g1 = tf.concat([fw_g1, bw_g1], 2)
logits = linear_logits([g1, p0], self.state_size, 0.0, scope='logits1',
mask=p_mask, is_train=is_train)
# TODO use batch _size
a1i = softsel(tf.reshape(g1, [batch_size, JX, 2 * self.state_size]),
tf.reshape(logits, [batch_size, JX]))
a1i = tf.tile(tf.expand_dims(a1i, 1), [1, JX, 1])
flat_logits1 = tf.reshape(logits, [-1, JX])
flat_yp = tf.nn.softmax(flat_logits1) # [-1, M*JX]
yp1 = tf.reshape(flat_yp, [-1, JX])
cell = BasicLSTMCell(self.state_size, state_is_tuple=True)
d_cell = SwitchableDropoutWrapper(cell, is_train, input_keep_prob=input_keep_prob)
# [N, M, JX, 2d]
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(d_cell, d_cell,
tf.concat([p0, g1, a1i, g1 * a1i], 2),
p_len, dtype='float', scope='g2')
g2 = tf.concat([fw_g2, bw_g2], 2)
logits2 = linear_logits([g2, p0], self.state_size, 0.0, scope='logits2',
mask=p_mask, is_train=is_train)
flat_logits2 = tf.reshape(logits2, [-1, JX])
flat_yp = tf.nn.softmax(flat_logits2) # [-1, M*JX]
yp2 = tf.reshape(flat_yp, [-1, JX])
return (yp1, flat_logits1), (yp2, flat_logits2)
def encode(self, inputs, masks, encoder_state_input, is_train, hparams):
"""
In a generalized encode function, you pass in your inputs,
masks, and an initial
hidden state input into this function.
:param inputs: Symbolic representations of your input
:param masks: this is to make sure tf.nn.dynamic_rnn doesn't iterate
through masked steps
:param encoder_state_input: (Optional) pass this as initial hidden state
to tf.nn.dynamic_rnn to build conditional representations
:return: an encoded representation of your input.
It can be context-level representation, word-level representation,
or both.
"""
context_embed, question_embed = inputs
p_mask, q_mask = masks
batch_size = hparams.batch_size,
input_keep_prob = hparams.input_keep_prob
cell = BasicLSTMCell(self.size, state_is_tuple=True)
d_cell = SwitchableDropoutWrapper(cell, is_train, input_keep_prob=input_keep_prob)
p_len = tf.reduce_sum(tf.cast(p_mask, 'int32'), 1) # [N]
q_len = tf.reduce_sum(tf.cast(q_mask, 'int32'), 1) # [N]
with tf.variable_scope('prepro'):
# [N, J, d], [N, d]
(fw_u, bw_u), ((_, fw_u_f), (_, bw_u_f)) = bidirectional_dynamic_rnn(
d_cell, d_cell, question_embed, q_len, dtype='float', scope='u1')
tf.get_variable_scope().reuse_variables()
# [N, JX, 2d]
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(cell, cell, context_embed,
p_len, dtype='float', scope='u1')
u = tf.concat([fw_u, bw_u], 2)
h = tf.concat([fw_h, bw_h], 2)
# Attention Layer
with tf.variable_scope('attention_layer'):
JQ = tf.shape(u)[1]
JX = tf.shape(h)[1]
h_aug = tf.tile(tf.expand_dims(h, 2), [1, 1, JQ, 1])
u_aug = tf.tile(tf.expand_dims(u, 1), [1, JX, 1, 1])
h_mask_aug = tf.tile(tf.expand_dims(p_mask, 2), [1, 1, JQ])
u_mask_aug = tf.tile(tf.expand_dims(q_mask, 1), [1, JX, 1])
hu_mask = tf.cast(h_mask_aug, tf.bool) & tf.cast(u_mask_aug, tf.bool)
hu_aug = h_aug * u_aug
u_logits = linear_logits([h_aug, u_aug, hu_aug], True, scope='u_logits',
mask=hu_mask, is_train=is_train)
u_a = softsel(u_aug, u_logits) # [N, JX, d]
h_a = softsel(h, tf.reduce_max(u_logits, 2)) # [N, d]
h_a = tf.tile(tf.expand_dims(h_a, 1), [1, JX, 1])
p0 = tf.concat([h, u_a, h * u_a, h * h_a], 2)
return p0
inputs = tf.split(1, sequence_length, in_onehot)
inputs = [tf.squeeze(input_, [1]) for input_ in inputs]
targets = tf.split(1, sequence_length, targ_ph)
# at this point, inputs is a list of length sequence_length
# each element of inputs is [batch_size,vocab_size]
# targets is a list of length sequence_length
# each element of targets is a 1D vector of length batch_size
# ------------------
# YOUR COMPUTATION GRAPH HERE
# create a BasicLSTMCell
cell = BasicLSTMCell(state_dim)
stacked_lstm = tf.nn.rnn_cell.MultiRNNCell([cell] * num_layers,
state_is_tuple=True)
initial_state = stacked_lstm.zero_state(batch_size, tf.float32)
# call seq2seq.rnn_decoder
outputs, final_state = tf.nn.seq2seq.rnn_decoder(inputs, initial_state,
stacked_lstm)
# transform the list of state outputs to a list of logits.
W = tf.Variable(tf.truncated_normal([state_dim, vocab_size], stddev=0.1))
