def linear(args,
output_size,
bias,
bias_start=0.0,
scope=None,
squeeze=False,
wd=0.0,
input_keep_prob=1.0,
is_train=None):
if args is None or (isinstance(args, (tuple, list)) and not args):
raise ValueError("`args` must be specified")
if not isinstance(args, (tuple, list)):
args = [args]
flat_args = [flatten(arg, 1) for arg in args] # for dense layer [(-1, d)]
if input_keep_prob < 1.0:
assert is_train is not None
flat_args = [
tf.cond(is_train, lambda: tf.nn.dropout(arg, input_keep_prob),
lambda: arg) # for dense layer [(-1, d)]
for arg in flat_args
]
flat_out = _linear(flat_args,
output_size,
bias,
bias_start=bias_start,
scope=scope) # dense
out = reconstruct(flat_out, args[0], 1) # ()
if squeeze:
out = tf.squeeze(out, [len(args[0].get_shape().as_list()) - 1])
if wd:
add_reg_without_bias()
return out
def cnn_for_sentence_encoding( # kim
rep_tensor, rep_mask, filter_sizes=(3,4,5), num_filters=200, scope=None,
is_train=None, keep_prob=1., wd=0.):
"""
:param rep_tensor:
:param rep_mask:
:param filter_sizes:
:param num_filters:
:param scope:
:param is_train:
:param keep_prob:
:param wd:
:return:
"""
bs, sl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(rep_tensor)[2]
ivec = rep_tensor.get_shape().as_list()[2]
with tf.variable_scope(scope or 'cnn_for_sentence_encoding'):
rep_tensor = mask_for_high_rank(rep_tensor, rep_mask)
rep_tensor_expand = tf.expand_dims(rep_tensor, 3)
rep_tensor_expand_dp = dropout(rep_tensor_expand, keep_prob, is_train)
# Create a convolution + maxpool layer for each filter size
pooled_outputs = []
for i, filter_size in enumerate(filter_sizes):
with tf.variable_scope("conv-maxpool-%s" % filter_size):
# Convolution Layer
filter_shape = [filter_size, ivec, 1, num_filters]
W = tf.get_variable('W', filter_shape, tf.float32)
b = tf.get_variable('b', [num_filters], tf.float32)
conv = tf.nn.conv2d(
rep_tensor_expand_dp,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
# Apply nonlinearity
h = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu") # bs, sl-fs+1, 1, fn
# Maxpooling over the outputs
# pooled = tf.nn.max_pool(
# h,
# ksize=[1, sl - filter_size + 1, 1, 1],
# strides=[1, 1, 1, 1],
# padding='VALID',
# name="pool")
pooled = tf.reduce_max(h, 1, True) # bs, 1, 1, fn
pooled_outputs.append(pooled)
# Combine all the pooled features
num_filters_total = num_filters * len(filter_sizes)
h_pool = tf.concat(pooled_outputs, 3)
h_pool_flat = tf.reshape(h_pool, [-1, num_filters_total])
if wd > 0.:
add_reg_without_bias()
return h_pool_flat
def contextual_bi_rnn(tensor_rep,
mask_rep,
hn,
cell_type,
only_final=False,
wd=0.,
keep_prob=1.,
is_train=None,
scope=None):
"""
fusing contextual information using bi-direction rnn
:param tensor_rep: [..., sl, vec]
:param mask_rep: [..., sl]
:param hn:
:param cell_type: 'gru', 'lstm', basic_lstm' and 'basic_rnn'
:param only_final: True or False
:param wd:
:param keep_prob:
:param is_train:
:param scope:
:return:
"""
with tf.variable_scope(scope or 'contextual_bi_rnn'): # correct
reuse = None if not tf.get_variable_scope().reuse else True
#print(reuse)
if cell_type == 'gru':
cell_fw = tf.contrib.rnn.GRUCell(hn, reuse=reuse)
cell_bw = tf.contrib.rnn.GRUCell(hn, reuse=reuse)
elif cell_type == 'lstm':
cell_fw = tf.contrib.rnn.LSTMCell(hn, reuse=reuse)
cell_bw = tf.contrib.rnn.LSTMCell(hn, reuse=reuse)
elif cell_type == 'basic_lstm':
