def average_attention_strategy(strategy, x, mask, state, layer, params):
strategy = strategy.lower()
is_training = ('decoder' not in state)
if strategy == "aan":
if is_training:
if params.aan_mask:
aan_bias = func.attention_bias(mask, "aan")
x_fwd = tf.matmul(aan_bias, x)
else:
aan_bias = tf.cumsum(mask, axis=1)
aan_bias = tf.where(tf.less_equal(aan_bias, 0.),
tf.ones_like(aan_bias), aan_bias)
aan_bias = tf.expand_dims(dtype.tf_to_float(aan_bias), 2)
x_fwd = tf.cumsum(x, axis=1) / aan_bias
else:
cache = state['decoder']['state']['layer_{}'.format(layer)]
x_fwd = (x + cache['aan']) / dtype.tf_to_float(state['time'] + 1)
cache['aan'] = x + cache['aan']
return x_fwd
else:
raise NotImplementedError("Not supported: {}".format(strategy))
def add_timing_signal(x, min_timescale=1.0, max_timescale=1.0e4,
time=None, name=None):
"""Transformer Positional Embedding"""
with tf.name_scope(name, default_name="add_timing_signal", values=[x]):
length = tf.shape(x)[1]
channels = tf.shape(x)[2]
if time is None:
position = dtype.tf_to_float(tf.range(length))
else:
# decoding position embedding
position = tf.expand_dims(time, 0)
num_timescales = channels // 2
log_timescale_increment = (
math.log(float(max_timescale) / float(min_timescale)) /
(dtype.tf_to_float(num_timescales) - 1)
)
inv_timescales = min_timescale * tf.exp(
dtype.tf_to_float(tf.range(num_timescales)) * -log_timescale_increment
)
scaled_time = (tf.expand_dims(position, 1) *
tf.expand_dims(inv_timescales, 0))
signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=1)
signal = tf.pad(signal, [[0, 0], [0, tf.mod(channels, 2)]])
signal = tf.reshape(signal, [1, length, channels])
return x + signal
def extract_encodes(source_memory, source_mask, l0_mask):
x_shp = util.shape_list(source_memory)
l0_mask = dtype.tf_to_float(tf.cast(l0_mask, tf.bool))
l0_mask = tf.squeeze(l0_mask, -1) * source_mask
# count retained encodings
k_value = tf.cast(tf.reduce_max(tf.reduce_sum(l0_mask, 1)), tf.int32)
# batch_size x k_value
_, topk_indices = tf.nn.top_k(l0_mask, k_value)
# prepare coordinate
x_pos = util.batch_coordinates(x_shp[0], k_value)
coord = tf.stack([x_pos, topk_indices], axis=2)
# gather retained features
g_x = tf.gather_nd(source_memory, coord)
g_mask = tf.gather_nd(l0_mask, coord)
# padding zero
g_x = tf.pad(g_x, [[0, 0], [1, 0], [0, 0]])
# generate counts, i.e. how many tokens are dropped
droped_number = tf.reduce_sum(source_mask, 1) - tf.reduce_sum(l0_mask, 1)
pad_mask = dtype.tf_to_float(tf.greater(droped_number, 0.))
droped_number = tf.where(tf.less_equal(droped_number, 0.),
tf.ones_like(droped_number), droped_number)
count_mask = tf.ones_like(g_mask)
count_mask = tf.concat([tf.expand_dims(droped_number, 1), count_mask], 1)
g_mask = tf.concat([tf.expand_dims(pad_mask, 1), g_mask], 1)
return g_x, g_mask, count_mask
def attention_bias(inputs, mode, inf=None, name=None):
""" A bias tensor used in attention mechanism"""
if inf is None:
inf = dtype.inf()
with tf.name_scope(name, default_name="attention_bias", values=[inputs]):
if mode == "causal":
length = inputs
lower_triangle = tf.matrix_band_part(
tf.ones([length, length]), -1, 0
)
ret = dtype.tf_to_float(- inf * (1.0 - lower_triangle))
return tf.reshape(ret, [1, 1, length, length])
elif mode == "masking":
mask = inputs
ret = (1.0 - mask) * - inf
return tf.expand_dims(tf.expand_dims(ret, 1), 1)
elif mode == "aan":
length = tf.shape(inputs)[1]
diagonal = tf.eye(length)
cum_factor = tf.expand_dims(tf.cumsum(diagonal, axis=0), 0)
mask = tf.expand_dims(inputs, 1) * tf.expand_dims(inputs, 2)
mask *= dtype.tf_to_float(cum_factor)
weight = tf.nn.softmax(mask + (1.0 - mask) * - inf)
weight *= mask
return weight
else:
raise ValueError("Unknown mode %s" % mode)
