import random
import itertools
import numpy as np
import tensorflow as tf
from tensorflow.python.ops.rnn_cell import BasicLSTMCell, GRUCell
from basic_cnn.read_data import DataSet
from basic_cnn.superhighway import SHCell
from my.tensorflow import exp_mask, get_initializer, VERY_SMALL_NUMBER
from my.tensorflow.nn import linear, double_linear_logits, linear_logits, softsel, dropout, get_logits, softmax, \
highway_network, multi_conv1d
from my.tensorflow.rnn import bidirectional_dynamic_rnn, dynamic_rnn
from my.tensorflow.rnn_cell import SwitchableDropoutWrapper, AttentionCell
def bi_attention(config, is_train, h, u, h_mask=None, u_mask=None, scope=None, tensor_dict=None):
"""
h_a:
all u attending on h
choosing an element of h that max-matches u
First creates confusion matrix between h and u
Then take max of the attention weights over u row
Finally softmax over
u_a:
each h attending on u
:param h: [N, M, JX, d]
:param u: [N, JQ, d]
:param h_mask: [N, M, JX]
:param u_mask: [N, B]
:param scope:
:return: [N, M, d], [N, M, JX, d]
"""
with tf.variable_scope(scope or "bi_attention"):
N, M, JX, JQ, d = config.batch_size, config.max_num_sents, config.max_sent_size, config.max_ques_size, config.hidden_size
JX = tf.shape(h)[2]
h_aug = tf.tile(tf.expand_dims(h, 3), [1, 1, 1, JQ, 1])
u_aug = tf.tile(tf.expand_dims(tf.expand_dims(u, 1), 1), [1, M, JX, 1, 1])
if h_mask is None:
and_mask = None
else:
h_mask_aug = tf.tile(tf.expand_dims(h_mask, 3), [1, 1, 1, JQ])
u_mask_aug = tf.tile(tf.expand_dims(tf.expand_dims(u_mask, 1), 1), [1, M, JX, 1])
and_mask = h_mask_aug & u_mask_aug
u_logits = get_logits([h_aug, u_aug], None, True, wd=config.wd, mask=and_mask,
is_train=is_train, func=config.logit_func, scope='u_logits') # [N, M, JX, JQ]
u_a = softsel(u_aug, u_logits) # [N, M, JX, d]
if tensor_dict is not None:
# a_h = tf.nn.softmax(h_logits) # [N, M, JX]
a_u = tf.nn.softmax(u_logits) # [N, M, JX, JQ]
# tensor_dict['a_h'] = a_h
tensor_dict['a_u'] = a_u
if config.bi:
h_a = softsel(h, tf.reduce_max(u_logits, 3)) # [N, M, d]
h_a = tf.tile(tf.expand_dims(h_a, 2), [1, 1, JX, 1])
else:
h_a = None
return u_a, h_a
def attention_layer(config, is_train, h, u, h_mask=None, u_mask=None, scope=None, tensor_dict=None):
with tf.variable_scope(scope or "attention_layer"):
u_a, h_a = bi_attention(config, is_train, h, u, h_mask=h_mask, u_mask=u_mask, tensor_dict=tensor_dict)
if config.bi:
p0 = tf.concat(3, [h , u_a, h * u_a, h * h_a])
else:
p0 = tf.concat(3, [h , u_a, h * u_a])
return p0
class Model(object):
def __init__(self, config, scope):
self.scope = scope
self.config = config
self.global_step = tf.get_variable('global_step', shape=[], dtype='int32',
initializer=tf.constant_initializer(0), trainable=False)
# Define forward inputs here
N, M, JX, JQ, VW, VC, 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.max_word_size
self.x = tf.placeholder('int32', [N, M, None], name='x')
self.cx = tf.placeholder('int32', [N, M, None, W], name='cx')
self.x_mask = tf.placeholder('bool', [N, M, None], name='x_mask')
self.q = tf.placeholder('int32', [N, JQ], name='q')
self.cq = tf.placeholder('int32', [N, JQ, W], name='cq')
self.q_mask = tf.placeholder('bool', [N, JQ], name='q_mask')
self.y = tf.placeholder('bool', [N, M, JX], name='y')
self.is_train = tf.placeholder('bool', [], name='is_train')
self.new_emb_mat = tf.placeholder('float', [None, config.word_emb_size], name='new_emb_mat')
# Define misc
self.tensor_dict = {}
# Forward outputs / loss inputs
self.logits = None
self.yp = None
self.var_list = None
# Loss outputs
self.loss = None
self._build_forward()
self._build_loss()
if config.mode == 'train':
self._build_ema()
self.summary = tf.merge_all_summaries()
self.summary = tf.merge_summary(tf.get_collection("summaries", scope=self.scope))
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]
dc, dw, dco = config.char_emb_size, config.word_emb_size, config.char_out_size
with tf.variable_scope("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
xx = tf.concat(3, [xx, Ax]) # [N, M, JX, di]
qq = tf.concat(2, [qq, Aq]) # [N, JQ, di]
# highway network
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"):
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)
(fw_g0, bw_g0), _ = bidirectional_dynamic_rnn(d_cell, d_cell, p0, x_len, dtype='float', scope='g0') # [N, M, JX, 2d]
g0 = tf.concat(3, [fw_g0, bw_g0])
# p1 = attention_layer(config, self.is_train, g0, u, h_mask=self.x_mask, u_mask=self.q_mask, scope="p1")
(fw_g1, bw_g1), _ = bidirectional_dynamic_rnn(d_cell, d_cell, g0, x_len, dtype='float', scope='g1') # [N, M, JX, 2d]