cell_fw = tf.contrib.rnn.BasicLSTMCell(hn, reuse=reuse)
cell_bw = tf.contrib.rnn.BasicLSTMCell(hn, reuse=reuse)
elif cell_type == 'basic_rnn':
cell_fw = tf.contrib.rnn.BasicRNNCell(hn, reuse=reuse)
cell_bw = tf.contrib.rnn.BasicRNNCell(hn, reuse=reuse)
else:
raise AttributeError('no cell type \'%s\'' % cell_type)
cell_dp_fw = SwitchableDropoutWrapper(cell_fw, is_train, keep_prob)
cell_dp_bw = SwitchableDropoutWrapper(cell_bw, is_train, keep_prob)
tensor_len = tf.reduce_sum(tf.cast(mask_rep, tf.int32), -1) # [bs]
(outputs_fw,
output_bw), _ = bidirectional_dynamic_rnn(cell_dp_fw,
cell_dp_bw,
tensor_rep,
tensor_len,
dtype=tf.float32)
rnn_outputs = tf.concat([outputs_fw, output_bw], -1) # [...,sl,2hn]
if wd > 0:
add_reg_without_bias()
if not only_final:
return rnn_outputs # [....,sl, 2hn]
else:
return get_last_state(rnn_outputs, mask_rep) # [...., 2hn]
def hierarchical_cnn_res_gate(
rep_tensor, rep_mask, n_gram=5, layer_num=5, hn=None, scope=None,
is_train=None, keep_prob=1., wd=0.):
# padding
if n_gram % 2 == 1:
padding_front = padding_back = int((n_gram - 1) / 2)
else:
padding_front = (n_gram - 1) // 2
padding_back = padding_front + 1
padding = [[0, 0], [padding_front, padding_back], [0, 0], [0, 0]]
# lengths
bs, sl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(rep_tensor)[2]
org_ivec = rep_tensor.get_shape().as_list()[2]
ivec = hn or org_ivec
with tf.variable_scope(scope or 'cnn_for_sentence_encoding'):
rep_tensor = mask_for_high_rank(rep_tensor, rep_mask) # bs, sl, hn
iter_rep = rep_tensor
layer_res_list = []
for layer_idx in range(layer_num):
with tf.variable_scope("conv_maxpool_%s" % layer_idx):
iter_rep_etd = tf.expand_dims(iter_rep, 3) # bs,sl,hn,1
iter_rep_etd_dp = dropout(iter_rep_etd, keep_prob, is_train)
# Convolution Layer
feature_size = org_ivec if layer_idx == 0 else ivec
filter_shape = [n_gram, feature_size, 1, 2 * ivec]
W = tf.get_variable('W', filter_shape, tf.float32)
b = tf.get_variable('b', [2 * ivec], tf.float32)
iter_rep_etd_pad = tf.pad(iter_rep_etd_dp, padding)
conv = tf.nn.conv2d(
iter_rep_etd_pad,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
map_res = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu") # bs,sl,1,2hn
map_res = tf.squeeze(map_res, [2]) # bs,sl,2*hn
# gate
map_res_a, map_res_b = tf.split(map_res, num_or_size_splits=2, axis=2)
iter_rep = map_res_a * tf.nn.sigmoid(map_res_b)
# res
if len(layer_res_list) > 0:
iter_rep = iter_rep + layer_res_list[-1]
layer_res_list.append(iter_rep)
if wd > 0.:
add_reg_without_bias()
return iter_rep
def cnn_for_context_fusion(
rep_tensor, rep_mask, filter_sizes=(3,4,5), num_filters=200, scope=None,
is_train=None, keep_prob=1., wd=0.):
bs, sl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(rep_tensor)[2]
ivec = rep_tensor.get_shape().as_list()[2]
with tf.variable_scope(scope or 'cnn_for_sentence_encoding'):
rep_tensor = mask_for_high_rank(rep_tensor, rep_mask)
rep_tensor_expand = tf.expand_dims(rep_tensor, 3) # bs, sl,
rep_tensor_expand_dp = dropout(rep_tensor_expand, keep_prob, is_train)
# Create a convolution + maxpool layer for each filter size
pooled_outputs = []
for i, filter_size in enumerate(filter_sizes):
with tf.variable_scope("conv-maxpool-%s" % filter_size):
# Convolution Layer
filter_shape = [filter_size, ivec, 1, num_filters]
W = tf.get_variable('W', filter_shape, tf.float32)
b = tf.get_variable('b', [num_filters], tf.float32)