def _step_fn(prev, x):
t, h_, c_, a_ = prev
if not one2one:
m, v = x[mask_pos], x[:mask_pos]
else:
c, c_c, m, v = x[-1], x[mask_pos + 1:-1], x[mask_pos], x[:mask_pos]
s = cell_lower(h_, v)
s = m * s + (1. - m) * h_
if not one2one:
vle = additive_attention(cell_lower.get_hidden(s),
memory,
mem_mask,
mem_shape[-1],
ln=ln,
num_heads=num_heads,
proj_memory=proj_memories,
scope="attention")
a, c = vle['weights'], vle['output']
c_c = cell_higher.fetch_states(c)
else:
a = tf.tile(tf.expand_dims(tf.range(time_steps), 0),
[batch_size, 1])
a = dtype.tf_to_float(tf.equal(a, t))
a = tf.tile(tf.expand_dims(a, 1), [1, num_heads, 1])
a = tf.reshape(a, tf.shape(init_weight))
h = cell_higher(s, c_c)
h = m * h + (1. - m) * s
return t + 1, h, c, a
def rnn(cell_name, x, d, mask=None, ln=False, init_state=None, sm=True):
"""Self implemented RNN procedure, supporting mask trick"""
# cell_name: gru, lstm or atr
# x: input sequence embedding matrix, [batch, seq_len, dim]
# d: hidden dimension for rnn
# mask: mask matrix, [batch, seq_len]
# ln: whether use layer normalization
# init_state: the initial hidden states, for cache purpose
# sm: whether apply swap memory during rnn scan
# dp: variational dropout
in_shape = util.shape_list(x)
batch_size, time_steps = in_shape[:2]
cell = get_cell(cell_name, d, ln=ln)
if init_state is None:
init_state = cell.get_init_state(shape=[batch_size])
if mask is None:
mask = dtype.tf_to_float(tf.ones([batch_size, time_steps]))
# prepare projected input
cache_inputs = cell.fetch_states(x)
cache_inputs = [tf.transpose(v, [1, 0, 2]) for v in list(cache_inputs)]
mask_ta = tf.transpose(tf.expand_dims(mask, -1), [1, 0, 2])
def _step_fn(prev, x):
t, h_ = prev
m = x[-1]
v = x[:-1]
h = cell(h_, v)
h = m * h + (1. - m) * h_
return t + 1, h
time = tf.constant(0, dtype=tf.int32, name="time")
step_states = (time, init_state)
step_vars = cache_inputs + [mask_ta]
outputs = tf.scan(_step_fn,
step_vars,
initializer=step_states,
parallel_iterations=32,
swap_memory=sm)
output_ta = outputs[1]
output_state = outputs[1][-1]
outputs = tf.transpose(output_ta, [1, 0, 2])
return (outputs, output_state), \
(cell.get_hidden(outputs), cell.get_hidden(output_state))
def _get_init_state(self, d, shape=None, x=None, scope=None):
# gen init state vector
# if no evidence x is provided, use zero initialization
if x is None:
assert shape is not None, "you should provide shape"
if not isinstance(shape, (tuple, list)):
shape = [shape]
shape = shape + [d]
return dtype.tf_to_float(tf.zeros(shape))
else:
return linear(x,
d,
bias=True,
ln=self.ln,
scope="{}_init".format(scope or self.scope))
def encoder(source, params):
mask = dtype.tf_to_float(tf.cast(source, tf.bool))
hidden_size = params.hidden_size
source, mask = util.remove_invalid_seq(source, mask)
embed_name = "embedding" if params.shared_source_target_embedding \
else "src_embedding"
src_emb = tf.get_variable(embed_name,
[params.src_vocab.size(), params.embed_size])
src_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(src_emb, source)
inputs = tf.nn.bias_add(inputs, src_bias)
inputs = util.valid_apply_dropout(inputs, params.dropout)
with tf.variable_scope("encoder"):
x = inputs
for layer in range(params.num_encoder_layer):
with tf.variable_scope("layer_{}".format(layer)):
# forward rnn
with tf.variable_scope('forward'):
outputs = rnn.rnn(params.cell,
x,
hidden_size,
mask=mask,
ln=params.layer_norm,
sm=params.swap_memory)
output_fw, state_fw = outputs[1]
if layer == 0:
# backward rnn
with tf.variable_scope('backward'):
if not params.caencoder:
outputs = rnn.rnn(params.cell,
tf.reverse(x, [1]),
hidden_size,
mask=tf.reverse(mask, [1]),
ln=params.layer_norm,
sm=params.swap_memory)