# use a linear transformation.
logits = [tf.matmul(i, W) for i in outputs]
def __init__(self, num_units, forget_bias=1.0, input_size=None):
BasicLSTMCell.__init__(self,
num_units,
forget_bias=forget_bias,
input_size=input_size)
self.matrix, self.bias = None, None
def _build_forward(self):
#config为预先配置好的参数等
config = self.config
N, M, JX, JQ, VW, VC, d, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.char_vocab_size, config.hidden_size, \
config.max_word_size
JX = tf.shape(self.x)[2]
JQ = tf.shape(self.q)[1]
M = tf.shape(self.x)[1]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
#嵌入层
with tf.variable_scope("emb"):
#字符嵌入层
if config.use_char_emb: #若需要字符嵌入层
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
char_emb_mat = tf.get_variable("char_emb_mat",
shape=[VC, dc],
dtype='float')
with tf.variable_scope("char"):
Acx = tf.nn.embedding_lookup(char_emb_mat,
self.cx) # [N, M, JX, W, dc]
Acq = tf.nn.embedding_lookup(char_emb_mat,
self.cq) # [N, JQ, W, dc]
Acx = tf.reshape(Acx, [-1, JX, W, dc])
Acq = tf.reshape(Acq, [-1, JQ, W, dc])
#CNN的滤波器参数
filter_sizes = list(
map(int, config.out_channel_dims.split(',')))
heights = list(map(int, config.filter_heights.split(',')))
assert sum(filter_sizes) == dco, (filter_sizes, dco)
with tf.variable_scope("conv"):
xx = multi_conv1d(Acx,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
if config.share_cnn_weights:
tf.get_variable_scope().reuse_variables()
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
else:
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="qq")
xx = tf.reshape(xx, [-1, M, JX, dco])
qq = tf.reshape(qq, [-1, JQ, dco])
#词嵌入层
if config.use_word_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
if config.mode == 'train':
word_emb_mat = tf.get_variable(
"word_emb_mat",
dtype='float',
shape=[VW, dw],
initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat",
shape=[VW, dw],
dtype='float')
if config.use_glove_for_unk: #若调用已训练好的词嵌入文件
word_emb_mat = tf.concat(
0, [word_emb_mat, self.new_emb_mat])
with tf.name_scope("word"):
#将文章主体context：x和问题query：q转换为词向量
#embedding_lookup(params, ids),根据ids寻找params中的第id行
Ax = tf.nn.embedding_lookup(word_emb_mat,
self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat,
self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
if config.use_char_emb: #若进行了字符嵌入，在指定维度上将字符嵌入和词嵌入进行拼接
xx = tf.concat(3, [xx, Ax]) # [N, M, JX, di]
qq = tf.concat(2, [qq, Aq]) # [N, JQ, di]
else:
xx = Ax
qq = Aq
# 经过两层highway network得到context vector∈ R^(d*T)和query vectorQ∈R^(d∗J)
if config.highway:
with tf.variable_scope("highway"):
xx = highway_network(xx,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
tf.get_variable_scope().reuse_variables()
qq = highway_network(qq,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
self.tensor_dict['xx'] = xx
self.tensor_dict['qq'] = qq
cell = BasicLSTMCell(d, state_is_tuple=True)
#SwitchableDropoutWrapper为自定义的DropoutWrapper类
d_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
#reduce_sum在指定的维度上求和（得到x和q的非空值总数），cast将输入的tensor映射到指定类型（此处为x_mask到int32）
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N]
#Contextual Embedding Layer：对上一层得到的X和Q分别使用BiLSTM进行处理，分别捕捉X和Q中各自单词间的局部关系
with tf.variable_scope("prepro"):
(fw_u, bw_u), ((_, fw_u_f), (_,
bw_u_f)) = bidirectional_dynamic_rnn(
d_cell,
d_cell,
qq,
q_len,
dtype='float',
scope='u1') # [N, J, d], [N, d]
#fw_u和bw_u分别为双向lstm的output
u = tf.concat(2, [fw_u, bw_u]) #[N, J, 2d]
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='u1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='h1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
self.tensor_dict['u'] = u
self.tensor_dict['h'] = h
#核心层Attention Flow Layer
with tf.variable_scope("main"):
if config.dynamic_att:
p0 = h
#expand_dims()在矩阵指定位置增加维度
#tile()对矩阵的指定维度进行复制
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]), [
N * M, JQ, 2 * d
]) #先在索引1的位置添加一个维度，然后复制M（context中最多的sentence数量）次，使u和h能具有相同的维度
q_mask = tf.reshape(
tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]),
[N * M, JQ])
first_cell = AttentionCell(
cell,
u,
mask=q_mask,
mapper='sim',
input_keep_prob=self.config.input_keep_prob,
is_train=self.is_train)
else:
p0 = attention_layer(config,
self.is_train,
h,
u,
h_mask=self.x_mask,
u_mask=self.q_mask,
scope="p0",
tensor_dict=self.tensor_dict)
first_cell = d_cell
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, p0, x_len, dtype='float',
scope='g0') # [N, M, JX, 2d]
g0 = tf.concat(3, [fw_g0, bw_g0])
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, g0, x_len, dtype='float',
scope='g1') # [N, M, JX, 2d]
g1 = tf.concat(3, [fw_g1, bw_g1])
logits = get_logits([g1, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits1')
a1i = softsel(tf.reshape(g1, [N, M * JX, 2 * d]),
tf.reshape(logits, [N, M * JX]))
a1i = tf.tile(tf.expand_dims(tf.expand_dims(a1i, 1), 1),
[1, M, JX, 1])
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(
d_cell,
d_cell,
tf.concat(3, [p0, g1, a1i, g1 * a1i]),
x_len,
dtype='float',
scope='g2') # [N, M, JX, 2d]
g2 = tf.concat(3, [fw_g2, bw_g2])
logits2 = get_logits([g2, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits2')
flat_logits = tf.reshape(logits, [-1, M * JX])
flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
yp = tf.reshape(flat_yp, [-1, M, JX])
flat_logits2 = tf.reshape(logits2, [-1, M * JX])
flat_yp2 = tf.nn.softmax(flat_logits2)
yp2 = tf.reshape(flat_yp2, [-1, M, JX])
self.tensor_dict['g1'] = g1
self.tensor_dict['g2'] = g2
self.logits = flat_logits
self.logits2 = flat_logits2
self.yp = yp
self.yp2 = yp2
def __init__(self, num_units, cell_type='lstm', scope=None):
self.cell_fw = GRUCell(
num_units) if cell_type == 'gru' else BasicLSTMCell(num_units)
self.cell_bw = GRUCell(
num_units) if cell_type == 'gru' else BasicLSTMCell(num_units)
self.scope = scope or "bi_rnn"
def _build_forward(self):
config = self.config
N, M, JX, JQ, VW, VC, d, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.char_vocab_size, config.hidden_size, \
config.max_word_size
JX = tf.shape(self.x)[2]
JQ = tf.shape(self.q)[1]
M = tf.shape(self.x)[1]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
with tf.variable_scope("emb"):
if config.use_char_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
char_emb_mat = tf.get_variable("char_emb_mat",
shape=[VC, dc],
dtype='float')
with tf.variable_scope("char"):
Acx = tf.nn.embedding_lookup(char_emb_mat,
self.cx) # [N, M, JX, W, dc]
