def traditional_attention(rep_tensor,
rep_mask,
scope=None,
keep_prob=1.,
is_train=None,
wd=0.,
activation='elu',
tensor_dict=None,
name=None):
bs, sl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(
rep_tensor)[2]
ivec = rep_tensor.get_shape()[2]
with tf.variable_scope(scope or 'traditional_attention'):
rep_tensor_map = bn_dense_layer(rep_tensor, ivec, True, 0.,
'bn_dense_map', activation, False, wd,
keep_prob, is_train)
rep_tensor_logits = get_logits([rep_tensor_map],
None,
False,
scope='self_attn_logits',
mask=rep_mask,
input_keep_prob=keep_prob,
is_train=is_train) # bs,sl
attn_res = softsel(rep_tensor, rep_tensor_logits, rep_mask) # bs,vec
# save attn
if tensor_dict is not None and name is not None:
tensor_dict[name] = tf.nn.softmax(rep_tensor_logits)
return attn_res
def visit_multi_dimensional_attention(rep_tensor,
keep_prob=1.,
is_train=None,
wd=0.,
activation='relu'):
# bs, sl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(rep_tensor)[2]
ivec = rep_tensor.get_shape()[2]
with tf.variable_scope('multi_dimensional_attention'):
map1 = bn_dense_layer(rep_tensor, ivec, True, 0., 'bn_dense_map1',
activation, False, wd, keep_prob, is_train)
map2 = bn_dense_layer(map1, ivec, True, 0., 'bn_dense_map2', 'linear',
False, wd, keep_prob, is_train)
soft = tf.nn.softmax(map2, 1) # bs,sl,vec
attn_output = tf.reduce_sum(soft * rep_tensor, 1) # bs, vec
return attn_output
def bi_sru_recurrent_network(
rep_tensor, rep_mask, is_train=None, keep_prob=1., wd=0.,
scope=None, hn=None, reuse=None):
"""
:param rep_tensor: [Tensor/tf.float32] rank is 3 with shape [batch_size/bs, max_sent_len/sl, vec]
:param rep_mask: [Tensor/tf.bool]rank is 2 with shape [bs,sl]
:param is_train: [Scalar Tensor/tf.bool]scalar tensor to indicate whether the mode is training or not
:param keep_prob: [float] dropout keep probability in the range of (0,1)
:param wd: [float]for L2 regularization, if !=0, add tensors to tf collection "reg_vars"
:param scope: [str]variable scope name
:param hn:
:param
:return: [Tensor/tf.float32] with shape [bs, sl, 2vec] for forward and backward
"""
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]
ivec = hn or ivec
with tf.variable_scope(scope or 'bi_sru_recurrent_network'):
# U_d = bn_dense_layer([rep_tensor], 6 * ivec, False, 0., 'get_frc', 'linear',
# False, wd, keep_prob, is_train) # bs, sl, 6vec
# U_d_fw, U_d_bw = tf.split(U_d, 2, 2)
with tf.variable_scope('forward'):
U_d_fw = bn_dense_layer([rep_tensor], 3 * ivec, False, 0., 'get_frc_fw', 'linear',
False, wd, keep_prob, is_train) # bs, sl, 6vec
U_fw = tf.concat([rep_tensor, U_d_fw], -1)
fw_SRUCell = SwitchableDropoutWrapper(SRUCell(ivec, tf.nn.tanh, reuse), is_train, keep_prob)
fw_output, _ = dynamic_rnn(
fw_SRUCell, U_fw, tf.reduce_sum(tf.cast(rep_mask, tf.int32), -1),
dtype=tf.float32, scope='forward_sru') # bs, sl, vec
with tf.variable_scope('backward'):
U_d_bw = bn_dense_layer([rep_tensor], 3 * ivec, False, 0., 'get_frc_bw', 'linear',
False, wd, keep_prob, is_train) # bs, sl, 6vec
U_bw = tf.concat([rep_tensor, U_d_bw], -1)
bw_SRUCell = SwitchableDropoutWrapper(SRUCell(ivec, tf.nn.tanh, reuse), is_train, keep_prob)
bw_output, _ = bw_dynamic_rnn(
bw_SRUCell, U_bw, tf.reduce_sum(tf.cast(rep_mask, tf.int32), -1),
dtype=tf.float32, scope='backward_sru') # bs, sl, vec
all_output = tf.concat([fw_output, bw_output], -1) # bs, sl, 2vec
return all_output
def first_level_sa(rep_tensor,
rep_mask,
keep_prob=1.,
is_train=None,
wd=0.,
activation='relu'):