# # pading in the sequence
if filter_size % 2 == 1:
padding_front = padding_back = int((filter_size - 1) / 2)
else:
padding_front = (filter_size - 1) // 2
padding_back = padding_front + 1
padding = [[0, 0], [padding_front, padding_back], [0, 0], [0, 0]]
rep_tensor_expand_dp_pad = tf.pad(rep_tensor_expand_dp, padding)
conv = tf.nn.conv2d(
rep_tensor_expand_dp_pad,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv")
# Apply nonlinearity
h = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu") # bs, sl, 1, fn
h_squeeze = tf.squeeze(h, [2]) # bs, sl, fn
pooled_outputs.append(h_squeeze)
# Combine all the pooled features
result = tf.concat(pooled_outputs, 2) # bs, sl, 3 * fn
if wd > 0.:
add_reg_without_bias()
return result
def multi_head_attention(rep_tensor,
rep_mask,
head_num=8,
hidden_units_num=64,
scope=None,
is_train=None,
keep_prob=1.,
wd=0.):
bs, sl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(
rep_tensor)[2]
ivec = rep_tensor.get_shape().as_list()[2]
with tf.variable_scope(scope or 'multi_head_attention'):
with tf.variable_scope('positional_encoding'):
seq_idxs = tf.tile(tf.expand_dims(tf.range(sl), 1),
[1, ivec]) # sl, ivec
feature_idxs = tf.tile(tf.expand_dims(tf.range(ivec), 0),
[sl, 1]) # sl, ivec
pos_enc = tf.where(
tf.equal(tf.mod(feature_idxs, 2), 0),
tf.sin(
tf.cast(seq_idxs, tf.float32) / tf.pow(
10000., 2.0 * tf.cast(feature_idxs, tf.float32) /
(1.0 * ivec))),
tf.cos(
tf.cast(seq_idxs, tf.float32) / tf.pow(
10000., 2.0 * tf.cast(feature_idxs - 1, tf.float32) /
(1.0 * ivec))),
)
rep_tensor_pos = mask_for_high_rank(rep_tensor + pos_enc,
rep_mask) # bs, sl, ivec
with tf.variable_scope('multi_head_attention'):
W = tf.get_variable('W', [3, head_num, ivec, hidden_units_num],
tf.float32)
rep_tile = tf.tile(
tf.expand_dims(tf.expand_dims(rep_tensor_pos, 0), 0),
[3, head_num, 1, 1, 1]) # 3,head_num,bs,sl,ivec
rep_tile_reshape = tf.reshape(
rep_tile, [3, head_num, bs * sl, ivec]) # head_num,bs*sl,ivec
maps = tf.reshape( # 3,head_num,bs*sl,hn -> 3,head_num,bs,sl,hn
tf.matmul(dropout(rep_tile_reshape, keep_prob, is_train), W),
[3, head_num, bs, sl, hidden_units_num])
Q_map, K_map, V_map = tf.split(maps, 3, 0)
Q_map = tf.squeeze(Q_map, [0]) # head_num,bs,sl,hn
K_map = tf.squeeze(K_map, [0]) # head_num,bs,sl,hn
V_map = tf.squeeze(V_map, [0]) # head_num,bs,sl,hn
# head_num,bs,sl,sl
# similarity_mat = tf.reduce_sum(Q_map_tile * K_map_tile, -1) / math.sqrt(1. * hidden_units_num)
similarity_mat = tf.matmul(Q_map, tf.transpose(
K_map, [0, 1, 3, 2])) / math.sqrt(1. * hidden_units_num)
# mask: bs,sl -> head_num,bs,sl
multi_mask = tf.tile(tf.expand_dims(rep_mask, 0),
[head_num, 1, 1]) # head_num,bs,sl
multi_mask_tile_1 = tf.expand_dims(multi_mask,
2) # head_num,bs,1,sl
multi_mask_tile_2 = tf.expand_dims(multi_mask,
3) # head_num,bs,sl,1
multi_mask_tile = tf.logical_and(
multi_mask_tile_1, multi_mask_tile_2) # head_num,bs,sl,sl
similarity_mat_masked = exp_mask(
similarity_mat, multi_mask_tile) # head_num,bs,sl,sl
prob_dist = tf.nn.softmax(
similarity_mat_masked) # head_num,bs,sl,sl
prob_dist_dp = dropout(prob_dist, keep_prob, is_train)
attn_res = tf.matmul(prob_dist_dp, V_map) # head_num,bs,sl,hn
attn_res_tran = tf.transpose(attn_res, [1, 2, 0, 3])
output = tf.reshape(attn_res_tran,
[bs, sl, head_num * hidden_units_num])
if wd > 0.:
add_reg_without_bias()
return output