Acq = tf.nn.embedding_lookup(char_emb_mat,
self.cq) # [N, JQ, W, dc]
Acx = tf.reshape(Acx, [-1, JX, W, dc])
Acq = tf.reshape(Acq, [-1, JQ, W, dc])
filter_sizes = list(
map(int, config.out_channel_dims.split(',')))
heights = list(map(int, config.filter_heights.split(',')))
assert sum(filter_sizes) == dco, (filter_sizes, dco)
with tf.variable_scope("conv"):
xx = multi_conv1d(Acx,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
if config.share_cnn_weights:
tf.get_variable_scope().reuse_variables()
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
else:
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="qq")
xx = tf.reshape(xx, [-1, M, JX, dco])
qq = tf.reshape(qq, [-1, JQ, dco])
if config.use_word_emb:
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
if config.mode == 'train':
word_emb_mat = tf.get_variable(
"word_emb_mat",
dtype='float',
shape=[VW, dw],
initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat",
shape=[VW, dw],
dtype='float')
if config.use_glove_for_unk:
word_emb_mat = tf.concat(
0, [word_emb_mat, self.new_emb_mat])
with tf.name_scope("word"):
Ax = tf.nn.embedding_lookup(word_emb_mat,
self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat,
self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
if config.use_char_emb:
xx = tf.concat(3, [xx, Ax]) # [N, M, JX, di]
qq = tf.concat(2, [qq, Aq]) # [N, JQ, di]
else:
xx = Ax
qq = Aq
# highway network
if config.highway:
with tf.variable_scope("highway"):
xx = highway_network(xx,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
tf.get_variable_scope().reuse_variables()
qq = highway_network(qq,
config.highway_num_layers,
True,
wd=config.wd,
is_train=self.is_train)
self.tensor_dict['xx'] = xx
self.tensor_dict['qq'] = qq
cell = BasicLSTMCell(d, state_is_tuple=True)
d_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N]
with tf.variable_scope("prepro"):
(fw_u, bw_u), ((_, fw_u_f), (_,
bw_u_f)) = bidirectional_dynamic_rnn(
d_cell,
d_cell,
qq,
q_len,
dtype='float',
scope='u1') # [N, J, d], [N, d]
u = tf.concat(2, [fw_u, bw_u])
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='u1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='h1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
self.tensor_dict['u'] = u
self.tensor_dict['h'] = h
with tf.variable_scope("main"):
if config.dynamic_att:
p0 = h
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]),
[N * M, JQ, 2 * d])
q_mask = tf.reshape(
tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]),
[N * M, JQ])
first_cell = AttentionCell(
cell,
u,
mask=q_mask,
mapper='sim',
input_keep_prob=self.config.input_keep_prob,
is_train=self.is_train)
else:
p0 = attention_layer(config,
self.is_train,
h,
u,
h_mask=self.x_mask,
u_mask=self.q_mask,
scope="p0",
tensor_dict=self.tensor_dict)
first_cell = d_cell
self.p = p0
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, p0, x_len, dtype='float',
scope='g0') # [N, M, JX, 2d]
g0 = tf.concat(3, [fw_g0, bw_g0])
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, g0, x_len, dtype='float',
scope='g1') # [N, M, JX, 2d]
g1 = tf.concat(3, [fw_g1, bw_g1])
with tf.variable_scope("output"):
if config.model_name == "basic":
logits = get_logits([g1, p0], d, True, wd=config.wd, \
input_keep_prob=config.input_keep_prob,
mask=self.x_mask, is_train=self.is_train, \
func=config.answer_func, scope='logits1')
a1i = softsel(tf.reshape(g1, [N, M * JX, 2 * d]), \
tf.reshape(logits, [N, M * JX]))
a1i = tf.tile(tf.expand_dims(tf.expand_dims(a1i, 1), 1), \
[1, M, JX, 1])
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(d_cell, d_cell, \
tf.concat(3, [p0, g1, a1i, g1 * a1i]),
x_len, dtype='float', scope='g2') # [N, M, JX, 2d]
g2 = tf.concat(3, [fw_g2, bw_g2])
logits2 = get_logits([g2, p0], d, True, wd=config.wd, \
input_keep_prob=config.input_keep_prob, mask=self.x_mask,
is_train=self.is_train, func=config.answer_func,
scope='logits2')
flat_logits = tf.reshape(logits, [-1, M * JX])
flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
yp = tf.reshape(flat_yp, [-1, M, JX])
flat_logits2 = tf.reshape(logits2, [-1, M * JX])
flat_yp2 = tf.nn.softmax(flat_logits2)
yp2 = tf.reshape(flat_yp2, [-1, M, JX])
self.tensor_dict['g1'] = g1
self.tensor_dict['g2'] = g2
self.logits = flat_logits
self.logits2 = flat_logits2
self.yp = yp
self.yp2 = yp2
elif config.model_name == "basic-class":
C = 3 if config.data_dir.startswith('data/snli') else 2
(fw_g2, bw_g2) = (fw_g1, bw_g1)
if config.classifier == 'maxpool':
g2 = tf.concat(3, [fw_g2, bw_g2]) # [N, M, JX, 2d]
g2 = tf.reduce_max(g2, 2) # [N, M, 2d]
g2_dim = 2 * d
elif config.classifier == 'sumpool':
g2 = tf.concat(3, [fw_g2, bw_g2])
g2 = tf.reduce_sum(g2, 2)
g2_dim = 2 * d
else:
fw_g2_ = tf.gather(tf.transpose(fw_g2, [2, 0, 1, 3]),
JX - 1)
bw_g2_ = tf.gather(tf.transpose(bw_g2, [2, 0, 1, 3]), 0)
g2 = tf.concat(2, [fw_g2_, bw_g2_])
g2_dim = 2 * d
g2_ = tf.reshape(g2, [N, g2_dim])
logits0 = linear(g2_,
C,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
is_train=self.is_train,
scope='classifier')
flat_yp0 = tf.nn.softmax(logits0)
yp0 = tf.reshape(flat_yp0, [N, M, C])
self.tensor_dict['g1'] = g1
self.logits0 = logits0
self.yp0 = yp0
self.logits = logits0
self.yp = yp0
def construct_model(images,
actions=None,
states=None,
iter_num=-1.0,
k=-1,
use_state=True,
num_masks=10,
stp=False,
cdna=True,
dna=False,
context_frames=2,
pix_distributions=None,
conf=None):
"""Build convolutional lstm video predictor using STP, CDNA, or DNA.
Args:
images: tensor of ground truth image sequences
actions: tensor of action sequences
states: tensor of ground truth state sequences
iter_num: tensor of the current training iteration (for sched. sampling)
k: constant used for scheduled sampling. -1 to feed in own prediction.
use_state: True to include state and action in prediction
num_masks: the number of different pixel motion predictions (and
the number of masks for each of those predictions)
stp: True to use Spatial Transformer Predictor (STP)
cdna: True to use Convoluational Dynamic Neural Advection (CDNA)
dna: True to use Dynamic Neural Advection (DNA)
context_frames: number of ground truth frames to pass in before
feeding in own predictions
pix_distrib: the initial one-hot distriubtion for designated pixels
Returns:
gen_images: predicted future image frames
gen_states: predicted future states
Raises:
ValueError: if more than one network option specified or more than 1 mask
specified for DNA model.