# bs, sw, cl, vec = tf.shape(rep_tensor)[0], tf.shape(rep_tensor)[1], tf.shape(rep_tensor)[2], tf.shape(rep_tensor)[3]
ivec = rep_tensor.get_shape()[3]
with tf.variable_scope('first_level_sa'):
map1 = bn_dense_layer(rep_tensor, ivec, True, 0., 'bn_dense_map1',
activation, False, wd, keep_prob, is_train)
map2 = bn_dense_layer(map1, ivec, True, 0., 'bn_dense_map2', 'linear',
False, wd, keep_prob, is_train)
map2_masked = exp_mask_for_high_rank(map2, rep_mask)
soft = tf.nn.softmax(map2_masked, 2) # bs,sk,code_len,vec
attn_output = tf.reduce_sum(soft * rep_tensor, 2) # bs, sk, vec
return attn_output
def normal_attention(rep_tensor,
rep_mask,
scope=None,
keep_prob=1.,
is_train=None,
wd=0.,
activation='elu',
tensor_dict=None,
name=None):
batch_size, code_len, vec_size = tf.shape(rep_tensor)[0], tf.shape(
rep_tensor)[1], tf.shape(rep_tensor)[2]
ivec = rep_tensor.get_shape()[2]
with tf.variable_scope(scope or 'normal_attention'):
rep_tensor_map = bn_dense_layer(rep_tensor, ivec, True, 0.,
'bn_dense_map', activation, False, wd,
keep_prob, is_train)
rep_tensor_logits = get_logits([rep_tensor_map],
None,
False,
scope='self_attn_logits',
mask=rep_mask,
input_keep_prob=keep_prob,
is_train=is_train) # bs,sl
attn_result = softsel(rep_tensor, rep_tensor_logits,
rep_mask) # bs,vec
# save attn
if tensor_dict is not None and name is not None:
tensor_dict[name] = tf.nn.softmax(rep_tensor_logits)
with tf.variable_scope('output'):
o_bias = tf.get_variable('o_bias', [ivec], tf.float32,
tf.constant_initializer(0.))
# input gate
fusion_gate = tf.nn.sigmoid(
linear(rep_tensor_map, ivec, True, 0., 'linear_fusion_i',
False, wd, keep_prob, is_train) +
linear(attn_result, ivec, True, 0., 'linear_fusion_a', False,
wd, keep_prob, is_train) + o_bias)
output = fusion_gate * rep_tensor_map + (1 -
fusion_gate) * attn_result
output = mask_for_high_rank(output, rep_mask) # bs,sl,vec
return output
def __call__(self, inputs, state, scope=None):
"""
:param inputs: [bs, vec]
:param state:
:param scope:
:return:
"""
with tf.variable_scope(scope or "SRU_cell"):
b_f = tf.get_variable('b_f', [self._num_units], dtype=tf.float32,
initializer=tf.constant_initializer(0))
b_r = tf.get_variable('b_r', [self._num_units], dtype=tf.float32,
initializer=tf.constant_initializer(0))
U_d = bn_dense_layer(inputs, 3 * self._num_units, False, 0., 'get_frc', 'linear') # bs, 3vec
x_t = tf.identity(inputs, 'x_t')
x_dt, f_t, r_t = tf.split(U_d, 3, 1)
f_t = tf.nn.sigmoid(f_t + b_f)
r_t = tf.nn.sigmoid(r_t + b_r)
c_t = f_t * state + (1 - f_t) * x_dt
h_t = r_t * self._activation(c_t) + (1 - r_t) * x_t
return h_t, c_t
def visit_sa_with_dense(rep_tensor,
keep_prob=1.,
is_train=None,
wd=0.,
activation='relu',
hn=None,
is_scale=True,
is_plus_sa=True):
batch_size, sw_len, vec_size = tf.shape(rep_tensor)[0], tf.shape(
rep_tensor)[1], tf.shape(rep_tensor)[2]
ivec = rep_tensor.get_shape().as_list()[2]
ivec = hn or ivec
with tf.variable_scope('temporal_attention'):
# mask generation
attn_mask = tf.cast(
tf.diag(-tf.ones([sw_len], tf.int32)) + 1,
tf.bool) # batch_size, code_len, code_len
# non-linear for context
rep_map = bn_dense_layer(rep_tensor, ivec, True, 0., 'bn_dense_map',
activation, False, wd, keep_prob, is_train)
rep_map_tile = tf.tile(tf.expand_dims(rep_map, 1),
[1, sw_len, 1, 1]) # bs,sl,sl,vec
rep_map_dp = dropout(rep_map, keep_prob, is_train)
# attention
with tf.variable_scope('attention'): # bs,sl,sl,vec
f_bias = tf.get_variable('f_bias', [ivec], tf.float32,
tf.constant_initializer(0.))