g1 = tf.concat(3, [fw_g1, bw_g1])
# logits = u_logits(config, self.is_train, g1, u, h_mask=self.x_mask, u_mask=self.q_mask, scope="logits")
# [N, M, JX]
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])
self.tensor_dict['g1'] = g1
self.logits = flat_logits
self.yp = yp
def _build_loss(self):
config = self.config
N, M, JX, JQ, VW, VC = \
config.batch_size, config.max_num_sents, config.max_sent_size, \
config.max_ques_size, config.word_vocab_size, config.char_vocab_size
JX = tf.shape(self.x)[2]
loss_mask = tf.reduce_max(tf.cast(self.q_mask, 'float'), 1)
losses = -tf.log(tf.reduce_sum(self.yp * tf.cast(self.y, 'float'), [1, 2]) + VERY_SMALL_NUMBER)
ce_loss = tf.reduce_mean(loss_mask * losses)
tf.add_to_collection('losses', ce_loss)
self.loss = tf.add_n(tf.get_collection('losses', scope=self.scope), name='loss')
tf.scalar_summary(self.loss.op.name, self.loss)
tf.add_to_collection('ema/scalar', self.loss)
def _build_ema(self):
ema = tf.train.ExponentialMovingAverage(self.config.decay)
ema_op = ema.apply(tf.get_collection("ema/scalar", scope=self.scope) + tf.get_collection("ema/histogram", scope=self.scope))
for var in tf.get_collection("ema/scalar", scope=self.scope):
ema_var = ema.average(var)
tf.scalar_summary(ema_var.op.name, ema_var)
for var in tf.get_collection("ema/histogram", scope=self.scope):
ema_var = ema.average(var)
tf.histogram_summary(ema_var.op.name, ema_var)
with tf.control_dependencies([ema_op]):
self.loss = tf.identity(self.loss)
def get_loss(self):
return self.loss
def get_global_step(self):
return self.global_step
def get_var_list(self):
return self.var_list
def get_feed_dict(self, batch, is_train, supervised=True):
assert isinstance(batch, DataSet)
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
feed_dict = {}
if config.len_opt:
"""
Note that this optimization results in variable GPU RAM usage (i.e. can cause OOM in the middle of training.)
First test without len_opt and make sure no OOM, and use len_opt
"""
if sum(len(para) for para in batch.data['x']) == 0:
new_JX = 1
else:
new_JX = max(len(para) for para in batch.data['x'])
JX = min(JX, new_JX)
# print(JX)
x = np.zeros([N, M, JX], dtype='int32')
cx = np.zeros([N, M, JX, W], dtype='int32')
x_mask = np.zeros([N, M, JX], dtype='bool')
q = np.zeros([N, JQ], dtype='int32')
cq = np.zeros([N, JQ, W], dtype='int32')
q_mask = np.zeros([N, JQ], dtype='bool')
feed_dict[self.x] = x
feed_dict[self.x_mask] = x_mask
feed_dict[self.cx] = cx
feed_dict[self.q] = q
feed_dict[self.cq] = cq
feed_dict[self.q_mask] = q_mask
feed_dict[self.is_train] = is_train
if config.use_glove_for_unk:
feed_dict[self.new_emb_mat] = batch.shared['new_emb_mat']
X = batch.data['x']
CX = batch.data['cx']
def _get_word(word):
if word.startswith("@"):
return 2
d = batch.shared['word2idx']
for each in (word, word.lower(), word.capitalize(), word.upper()):
if each in d:
return d[each]
if config.use_glove_for_unk:
d2 = batch.shared['new_word2idx']
for each in (word, word.lower(), word.capitalize(), word.upper()):
if each in d2:
return d2[each] + len(d)
return 1
def _get_char(char):
d = batch.shared['char2idx']
if char in d:
return d[char]
return 1
if supervised:
y = np.zeros([N, M, JX], dtype='int32')
feed_dict[self.y] = y
for i, (xi, yi) in enumerate(zip(batch.data['x'], batch.data['y'])):
count = 0
for j, xij in enumerate(xi):
for k, xijk in enumerate(xij):
if xijk == yi:
y[i, j, k] = True
count += 1
assert count > 0
for i, xi in enumerate(X):
for j, xij in enumerate(xi):
for k, xijk in enumerate(xij):
each = _get_word(xijk)
x[i, j, k] = each
x_mask[i, j, k] = True
for i, cxi in enumerate(CX):
for j, cxij in enumerate(cxi):
for k, cxijk in enumerate(cxij):
for l, cxijkl in enumerate(cxijk):
cx[i, j, k, l] = _get_char(cxijkl)
if l + 1 == config.max_word_size:
break
for i, qi in enumerate(batch.data['q']):
for j, qij in enumerate(qi):
q[i, j] = _get_word(qij)
q_mask[i, j] = True
for i, cqi in enumerate(batch.data['cq']):
for j, cqij in enumerate(cqi):
for k, cqijk in enumerate(cqij):
cq[i, j, k] = _get_char(cqijk)
if k + 1 == config.max_word_size:
break
return feed_dict
def get_multi_gpu_models(config):
models = []
for gpu_idx in range(config.num_gpus):
with tf.name_scope("model_{}".format(gpu_idx)) as scope, tf.device("/gpu:{}".format(gpu_idx)):
model = Model(config, scope)
tf.get_variable_scope().reuse_variables()
models.append(model)
return models