"""
if 'dna_size' in conf.keys():
DNA_KERN_SIZE = conf['dna_size']
else:
DNA_KERN_SIZE = 5
print 'constructing network with less layers...'
if stp + cdna + dna != 1:
raise ValueError('More than one, or no network option specified.')
batch_size, img_height, img_width, color_channels = images[0].get_shape(
)[0:4]
batch_size = int(batch_size)
lstm_func = basic_conv_lstm_cell
# Generated robot states and images.
gen_states, gen_images, gen_masks, inf_low_state, pred_low_state = [], [], [], [], []
current_state = states[0]
gen_pix_distrib = []
summaries = []
if k == -1:
feedself = True
else:
# Scheduled sampling:
# Calculate number of ground-truth frames to pass in.
num_ground_truth = tf.to_int32(
tf.round(
tf.to_float(batch_size) * (k / (k + tf.exp(iter_num / k)))))
feedself = False
# LSTM state sizes and states.
lstm_size = np.int32(np.array([16, 32, 64, 100, 10]))
lstm_state1, lstm_state2, lstm_state3 = None, None, None
single_lstm1 = BasicLSTMCell(lstm_size[3], state_is_tuple=True)
single_lstm2 = BasicLSTMCell(lstm_size[4], state_is_tuple=True)
low_dim_lstm = MultiRNNCell([single_lstm1, single_lstm2],
state_is_tuple=True)
low_dim_lstm_state = low_dim_lstm.zero_state(batch_size, tf.float32)
dim_low_state = int(lstm_size[-1])
t = -1
for image, action in zip(images[:-1], actions[:-1]):
t += 1
print 'building timestep ', t
# Reuse variables after the first timestep.
reuse = bool(gen_images)
done_warm_start = len(gen_images) > context_frames - 1
with slim.arg_scope([
lstm_func, slim.layers.conv2d, slim.layers.fully_connected,
tf_layers.layer_norm, slim.layers.conv2d_transpose
],
reuse=reuse):
if feedself and done_warm_start:
# Feed in generated image.
prev_image = gen_images[-1]
if pix_distributions != None:
prev_pix_distrib = gen_pix_distrib[-1]
elif done_warm_start:
# Scheduled sampling
prev_image = scheduled_sample(image, gen_images[-1],
batch_size, num_ground_truth)
else:
# Always feed in ground_truth
prev_image = image
if pix_distributions != None:
prev_pix_distrib = pix_distributions[t]
prev_pix_distrib = tf.expand_dims(prev_pix_distrib, -1)
# Predicted state is always fed back in
state_action = tf.concat(1, [action, current_state]) # 6x
import pdb
pdb.set_trace()
enc0 = slim.layers.conv2d( #32x32x32
prev_image,
32,
kernel_size=[5, 5],
stride=2,
scope='scale1_conv1',
normalizer_fn=tf_layers.layer_norm,
normalizer_params={'scope': 'layer_norm1'})
hidden1, lstm_state1 = lstm_func( #32x32
enc0, lstm_state1, lstm_size[0], scope='state1')
hidden1 = tf_layers.layer_norm(hidden1, scope='layer_norm2')
enc1 = slim.layers.conv2d( #16x16
hidden1,
hidden1.get_shape()[3], [3, 3],
stride=2,
scope='conv2')
hidden2, lstm_state2 = lstm_func( #16x16x32
enc1, lstm_state2, lstm_size[1], scope='state3')
hidden2 = tf_layers.layer_norm(hidden2, scope='layer_norm4')
enc2 = slim.layers.conv2d( #8x8x32
hidden2,
hidden2.get_shape()[3], [3, 3],
stride=2,
scope='conv3')
# Pass in state and action.
smear = tf.reshape(
state_action,
[batch_size, 1, 1,
int(state_action.get_shape()[1])])
smear = tf.tile( #8x8x6
smear,
[1, int(enc2.get_shape()[1]),
int(enc2.get_shape()[2]), 1])
if use_state:
enc2 = tf.concat(3, [enc2, smear])
enc3 = slim.layers.conv2d( #8x8x32
enc2,
hidden2.get_shape()[3], [1, 1],
stride=1,
scope='conv4')
hidden3, lstm_state3 = lstm_func( #8x8x64
enc3, lstm_state3, lstm_size[2], scope='state5') # last 8x8
hidden3 = tf_layers.layer_norm(hidden3, scope='layer_norm6')
enc3 = slim.layers.conv2d( # 8x8x32
hidden3, 16, [1, 1], stride=1, scope='conv5')
enc3_flat = tf.reshape(enc3, [batch_size, -1])
if 'use_low_dim_lstm' in conf:
with tf.variable_scope('low_dim_lstm', reuse=reuse):
hidden4, low_dim_lstm_state = low_dim_lstm(
enc3_flat, low_dim_lstm_state)
low_dim_state = hidden4
else:
enc_fully1 = slim.layers.fully_connected(enc3_flat,
400,
scope='enc_fully1')
enc_fully2 = slim.layers.fully_connected(enc_fully1,
100,
scope='enc_fully2')
low_dim_state = enc_fully2
# inferred low dimensional state:
inf_low_state.append(low_dim_state)
pred_low_state.append(project_fwd_lowdim(low_dim_state))
smear = tf.reshape(low_dim_state,
[batch_size, 1, 1, dim_low_state])
smear = tf.tile( # 8x8xdim_hidden_state
smear,
[1, int(enc2.get_shape()[1]),
int(enc2.get_shape()[2]), 1])
enc4 = slim.layers.conv2d_transpose( #16x16x32
smear,
hidden3.get_shape()[3],
3,
stride=2,
scope='convt1')
enc5 = slim.layers.conv2d_transpose( #32x32x32
enc4,
enc0.get_shape()[3],
3,
stride=2,
scope='convt2')
enc6 = slim.layers.conv2d_transpose( #64x64x16
enc5,
16,
3,
stride=2,
scope='convt3',
normalizer_fn=tf_layers.layer_norm,
normalizer_params={'scope': 'layer_norm9'})
# Using largest hidden state for predicting untied conv kernels.
enc7 = slim.layers.conv2d_transpose(enc6,
DNA_KERN_SIZE**2,
1,
stride=1,
scope='convt4')
# Only one mask is supported (more should be unnecessary).
if num_masks != 1:
raise ValueError('Only one mask is supported for DNA model.')