dependent = linear(
rep_map_dp, ivec, False,
scope='linear_dependent') # batch_size, code_len, vec_size
dependent_etd = tf.expand_dims(
dependent, 1) # batch_size, code_len,code_len, vec_size
head = linear(
rep_map_dp, ivec, False,
scope='linear_head') # batch_size, code_len, vec_size
head_etd = tf.expand_dims(
head, 2) # batch_size, code_len,code_len, vec_size
if is_plus_sa:
attention_fact = dependent_etd + head_etd + f_bias
else:
return rep_map
if is_scale:
logits = scaled_tanh(attention_fact, 5.0) # bs,sl,sl,vec
else:
logits = linear(tf.nn.tanh(attention_fact),
ivec,
True,
scope='linear_attn_fact')
logits_masked = exp_mask_for_high_rank(logits, attn_mask)
attn_score = tf.nn.softmax(logits_masked, 2) # bs,sl,sl,vec
attn_score = mask_for_high_rank(attn_score, attn_mask)
attn_result = tf.reduce_sum(attn_score * rep_map_tile,
2) # bs,sl,vec
with tf.variable_scope('output'):
o_bias = tf.get_variable('o_bias', [ivec], tf.float32,
tf.constant_initializer(0.))
# input gate
fusion_gate = tf.nn.sigmoid(
linear(rep_map, ivec, True, 0., 'linear_fusion_i', False, wd,
keep_prob, is_train) +
linear(attn_result, ivec, True, 0., 'linear_fusion_a', False,
wd, keep_prob, is_train) + o_bias)
output = fusion_gate * rep_map + (1 - fusion_gate) * attn_result
return output
def directional_attention_with_dense(rep_tensor,
rep_mask,
direction=None,
scope=None,
keep_prob=1.,
is_train=None,
wd=0.,
activation='elu',
tensor_dict=None,
name=None,
hn=None):
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]
ivec = hn or ivec
with tf.variable_scope(scope or 'directional_attention_%s' % direction
or 'diag'):
# mask generation
sl_indices = tf.range(sl, dtype=tf.int32)
sl_col, sl_row = tf.meshgrid(sl_indices, sl_indices)
if direction is None:
direct_mask = tf.cast(
tf.diag(-tf.ones([sl], tf.int32)) + 1, tf.bool)
else:
if direction == 'forward':
direct_mask = tf.greater(sl_row, sl_col)
else:
direct_mask = tf.greater(sl_col, sl_row)
direct_mask_tile = tf.tile(tf.expand_dims(direct_mask, 0),
[bs, 1, 1]) # bs,sl,sl
rep_mask_tile = tf.tile(tf.expand_dims(rep_mask, 1),
[1, sl, 1]) # bs,sl,sl
attn_mask = tf.logical_and(direct_mask_tile, rep_mask_tile) # bs,sl,sl
# non-linear
rep_map = bn_dense_layer(rep_tensor, ivec, True, 0., 'bn_dense_map',
activation, False, wd, keep_prob, is_train)
rep_map_tile = tf.tile(tf.expand_dims(rep_map, 1),
[1, sl, 1, 1]) # bs,sl,sl,vec
rep_map_dp = dropout(rep_map, keep_prob, is_train)
# attention
with tf.variable_scope('attention'): # bs,sl,sl,vec
f_bias = tf.get_variable('f_bias', [ivec], tf.float32,
tf.constant_initializer(0.))
dependent = linear(rep_map_dp,
ivec,
False,
scope='linear_dependent') # bs,sl,vec
dependent_etd = tf.expand_dims(dependent, 1) # bs,1,sl,vec
head = linear(rep_map_dp, ivec, False,
scope='linear_head') # bs,sl,vec
head_etd = tf.expand_dims(head, 2) # bs,sl,1,vec
logits = scaled_tanh(dependent_etd + head_etd + f_bias,
5.0) # bs,sl,sl,vec
logits_masked = exp_mask_for_high_rank(logits, attn_mask)
attn_score = tf.nn.softmax(logits_masked, 2) # bs,sl,sl,vec
attn_score = mask_for_high_rank(attn_score, attn_mask)
attn_result = tf.reduce_sum(attn_score * rep_map_tile,
2) # bs,sl,vec
with tf.variable_scope('output'):
o_bias = tf.get_variable('o_bias', [ivec], tf.float32,
tf.constant_initializer(0.))
# input gate
fusion_gate = tf.nn.sigmoid(
linear(rep_map, ivec, True, 0., 'linear_fusion_i', False, wd,
keep_prob, is_train) +
linear(attn_result, ivec, True, 0., 'linear_fusion_a', False,
wd, keep_prob, is_train) + o_bias)
output = fusion_gate * rep_map + (1 - fusion_gate) * attn_result
output = mask_for_high_rank(output, rep_mask)
# save attn
if tensor_dict is not None and name is not None:
tensor_dict[name + '_dependent'] = dependent
tensor_dict[name + '_head'] = head
tensor_dict[name] = attn_score
tensor_dict[name + '_gate'] = fusion_gate
return output