output_bw, state_bw = outputs[1]
else:
outputs = rnn.cond_rnn(params.cell,
tf.reverse(x, [1]),
tf.reverse(output_fw, [1]),
hidden_size,
mask=tf.reverse(mask, [1]),
ln=params.layer_norm,
sm=params.swap_memory,
num_heads=params.num_heads,
one2one=True)
output_bw, state_bw = outputs[1]
output_bw = tf.reverse(output_bw, [1])
if not params.caencoder:
y = tf.concat([output_fw, output_bw], -1)
z = tf.concat([state_fw, state_bw], -1)
else:
y = output_bw
z = state_bw
else:
y = output_fw
z = state_fw
y = func.linear(y, params.embed_size, ln=False, scope="ff")
# short cut via residual connection
if x.get_shape()[-1].value == y.get_shape()[-1].value:
x = func.residual_fn(x, y, dropout=params.dropout)
else:
x = y
if params.layer_norm:
x = func.layer_norm(x, scope="ln")
if params.embed_size != hidden_size:
x = func.layer_norm(func.linear(x, hidden_size, scope="x_map"))
with tf.variable_scope("decoder_initializer"):
decoder_cell = rnn.get_cell(params.cell,
hidden_size,
ln=params.layer_norm)
return {
"encodes": x,
"decoder_initializer": {
"layer_{}".format(l):
decoder_cell.get_init_state(x=z, scope="layer_{}".format(l))
for l in range(params.num_decoder_layer)
},
"mask": mask
}
def encoder(source, params):
mask = dtype.tf_to_float(tf.cast(source, tf.bool))
hidden_size = params.hidden_size
source, mask = util.remove_invalid_seq(source, mask)
embed_name = "embedding" if params.shared_source_target_embedding \
else "src_embedding"
src_emb = tf.get_variable(embed_name,
[params.src_vocab.size(), params.embed_size])
src_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(src_emb, source)
inputs = tf.nn.bias_add(inputs, src_bias)
inputs = util.valid_apply_dropout(inputs, params.dropout)
with tf.variable_scope("encoder"):
# forward rnn
with tf.variable_scope('forward'):
outputs = rnn.rnn(params.cell,
inputs,
hidden_size,
mask=mask,
ln=params.layer_norm,
sm=params.swap_memory)
output_fw, state_fw = outputs[1]
# backward rnn
with tf.variable_scope('backward'):
if not params.caencoder:
outputs = rnn.rnn(params.cell,
tf.reverse(inputs, [1]),
hidden_size,
mask=tf.reverse(mask, [1]),
ln=params.layer_norm,
sm=params.swap_memory)
output_bw, state_bw = outputs[1]
else:
outputs = rnn.cond_rnn(params.cell,
tf.reverse(inputs, [1]),
tf.reverse(output_fw, [1]),
hidden_size,
mask=tf.reverse(mask, [1]),
ln=params.layer_norm,
sm=params.swap_memory,
num_heads=params.num_heads,
one2one=True)
output_bw, state_bw = outputs[1]
output_bw = tf.reverse(output_bw, [1])
if not params.caencoder:
source_encodes = tf.concat([output_fw, output_bw], -1)
source_feature = tf.concat([state_fw, state_bw], -1)
else:
source_encodes = output_bw
source_feature = state_bw
with tf.variable_scope("decoder_initializer"):
decoder_init = rnn.get_cell(
params.cell, hidden_size,
ln=params.layer_norm).get_init_state(x=source_feature)
decoder_init = tf.tanh(decoder_init)
return {
"encodes": source_encodes,
"decoder_initializer": decoder_init,
"mask": mask
}
def cond_rnn(cell_name,
x,
memory,
d,
init_state=None,
mask=None,
mem_mask=None,
ln=False,
sm=True,
one2one=False,
num_heads=1):
"""Self implemented conditional-RNN procedure, supporting mask trick"""
# cell_name: gru, lstm or atr
# x: input sequence embedding matrix, [batch, seq_len, dim]
# memory: the conditional part
# d: hidden dimension for rnn
# mask: mask matrix, [batch, seq_len]
# mem_mask: memory mask matrix, [batch, mem_seq_len]
# ln: whether use layer normalization
# init_state: the initial hidden states, for cache purpose
# sm: whether apply swap memory during rnn scan
# one2one: whether the memory is one-to-one mapping for x
# num_heads: number of attention heads, multi-head attention
# dp: variational dropout
in_shape = util.shape_list(x)