transformed = [dna_transformation(prev_image, enc7, DNA_KERN_SIZE)]
if 'use_masks' in conf:
masks = slim.layers.conv2d_transpose(enc6,
num_masks + 1,
1,
stride=1,
scope='convt7')
masks = tf.reshape(
tf.nn.softmax(tf.reshape(masks, [-1, num_masks + 1])), [
int(batch_size),
int(img_height),
int(img_width), num_masks + 1
])
mask_list = tf.split(3, num_masks + 1, masks)
output = mask_list[0] * prev_image
for layer, mask in zip(transformed, mask_list[1:]):
output += layer * mask
else:
mask_list = None
output = transformed
gen_images.append(output)
gen_masks.append(mask_list)
if dna and pix_distributions != None:
transf_distrib = [
dna_transformation(prev_pix_distrib, enc7, DNA_KERN_SIZE)
]
if pix_distributions != None:
pix_distrib_output = mask_list[0] * prev_pix_distrib
mult_list = []
for i in range(num_masks):
mult_list.append(transf_distrib[i] * mask_list[i + 1])
pix_distrib_output += mult_list[i]
gen_pix_distrib.append(pix_distrib_output)
# pred_low_state_stopped = tf.stop_gradient(pred_low_state)
state_enc1 = slim.layers.fully_connected(
# pred_low_state[-1],
low_dim_state,
100,
scope='state_enc1')
state_enc2 = slim.layers.fully_connected(
state_enc1,
# int(current_state.get_shape()[1]),
4,
scope='state_enc2',
activation_fn=None)
current_state = tf.squeeze(state_enc2)
gen_states.append(current_state)
if pix_distributions != None:
return gen_images, gen_states, gen_masks, gen_pix_distrib, inf_low_state, pred_low_state
else:
return gen_images, gen_states, gen_masks, None, inf_low_state, pred_low_state
def __init__(self, model_parameters, training_parameters, directories,
**kwargs):
""" Initialization of the RNN Model as TensorFlow computational graph
"""
self.model_parameters = model_parameters
self.training_parameters = training_parameters
self.directories = directories
# Define model hyperparameters Tensors
with tf.name_scope("Parameters"):
self.learning_rate = tf.placeholder(tf.float32,
name="learning_rate")
self.momentum = tf.placeholder(tf.float32, name="momentum")
self.input_keep_probability = tf.placeholder(
tf.float32, name="input_keep_probability")
self.output_keep_probability = tf.placeholder(
tf.float32, name="output_keep_probability")
self.is_training = tf.placeholder(tf.bool)
# Define input, output and initialization Tensors
with tf.name_scope("Input"):
self.inputs = tf.placeholder("float", [
None, self.model_parameters.sequence_length,
self.model_parameters.input_dimension
],
name='input_placeholder')
self.targets = tf.placeholder(
"float", [None, self.model_parameters.sequence_length, 1],
name='labels_placeholder')
self.init = tf.placeholder(
tf.float32,
shape=[None, self.model_parameters.state_size],
name="init")
# Define the TensorFlow RNN computational graph
with tf.name_scope("LSTMRNN_RNN"):
cells = []
# Define the layers
for _ in range(self.model_parameters.n_layers):
if self.model_parameters.model == 'rnn':
cell = BasicRNNCell(self.model_parameters.state_size)
elif self.model_parameters.model == 'gru':
cell = GRUCell(self.model_parameters.state_size)
elif self.model_parameters.model == 'lstm':
cell = BasicLSTMCell(self.model_parameters.state_size,
state_is_tuple=True)
elif self.model_parameters.model == 'nas':
cell = NASCell(self.model_parameters.state_size)
else:
raise Exception("model type not supported: {}".format(
self.model_parameters.model))
if (self.model_parameters.output_keep_probability < 1.0
or self.model_parameters.input_keep_probability < 1.0):
if self.model_parameters.output_keep_probability < 1.0:
cell = DropoutWrapper(
cell,
output_keep_prob=self.output_keep_probability)
if self.model_parameters.input_keep_probability < 1.0:
cell = DropoutWrapper(
cell, input_keep_prob=self.input_keep_probability)
cells.append(cell)
cell = MultiRNNCell(cells)
# Simulate time steps and get RNN cell output
self.outputs, self.next_state = tf.nn.dynamic_rnn(cell,
self.inputs,
dtype=tf.float32)
# Define cost Tensors
with tf.name_scope("LSTMRNN_Cost"):
# Flatten to apply same weights to all time steps
self.flattened_outputs = tf.reshape(
self.outputs, [-1, self.model_parameters.state_size],
name="flattened_outputs")
self.output_w = tf.Variable(tf.truncated_normal(
[self.model_parameters.state_size, 1], stddev=0.01),
name="output_weights")
self.variable_summaries(self.output_w, 'output_weights')
self.output_b = tf.Variable(tf.constant(0.1), name="output_biases")
self.variable_summaries(self.output_w, 'output_biases')
# Define decision threshold Tensor
self.decision_threshold = tf.Variable(
self.model_parameters.threshold, name="decision_threshold")
# Define moving average step Tensor
self.ma_step = tf.Variable(self.model_parameters.ma_step,
name="ma_step")
# Softmax activation layer, using RNN inner loop last output
# logits and labels must have the same shape [batch_size, num_classes]
self.logits = tf.add(tf.matmul(self.flattened_outputs,
self.output_w),
self.output_b,
name="logits")
self.logits_bn = self.batch_norm_wrapper(
inputs=self.logits, is_training=self.is_training)
tf.summary.histogram('logits', self.logits)
tf.summary.histogram('logits_bn', self.logits_bn)
self.predictions = tf.reshape(
self.logits, [-1, self.model_parameters.sequence_length, 1],
name="predictions")
self.shaped_predictions = tf.reshape(self.predictions, [-1],
name="shaped_predictions")
self.tmp_smoothed_predictions = tf.concat(
[
self.shaped_predictions,
tf.fill(
tf.expand_dims(self.ma_step - 1, 0),
