def symbols_to_logits_fn(ids, dec_state):
dec = []
dec_c, dec_h = [], []
# (batch x beam_size x decoded_seq)
ids = tf.reshape(ids, [Hp.batch_size, beam_size, -1])
print("dec_state ", dec_state[0].get_shape().as_list())
for ind in range(beam_size):
with tf.variable_scope('dec_lstm', reuse=ind > 0
or reuse_vars):
w_input = ids[:, ind, -1].sg_lookup(emb=emb_word)
dec_state0 = tf.contrib.rnn.LSTMStateTuple(
c=dec_state.c[:, ind, :], h=dec_state.h[:, ind, :])
dec_out, dec_state_i = dec_cell(w_input, dec_state0)
dec_out = tf.expand_dims(dec_out, 1)
dec_i = dec_out.sg_conv1d_gpus(size=1,
dim=Hp.word_vs,
name="out_conv",
act="linear",
dev=dev,
reuse=ind > 0 or reuse_vars)
dec.append(tf.squeeze(dec_i, 1))
dec_c.append(dec_state_i[0])
dec_h.append(dec_state_i[1])
return tf.stack(dec, 1), tf.contrib.rnn.LSTMStateTuple(
tf.stack(dec_c, 1), tf.stack(dec_h, 1))
def read_and_decode(filename_queue, batch_size):
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
features = tf.parse_single_example(
serialized_example,
# Defaults are not specified since both keys are required.
features={
'height':
tf.FixedLenFeature([], tf.int64),
'word_opinion':
tf.FixedLenFeature([Hp.w_maxlen * 6],
dtype=tf.int64,
default_value=[-1] * Hp.w_maxlen * 6),
'char_opinion':
tf.FixedLenFeature([], tf.string)
})
char_opinion = tf.decode_raw(features['char_opinion'], tf.uint8)
height = tf.cast(features['height'], tf.int32)
word_opinion = tf.cast(features['word_opinion'], tf.int32)
char_opinion = tf.reshape(char_opinion, tf.stack([6, Hp.c_maxlen]))
word_opinion = tf.reshape(word_opinion, tf.stack([6, Hp.w_maxlen]))
words, chars = tf.train.batch([word_opinion, char_opinion],
batch_size=batch_size,
capacity=3 * batch_size,
num_threads=1)
return (words, chars)
def sg_upconv(tensor, opt):
r"""Applies a up convolution (or convolution transpose).
Args:
tensor: A 4-D `Tensor` (automatically passed by decorator).
opt:
size: A tuple/list of integers of length 2 representing `[kernel height, kernel width]`.
Can be an integer if both values are the same.
If not specified, (4, 4) is set implicitly.
Default value is [1, 2, 2, 1].
stride: A tuple/list of integers of length 2 or 4 representing stride dimensions.
If the length is 2, i.e., (a, b), the stride is `[1, a, b, 1]`.
If the length is 4, i.e., (a, b, c, d), the stride is `[a, b, c, d]`.
Can be an integer. If the length is a, the stride is `[1, a, a, 1]`.
in_dim: A positive `integer`. The size of input dimension.
dim: A positive `integer`. The size of output dimension.
pad: Either `SAME` (Default) or `VALID`.
bias: Boolean. If True, biases are added.
Returns:
A `Tensor` with the same type as `tensor`.
"""
# default options
opt += tf.sg_opt(size=(4, 4), stride=(1, 2, 2, 1), pad='SAME')
opt.size = opt.size if isinstance(opt.size,
(tuple, list)) else [opt.size, opt.size]
opt.stride = opt.stride if isinstance(
opt.stride, (tuple, list)) else [1, opt.stride, opt.stride, 1]
opt.stride = [1, opt.stride[0], opt.stride[1], 1] if len(
opt.stride) == 2 else opt.stride
# parameter tf.sg_initializer
w = tf.sg_initializer.he_uniform(
'W', (opt.size[0], opt.size[1], opt.dim, opt.in_dim))
b = tf.sg_initializer.constant('b', opt.dim) if opt.bias else 0
# tedious shape handling for conv2d_transpose
shape = tensor.get_shape().as_list()
out_shape = [
tf.shape(tensor)[0], shape[1] * opt.stride[1],
shape[2] * opt.stride[2], opt.dim
]