self.shaped_predictions[
tf.shape(self.shaped_predictions)[0] - 1])
],
axis=0,
name="tmp_smoothed_predictions")
self.ma_loop_idx = tf.constant(0, dtype='int32')
self.shaped_smoothed_predictions = tf.zeros([0], dtype='float32')
_, self.shaped_smoothed_predictions = tf.while_loop(
lambda i, _: i < tf.shape(self.shaped_predictions)[0],
self.ma_while_body,
[self.ma_loop_idx, self.shaped_smoothed_predictions],
shape_invariants=[tf.TensorShape([]),
tf.TensorShape([None])])
self.smoothed_predictions = tf.reshape(
self.shaped_smoothed_predictions,
[-1, self.model_parameters.sequence_length, 1],
name="smoothed_predictions")
self.soft_predictions_summary = tf.summary.tensor_summary(
"soft_predictions", self.smoothed_predictions)
# self.soft_predictions_summary = tf.summary.tensor_summary("soft_predictions", self.predictions)
# self.shaped_logits = tf.reshape(self.logits,
# [-1, self.model_parameters.sequence_length, 1],
# name="shaped_logits")
# Cross-Entropy
# self.cost = tf.reduce_mean(-tf.reduce_sum(
# self.targets * tf.log(self.predictions),
# reduction_indices=[2]), name="cross_entropy")
# self.cross_entropy = tf.reduce_mean(
# tf.nn.sigmoid_cross_entropy_with_logits(_sentinel=None,
# labels=self.targets,
# logits=self.predictions),
# name="cross_entropy")
# self.cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(
# _sentinel=None,
# labels=self.targets,
# logits=self.shaped_logits,
# name="cross_entropy")
# Root Mean Squared Error
# self.mean_squared_error = tf.losses.mean_squared_error(
# labels=self.targets,
# predictions=self.predictions)
self.cost = tf.sqrt(
tf.reduce_mean(
tf.squared_difference(self.smoothed_predictions,
self.targets)))
# self.cost = tf.sqrt(tf.reduce_mean(
# tf.squared_difference(
# self.predictions, self.targets)))
tf.summary.scalar('training_cost', self.cost)
# self.cost = tf.reduce_mean(
# self.cross_entropy,
# name="cost")
voicing_condition = tf.greater(
self.smoothed_predictions,
tf.fill(tf.shape(self.smoothed_predictions),
self.decision_threshold),
name="thresholding")
# voicing_condition = tf.greater(self.predictions,
# tf.fill(tf.shape(self.predictions), self.decision_threshold),
# name="thresholding")
self.label_predictions = tf.where(
voicing_condition,
tf.ones_like(self.smoothed_predictions),
tf.zeros_like(self.smoothed_predictions),
name="label_predictions")
# self.label_predictions = tf.where(voicing_condition,
# tf.ones_like(self.predictions) ,
# tf.zeros_like(self.predictions),
# name="label_predictions")
self.hard_predictions_summary = tf.summary.tensor_summary(
"hard_predictions", self.label_predictions)
self.correct_prediction = tf.equal(self.label_predictions,
self.targets,
name="correct_predictions")
self.r = tf.reshape(self.targets, [-1])
self.h = tf.reshape(self.label_predictions, [-1])
# Defined outside the while loop to avoid problems
self.dump_one = tf.constant(1, dtype=tf.int32, shape=[])
self.temp_pk_miss = tf.Variable([0], tf.int32, name='temp_pk_miss')
self.temp_pk_falsealarm = tf.Variable([0],
tf.int32,
name='temp_pk_falsealarm')
self.loop_idx = tf.constant(0, dtype=tf.int32, name='loop_idx')
self.loop_vars = self.loop_idx, self.temp_pk_miss, self.temp_pk_falsealarm
_, self.all_temp_pk_miss, self.all_temp_pk_falsealarm = tf.while_loop(
self.while_condition,
self.while_body,
self.loop_vars,
shape_invariants=(self.loop_idx.get_shape(),
tf.TensorShape([None]),
tf.TensorShape([None])))
self.pk_miss = tf.reduce_mean(
tf.cast(self.all_temp_pk_miss, tf.float32))
tf.summary.scalar('p_miss', self.pk_miss)
self.pk_falsealarm = tf.reduce_mean(
tf.cast(self.all_temp_pk_falsealarm, tf.float32))
tf.summary.scalar('p_falsealarm', self.pk_falsealarm)
self.pk = tf.reduce_mean(tf.cast(
tf.add(self.all_temp_pk_miss, self.all_temp_pk_falsealarm),
tf.float32),
name='pk')
tf.summary.scalar('pk', self.pk)
self.accuracy = tf.reduce_mean(tf.cast(self.correct_prediction,
tf.float32),
name="accuracy")
tf.summary.scalar('accuracy', self.accuracy)
self.recall, self.update_op_recall = tf.metrics.recall(
labels=self.targets,
predictions=self.label_predictions,
name="recall")
tf.summary.scalar('recall', self.recall)
self.precision, self.update_op_precision = tf.metrics.precision(
labels=self.targets,
predictions=self.label_predictions,
name="precision")
tf.summary.scalar('precision', self.precision)
# Define Training Tensors
with tf.name_scope("LSTMRNN_Train"):
# Momentum optimisation
self.optimizer = tf.train.MomentumOptimizer(
learning_rate=self.learning_rate,
momentum=self.momentum,
name="optimizer")
self.train_step = self.optimizer.minimize(self.cost,
name="train_step")
# Initializing the variables
self.initializer = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
name="b_h",
shape=[self._num_units],
initializer=tf.contrib.layers.variance_scaling_initializer())
r_t = tf.sign(tf.matmul(inputs, W_xr) + tf.matmul(state, W_hr) + b_r)
z_t = tf.sign(tf.matmul(inputs, W_xz) + tf.matmul(state, W_hz) + b_z)
h_hat_t = tf.tanh(
tf.matmul(inputs, W_xh) + tf.matmul(r_t * state, W_hh) + b_h)
h_t = z_t * state + (1 - z_t) * h_hat_t
return h_t, h_t
multi_cell = MultiRNNCell(
[BasicLSTMCell(state_dim) for i in range(num_layers)])
#multi_cell = MultiRNNCell([mygru(state_dim) for i in range(num_layers)])
initial_state = multi_cell.zero_state(batch_size, dtype=tf.float32)
# call seq2seq.rnn_decoder
outputs, final_state = rnn_decoder(inputs, initial_state, multi_cell)
# transform the list of state outputs to a list of logits.
# use a linear transformation.
weights = tf.get_variable(
name="W",
shape=[state_dim, vocab_size],
initializer=tf.contrib.layers.variance_scaling_initializer())
bias = tf.get_variable(
name="b",
shape=[vocab_size],
def __init__(self, num_units, forget_bias=1.0, input_size=None):
BasicLSTMCell.__init__(self, num_units, forget_bias=forget_bias, input_size=input_size)
self.matrix, self.bias = None, None
def _build_forward(self):
config = self.config
N, M, JX, JQ, VW, d, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.hidden_size, \
config.max_word_size
JX = tf.shape(self.x)[2]
JQ = tf.shape(self.q)[1]
M = tf.shape(self.x)[1]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
with tf.variable_scope("emb"):
print('word embedding')
# if config.use_char_emb:
# with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
# char_emb_mat = tf.get_variable("char_emb_mat", shape=[VC, dc], dtype='float')
# with tf.variable_scope("char"):
# Acx = tf.nn.embedding_lookup(char_emb_mat, self.cx) # [N, M, JX, W, dc]
# Acq = tf.nn.embedding_lookup(char_emb_mat, self.cq) # [N, JQ, W, dc]
# Acx = tf.reshape(Acx, [-1, JX, W, dc])
# Acq = tf.reshape(Acq, [-1, JQ, W, dc])
# filter_sizes = list(map(int, config.out_channel_dims.split(',')))
# heights = list(map(int, config.filter_heights.split(',')))
# assert sum(filter_sizes) == dco, (filter_sizes, dco)
# with tf.variable_scope("conv"):
# xx = multi_conv1d(Acx, filter_sizes, heights, "VALID", self.is_train, config.keep_prob, scope="xx")
# if config.share_cnn_weights:
# tf.get_variable_scope().reuse_variables()
# qq = multi_conv1d(Acq, filter_sizes, heights, "VALID", self.is_train, config.keep_prob, scope="xx")
# else:
# qq = multi_conv1d(Acq, filter_sizes, heights, "VALID", self.is_train, config.keep_prob, scope="qq")
# xx = tf.reshape(xx, [-1, M, JX, dco])
# qq = tf.reshape(qq, [-1, JQ, dco])
if config.use_word_emb:
with tf.variable_scope("emb_var"), tf.device("/gpu:7"):
if config.mode == 'train':
word_emb_mat = tf.get_variable("word_emb_mat", dtype='float', shape=[VW, dw], initializer=get_initializer(config.emb_mat))
else:
word_emb_mat = tf.get_variable("word_emb_mat", shape=[VW, dw], dtype='float')
# if config.use_glove_for_unk:
# word_emb_mat = tf.concat(0, [word_emb_mat])
print(word_emb_mat)
with tf.name_scope("word"):
print('embedding lookup')
Ax = tf.nn.embedding_lookup(word_emb_mat, self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb_mat, self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
print('embedding lookup ready')
# if config.use_char_emb:
# xx = tf.concat(3, [xx, Ax]) # [N, M, JX, di]
# qq = tf.concat(2, [qq, Aq]) # [N, JQ, di]
# else:
xx = Ax
qq = Aq
# highway network
#if config.highway:
# with tf.variable_scope("highway"):
# xx = highway_network(xx, config.highway_num_layers, True, wd=config.wd, is_train=self.is_train)
# tf.get_variable_scope().reuse_variables()
# qq = highway_network(qq, config.highway_num_layers, True, wd=config.wd, is_train=self.is_train)
self.tensor_dict['xx'] = xx
self.tensor_dict['qq'] = qq
print('context emmbedding')
cell = BasicLSTMCell(d, state_is_tuple=True)
d_cell = SwitchableDropoutWrapper(cell, self.is_train, input_keep_prob=config.input_keep_prob)
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N]
print('prepro')
with tf.variable_scope("prepro"):
(fw_u, bw_u), ((_, fw_u_f), (_, bw_u_f)) = bidirectional_dynamic_rnn(d_cell, d_cell, qq, q_len, dtype='float32', scope='u1') # [N, J, d], [N, d]
u = tf.concat(2, [fw_u, bw_u])
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(cell, cell, xx, x_len, dtype='float', scope='u1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(cell, cell, xx, x_len, dtype='float', scope='h1') # [N, M, JX, 2d]
h = tf.concat(3, [fw_h, bw_h]) # [N, M, JX, 2d]
self.tensor_dict['u'] = u
self.tensor_dict['h'] = h
print('main pro')
with tf.variable_scope("main"):
if config.dynamic_att:
p0 = h
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]), [N * M, JQ, 2 * d])
q_mask = tf.reshape(tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]), [N * M, JQ])
first_cell = AttentionCell(cell, u, mask=q_mask, mapper='sim',
input_keep_prob=self.config.input_keep_prob, is_train=self.is_train)
else:
p0 = attention_layer(config, self.is_train, h, u, h_mask=self.x_mask, u_mask=self.q_mask, scope="p0", tensor_dict=self.tensor_dict)
first_cell = d_cell
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(first_cell, first_cell, p0, x_len, dtype='float', scope='g0') # [N, M, JX, 2d]
g0 = tf.concat(3, [fw_g0, bw_g0])
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(first_cell, first_cell, g0, x_len, dtype='float', scope='g1') # [N, M, JX, 2d]
g1 = tf.concat(3, [fw_g1, bw_g1])
logits = get_logits([g1, p0], [config.batch_size,config.max_num_sents] , True, wd=config.wd, input_keep_prob=config.input_keep_prob,
mask=self.x_mask, is_train=self.is_train, func='sigmoid', scope='logits1')
# a1i = softsel(tf.reshape(g1, [N, M * JX, 2 * d]), tf.reshape(logits, [N, M * JX]))
# a1i = tf.tile(tf.expand_dims(tf.expand_dims(a1i, 1), 1), [1, M, JX, 1])
# (fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(d_cell, d_cell, tf.concat(3, [p0, g1, a1i, g1 * a1i]),
# x_len, dtype='float', scope='g2') # [N, M, JX, 2d]
# g2 = tf.concat(3, [fw_g2, bw_g2])
# logits2 = get_logits([g2, p0], d, True, wd=config.wd, input_keep_prob=config.input_keep_prob,
# mask=self.x_mask,
# is_train=self.is_train, func=config.answer_func, scope='logits2')
# flat_logits = tf.reshape(logits, [-1, M * JX])
# flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
# yp = tf.reshape(flat_yp, [-1, M, JX])
yp = tf.greater(0.5, logits)