# apply convolution
out = tf.nn.conv2d_transpose(tensor,
w,
output_shape=tf.stack(out_shape),
strides=opt.stride,
padding=opt.pad) + b
# reset shape is needed because conv2d_transpose() erase all shape information.
# noinspection PyUnresolvedReferences
out.set_shape([None, out_shape[1], out_shape[2], opt.dim])
return out
def iou(self, boxes1, boxes2):
"""
Note: Modified from https://github.com/nilboy/tensorflow-yolo/blob/python2.7/yolo/net/yolo_net.py
calculate ious
Args:
boxes1: 4-D tensor [CELL_SIZE, CELL_SIZE, BOXES_PER_CELL, 4] ====> (x_center, y_center, w, h)
boxes2: 1-D tensor [4] ===> (x_center, y_center, w, h)
Return:
iou: 3-D tensor [CELL_SIZE, CELL_SIZE, BOXES_PER_CELL]
"""
boxes1 = tf.stack([
boxes1[:, 0] - boxes1[:, 2] / 2, boxes1[:, 1] - boxes1[:, 3] / 2,
boxes1[:, 0] + boxes1[:, 2] / 2, boxes1[:, 1] + boxes1[:, 3] / 2
])
boxes2 = tf.stack([
boxes2[:, 0] - boxes2[:, 2] / 2, boxes2[:, 1] - boxes2[:, 3] / 2,
boxes2[:, 0] + boxes2[:, 2] / 2, boxes2[:, 1] + boxes2[:, 3] / 2
])
#calculate the left up point
lu = tf.maximum(boxes1[0:2], boxes2[0:2])
rd = tf.minimum(boxes1[2:], boxes2[2:])
#intersection
intersection = rd - lu
inter_square = tf.multiply(intersection[0], intersection[1])
mask = tf.cast(intersection[0] > 0, tf.float32) * tf.cast(
intersection[1] > 0, tf.float32)
inter_square = tf.multiply(mask, inter_square)
#calculate the boxs1 square and boxs2 square
square1 = tf.multiply((boxes1[2] - boxes1[0]), (boxes1[3] - boxes1[1]))
square2 = tf.multiply((boxes2[2] - boxes2[0]), (boxes2[3] - boxes2[1]))
return inter_square / (square1 + square2 - inter_square +
1e-6), inter_square
def sg_upconv1d(tensor, opt):
r"""Applies 1-D a up convolution (or convolution transpose).
Args:
tensor: A 3-D `Tensor` (automatically passed by decorator).
opt:
size: A positive `integer` representing `[kernel width]`. As a default it is set to 4
stride: A positive `integer` representing stride dimension. As a default it is set to 2
in_dim: A positive `integer`. The size of input dimension.
dim: A positive `integer`. The size of output dimension.
pad: Either `SAME` (Default) or `VALID`.
bias: Boolean. If True, biases are added.
regularizer: A (Tensor -> Tensor or None) function; the result of applying it on a newly created variable
will be added to the collection tf.GraphKeys.REGULARIZATION_LOSSES and can be used for regularization
summary: If True, summaries are added. The default is True.
Returns:
A `Tensor` with the same type as `tensor`.
"""
# default options
opt += tf.sg_opt(size=4, stride=2, pad='SAME')
opt.size = [opt.size, 1]
opt.stride = [1, opt.stride, 1, 1]
# parameter tf.sg_initializer
w = tf.sg_initializer.he_uniform('W', (opt.size[0], opt.size[1], opt.dim, opt.in_dim),
regularizer=opt.regularizer, summary=opt.summary)
b = tf.sg_initializer.constant('b', opt.dim, summary=opt.summary) if opt.bias else 0
# make 4-D tensor
tensor = tensor.sg_expand_dims(axis=2)
# tedious shape handling for conv2d_transpose
shape = tensor.get_shape().as_list()
if opt.pad == "SAME":
out_shape = [tf.shape(tensor)[0], shape[1] * opt.stride[1], shape[2] * opt.stride[2], opt.dim]
else:
out_shape = [tf.shape(tensor)[0], (shape[1] - 1) * opt.stride[1] + opt.size[0],
(shape[2] - 1) * opt.stride[2] + opt.size[1], opt.dim]
# apply convolution
out = tf.nn.conv2d_transpose(tensor, w, output_shape=tf.stack(out_shape),
strides=opt.stride, padding=opt.pad) + b
# reset shape is needed because conv2d_transpose() erase all shape information.
# noinspection PyUnresolvedReferences
out.set_shape([None, out_shape[1], out_shape[2], opt.dim])
# squeeze
out = out.sq_squeeze(axis=2)
return out
def sg_upconv1d(tensor, opt):
r"""Applies 1-D a up convolution (or convolution transpose).