# flat_logits2 = tf.reshape(logits2, [-1, M * JX])
# flat_yp2 = tf.nn.softmax(flat_logits2)
# yp2 = tf.reshape(flat_yp2, [-1, M, JX])
self.tensor_dict['g1'] = g1
# self.tensor_dict['g2'] = g2
self.logits = logits
# self.logits2 = flat_logits2
self.yp = yp
def build_forward(self):
config = self.config
N, M, JX, JQ, VW , VC, d, W = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.char_vocab_size, config.hidden_size, \
config.max_word_size
JX = tf.shape(self.x)[2]
JQ = tf.shape(self.q)[1]
M = tf.shape(self.x)[1]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
word_emb = tf.get_variable(
"word_emb_mat",
shape=[config.word_vocab_size, config.word_emb_size],
dtype='float',
initializer=self.emb_mat)
with tf.variable_scope("embedding"):
with tf.variable_scope("emb_var"), tf.device("/cpu:0"):
char_emb_mat = tf.get_variable("char_emb_mat",
shape=[VC, dc],
dtype='float')
with tf.variable_scope("char"), tf.device("/cpu:0"):
Acx = tf.nn.embedding_lookup(char_emb_mat,
self.cx) # [N, M, JX, W, dc]
Acq = tf.nn.embedding_lookup(char_emb_mat,
self.cq) # [N, JQ, W, dc]
Acx = tf.reshape(Acx, [-1, JX, W, dc])
Acq = tf.reshape(Acq, [-1, JQ, W, dc])
filter_sizes = list(
map(int, config.out_channel_dims.split(',')))
heights = list(map(int, config.filter_heights.split(',')))
assert sum(filter_sizes) == dco, (filter_sizes, dco)
with tf.variable_scope("conv"):
xx = multi_conv1d(Acx,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
if config.share_cnn_weights:
tf.get_variable_scope().reuse_variables()
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="xx")
else:
qq = multi_conv1d(Acq,
filter_sizes,
heights,
"VALID",
self.is_train,
config.keep_prob,
scope="qq")
xx = tf.reshape(xx, [-1, M, JX, dco])
qq = tf.reshape(qq, [-1, JQ, dco])
with tf.name_scope("word"):
Ax = tf.nn.embedding_lookup(word_emb, self.x) # [N, M, JX, d]
Aq = tf.nn.embedding_lookup(word_emb, self.q) # [N, JQ, d]
self.tensor_dict['x'] = Ax
self.tensor_dict['q'] = Aq
if config.use_char_emb:
xx = tf.concat([xx, Ax], 3) # [N, M, JX, di]
qq = tf.concat([qq, Aq], 2) # [N, JQ, di]
else:
xx = Ax
qq = Aq
self.tensor_dict['xx'] = xx
self.tensor_dict['qq'] = qq
'''x_len means the length of sequences '''
x_len = tf.reduce_sum(tf.cast(self.x_mask, 'int32'), 2) # [N, M]
q_len = tf.reduce_sum(tf.cast(self.q_mask, 'int32'), 1) # [N]
with tf.variable_scope("Encoding"):
cell = BasicLSTMCell(d, state_is_tuple=True)
encoding_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
(fw_u, bw_u), ((_, fw_u_f), (_,
bw_u_f)) = bidirectional_dynamic_rnn(
encoding_cell,
encoding_cell,
qq,
q_len,
dtype='float',
scope='u1') # [N, J, d], [N, d]
u = tf.concat([fw_u, bw_u], 2)
if config.share_lstm_weights:
tf.get_variable_scope().reuse_variables()
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='u1') # [N, M, JX, 2d]
h = tf.concat([fw_h, bw_h], 3) # [N, M, JX, 2d]
else:
(fw_h, bw_h), _ = bidirectional_dynamic_rnn(
cell, cell, xx, x_len, dtype='float',
scope='h1') # [N, M, JX, 2d]
h = tf.concat([fw_h, bw_h], 3) # [N, M, JX, 2d]
self.tensor_dict['u'] = u
self.tensor_dict['h'] = h
with tf.variable_scope("main"):
if config.dynamic_att:
p0 = h
u = tf.reshape(tf.tile(tf.expand_dims(u, 1), [1, M, 1, 1]),
[N * M, JQ, 2 * d])
q_mask = tf.reshape(
tf.tile(tf.expand_dims(self.q_mask, 1), [1, M, 1]),
[N * M, JQ])
first_cell = AttentionCell(
cell,
u,
mask=q_mask,
mapper='sim',
input_keep_prob=self.config.input_keep_prob,
is_train=self.is_train)
else:
## G
p0 = attention_layer(config,
self.is_train,
h,
u,
h_mask=self.x_mask,
u_mask=self.q_mask,
scope="p0",
tensor_dict=self.tensor_dict)
cell = BasicLSTMCell(d, state_is_tuple=True)
first_cell = SwitchableDropoutWrapper(
cell,
self.is_train,
input_keep_prob=config.input_keep_prob)
'''the following can be simplified, by using multi-layer rnn'''
## 2 layers of bi rnn
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, p0, x_len, dtype='float',
scope='g0') # [N, M, JX, 2d]
g0 = tf.concat([fw_g0, bw_g0], 3)
cell = BasicLSTMCell(d, state_is_tuple=True)
first_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(
first_cell, first_cell, g0, x_len, dtype='float',
scope='g1') # [N, M, JX, 2d]
##M
g1 = tf.concat([fw_g1, bw_g1], 3)
logits = get_logits([g1, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits1')
a1i = softsel(tf.reshape(g1, [N, M * JX, 2 * d]),
tf.reshape(logits, [N, M * JX]))
a1i = tf.tile(tf.expand_dims(tf.expand_dims(a1i, 1), 1),
[1, M, JX, 1])
cell = BasicLSTMCell(d, state_is_tuple=True)
M2_operate_cell = SwitchableDropoutWrapper(
cell, self.is_train, input_keep_prob=config.input_keep_prob)
(fw_g2, bw_g2), _ = bidirectional_dynamic_rnn(
M2_operate_cell,
M2_operate_cell,
tf.concat([p0, g1, a1i, g1 * a1i], 3),
x_len,
dtype='float',
scope='g2') # [N, M, JX, 2d]
## M^2
g2 = tf.concat([fw_g2, bw_g2], 3)
logits2 = get_logits([g2, p0],
d,
True,
wd=config.wd,
input_keep_prob=config.input_keep_prob,
mask=self.x_mask,
is_train=self.is_train,
func=config.answer_func,
scope='logits2')
flat_logits = tf.reshape(logits, [-1, M * JX])
flat_yp = tf.nn.softmax(flat_logits) # [-1, M*JX]
yp = tf.reshape(flat_yp, [-1, M, JX])
flat_logits2 = tf.reshape(logits2, [-1, M * JX])
flat_yp2 = tf.nn.softmax(flat_logits2)
yp2 = tf.reshape(flat_yp2, [-1, M, JX])
self.tensor_dict['g1'] = g1
self.tensor_dict['g2'] = g2
self.logits = flat_logits
self.logits2 = flat_logits2
self.yp = yp
self.yp2 = yp2
import tensorflow as tf from tensorflow.python.ops.rnn_cell import BasicLSTMCell from tensorflow.python.ops.rnn_cell import MultiRNNCell num_units = [128, 64] cells = [BasicLSTMCell(num_units=n) for n in num_units] stacked_rnn_cell = MultiRNNCell(cells)