Args:
tensor: A 3-D `Tensor` (automatically passed by decorator).
opt:
size: A positive `integer` representing `[kernel width]`. As a default it is set to 4
stride: A positive `integer` representing stride dimension. As a default it is set to 2
in_dim: A positive `integer`. The size of input dimension.
dim: A positive `integer`. The size of output dimension.
pad: Either `SAME` (Default) or `VALID`.
bias: Boolean. If True, biases are added.
Returns:
A `Tensor` with the same type as `tensor`.
"""
# default options
opt += tf.sg_opt(size=4, stride=2, pad='SAME')
opt.size = [opt.size, 1]
opt.stride = [1, opt.stride, 1, 1]
# parameter tf.sg_initializer
w = tf.sg_initializer.he_uniform(
'W', (opt.size[0], opt.size[1], opt.dim, opt.in_dim))
b = tf.sg_initializer.constant('b', opt.dim) if opt.bias else 0
# make 4-D tensor
tensor = tensor.sg_expand_dims(dim=2)
# tedious shape handling for conv2d_transpose
shape = tensor.get_shape().as_list()
out_shape = [
tf.shape(tensor)[0], shape[1] * opt.stride[1],
shape[2] * opt.stride[2], opt.dim
]
# apply convolution
out = tf.nn.conv2d_transpose(tensor,
w,
output_shape=tf.stack(out_shape),
strides=opt.stride,
padding=opt.pad) + b
# reset shape is needed because conv2d_transpose() erase all shape information.
# noinspection PyUnresolvedReferences
out.set_shape([None, out_shape[1], out_shape[2], opt.dim])
# squeeze
out = out.sq_squeeze(dim=2)
return out
def q_process(t1, t2):
'''
Processes each training sample so that it fits in the queue.
'''
# Lstrip zeros
zeros = tf.equal(t1, tf.zeros_like(t1)).sg_int().sg_sum()
t1 = t1[zeros:]
t2 = t2[zeros:]
# zero-PrePadding
t1 = tf.concat([tf.zeros([Hyperparams.seqlen-1], tf.int32), t1], 0)# 49 zero-prepadding
t2 = tf.concat([tf.zeros([Hyperparams.seqlen-1], tf.int32), t2], 0)# 49 zero-prepadding
# radom crop
stacked = tf.stack((t1, t2))
cropped = tf.random_crop(stacked, [2, Hyperparams.seqlen])
t1, t2 = cropped[0], cropped[1]
t2 = t2[-1]
return t1, t2
def tower_infer_dec(chars,
scope,
rnn_cell,
dec_cell,
word_emb,
rnn_state,
out_reuse_vars=False,
dev='/cpu:0'):
with tf.device(dev):
with tf.variable_scope('embatch_size', reuse=True):
# (vocab_size, latent_dim)
emb_char = tf.sg_emb(name='emb_char',
voca_size=Hp.char_vs,
dim=Hp.hd,
dev=dev)
emb_word = tf.sg_emb(name='emb_word',
emb=word_emb,
voca_size=Hp.word_vs,
dim=300,
dev=dev)
print(chars)
ch = chars
ch = tf.reverse_sequence(input=ch,
seq_lengths=[Hp.c_maxlen] * Hp.batch_size,
seq_dim=1)
reuse_vars = reuse_vars_enc = True
# -------------------------- BYTENET ENCODER --------------------------
with tf.variable_scope('encoder'):
# embed table lookup
enc = ch.sg_lookup(emb=emb_char) #(batch, sentlen, latentdim)
# loop dilated conv block
for i in range(Hp.num_blocks):
enc = (enc.sg_res_block(size=5,
rate=1,
name="enc1_%d" % (i),
is_first=True,
reuse_vars=reuse_vars,
dev=dev).sg_res_block(
size=5,
rate=2,
name="enc2_%d" % (i),
reuse_vars=reuse_vars,
dev=dev).sg_res_block(
size=5,
rate=4,
name="enc4_%d" % (i),
reuse_vars=reuse_vars,
dev=dev).sg_res_block(
size=5,
rate=8,
name="enc8_%d" % (i),
reuse_vars=reuse_vars,
dev=dev).sg_res_block(
size=5,
rate=16,
name="enc16_%d" % (i),
reuse_vars=reuse_vars,
dev=dev))
byte_enc = enc
# -------------------------- QCNN + QPOOL ENCODER #1 --------------------------
with tf.variable_scope('quazi'):
#quasi cnn layer ZFO [batch * 3, seqlen, dim2 ]
conv = byte_enc.sg_quasi_conv1d(is_enc=True,
size=4,
name="qconv_1",
dev=dev,
reuse_vars=reuse_vars)
# c = f * c + (1 - f) * z, h = o*c [batch * 4, seqlen, hd]
pool0 = conv.sg_quasi_rnn(is_enc=False,
att=False,
name="qrnn_1",
reuse_vars=reuse_vars,
dev=dev)
qpool_last = pool0[:, -1, :]
# -------------------------- MAXPOOL along time dimension --------------------------
inpt_maxpl = tf.expand_dims(byte_enc, 1) # [batch, 1, seqlen, channels]
maxpool = tf.nn.max_pool(inpt_maxpl, [1, 1, Hp.c_maxlen, 1], [1, 1, 1, 1],
'VALID')
maxpool = tf.squeeze(maxpool, [1, 2])
# -------------------------- HIGHWAY --------------------------
concat = qpool_last + maxpool
with tf.variable_scope('highway', reuse=reuse_vars):
input_lstm = highway(concat, concat.get_shape()[-1], num_layers=1)
# -------------------------- CONTEXT LSTM --------------------------
input_lstm = tf.nn.dropout(input_lstm, Hp.keep_prob)
with tf.variable_scope('contx_lstm', reuse=reuse_vars):
output, rnn_state = rnn_cell(input_lstm, rnn_state)
beam_size = 8
reuse_vars = out_reuse_vars
greedy = False
if greedy:
dec_state = rnn_state
dec_out = []
d_out = tf.constant([1] * Hp.batch_size)
for idx in range(Hp.w_maxlen):
w_input = d_out.sg_lookup(emb=emb_word)
dec_state = tf.contrib.rnn.LSTMStateTuple(c=dec_state.c,
h=dec_state.h)
with tf.variable_scope('dec_lstm', reuse=idx > 0 or reuse_vars):
d_out, dec_state = dec_cell(w_input, dec_state)
dec_out.append(d_out)
d_out = tf.expand_dims(d_out, 1).sg_conv1d_gpus(size=1,
dim=Hp.word_vs,
name="out_conv",
act="linear",
dev=dev,
reuse=idx > 0
or reuse_vars)
d_out = tf.squeeze(d_out).sg_argmax()
dec_out = tf.stack(dec_out, 1)
dec = dec_out.sg_conv1d_gpus(size=1,
dim=Hp.word_vs,
name="out_conv",
act="linear",
dev=dev,
reuse=True)
return dec.sg_argmax(), rnn_state
else:
# ------------------ BEAM SEARCH --------------------
dec_state = tf.contrib.rnn.LSTMStateTuple(
tf.tile(tf.expand_dims(rnn_state[0], 1), [1, beam_size, 1]),
tf.tile(tf.expand_dims(rnn_state[1], 1), [1, beam_size, 1]))
initial_ids = tf.constant([1] * Hp.batch_size)
def symbols_to_logits_fn(ids, dec_state):
dec = []
dec_c, dec_h = [], []
# (batch x beam_size x decoded_seq)
ids = tf.reshape(ids, [Hp.batch_size, beam_size, -1])
print("dec_state ", dec_state[0].get_shape().as_list())
for ind in range(beam_size):
with tf.variable_scope('dec_lstm', reuse=ind > 0
or reuse_vars):
w_input = ids[:, ind, -1].sg_lookup(emb=emb_word)
dec_state0 = tf.contrib.rnn.LSTMStateTuple(
c=dec_state.c[:, ind, :], h=dec_state.h[:, ind, :])
dec_out, dec_state_i = dec_cell(w_input, dec_state0)
dec_out = tf.expand_dims(dec_out, 1)
dec_i = dec_out.sg_conv1d_gpus(size=1,
dim=Hp.word_vs,
name="out_conv",
act="linear",
dev=dev,
reuse=ind > 0 or reuse_vars)
dec.append(tf.squeeze(dec_i, 1))
dec_c.append(dec_state_i[0])
dec_h.append(dec_state_i[1])
return tf.stack(dec, 1), tf.contrib.rnn.LSTMStateTuple(
tf.stack(dec_c, 1), tf.stack(dec_h, 1))
final_ids, final_probs = beam_search.beam_search(symbols_to_logits_fn,
dec_state,
initial_ids,
beam_size,
Hp.w_maxlen - 1,
Hp.word_vs,
3.5,
eos_id=2)
return final_ids[:, 0, :], rnn_state