batch_size, time_steps = in_shape[:2]
mem_shape = util.shape_list(memory)
cell_lower = get_cell(cell_name,
d,
ln=ln,
scope="{}_lower".format(cell_name))
cell_higher = get_cell(cell_name,
d,
ln=ln,
scope="{}_higher".format(cell_name))
if init_state is None:
init_state = cell_lower.get_init_state(shape=[batch_size])
if mask is None:
mask = dtype.tf_to_float(tf.ones([batch_size, time_steps]))
if mem_mask is None:
mem_mask = dtype.tf_to_float(tf.ones([batch_size, mem_shape[1]]))
# prepare projected encodes and inputs
cache_inputs = cell_lower.fetch_states(x)
cache_inputs = [tf.transpose(v, [1, 0, 2]) for v in list(cache_inputs)]
if not one2one:
proj_memories = linear(memory,
mem_shape[-1],
bias=False,
ln=ln,
scope="context_att")
else:
cache_memories = cell_higher.fetch_states(memory)
cache_memories = [
tf.transpose(v, [1, 0, 2]) for v in list(cache_memories)
]
mask_ta = tf.transpose(tf.expand_dims(mask, -1), [1, 0, 2])
init_context = dtype.tf_to_float(tf.zeros([batch_size, mem_shape[-1]]))
init_weight = dtype.tf_to_float(
tf.zeros([batch_size, num_heads, mem_shape[1]]))
mask_pos = len(cache_inputs)
def _step_fn(prev, x):
t, h_, c_, a_ = prev
if not one2one:
m, v = x[mask_pos], x[:mask_pos]
else:
c, c_c, m, v = x[-1], x[mask_pos + 1:-1], x[mask_pos], x[:mask_pos]
s = cell_lower(h_, v)
s = m * s + (1. - m) * h_
if not one2one:
vle = additive_attention(cell_lower.get_hidden(s),
memory,
mem_mask,
mem_shape[-1],
ln=ln,
num_heads=num_heads,
proj_memory=proj_memories,
scope="attention")
a, c = vle['weights'], vle['output']
c_c = cell_higher.fetch_states(c)
else:
a = tf.tile(tf.expand_dims(tf.range(time_steps), 0),
[batch_size, 1])
a = dtype.tf_to_float(tf.equal(a, t))
a = tf.tile(tf.expand_dims(a, 1), [1, num_heads, 1])
a = tf.reshape(a, tf.shape(init_weight))
h = cell_higher(s, c_c)
h = m * h + (1. - m) * s
return t + 1, h, c, a
time = tf.constant(0, dtype=tf.int32, name="time")
step_states = (time, init_state, init_context, init_weight)
step_vars = cache_inputs + [mask_ta]
if one2one:
step_vars += cache_memories + [memory]
outputs = tf.scan(_step_fn,
step_vars,
initializer=step_states,
parallel_iterations=32,
swap_memory=sm)
output_ta = outputs[1]
context_ta = outputs[2]
attention_ta = outputs[3]
outputs = tf.transpose(output_ta, [1, 0, 2])
output_states = outputs[:, -1]
contexts = tf.transpose(context_ta, [1, 0, 2])
attentions = tf.transpose(attention_ta, [1, 2, 0, 3])
return (outputs, output_states), \
(cell_higher.get_hidden(outputs), cell_higher.get_hidden(output_states)), \
contexts, attentions
def dot_attention(query, memory, mem_mask, hidden_size,
ln=False, num_heads=1, cache=None, dropout=None,
use_relative_pos=False, max_relative_position=16,
out_map=True, scope=None, fuse_mask=None,
decode_step=None):
"""
dotted attention model
:param query: [batch_size, qey_len, dim]
:param memory: [batch_size, seq_len, mem_dim] or None
:param mem_mask: [batch_size, seq_len]
:param hidden_size: attention space dimension
:param ln: whether use layer normalization
:param num_heads: attention head number
:param dropout: attention dropout, default disable
:param out_map: output additional mapping
:param cache: cache-based decoding
:param fuse_mask: aan mask during training, and timestep for testing
:param max_relative_position: maximum position considered for relative embedding
:param use_relative_pos: whether use relative position information
:param decode_step: the time step of current decoding, 0-based
:param scope:
:return: a value matrix, [batch_size, qey_len, mem_dim]
"""
with tf.variable_scope(scope or "dot_attention", reuse=tf.AUTO_REUSE,
dtype=tf.as_dtype(dtype.floatx())):
if fuse_mask is not None:
assert memory is not None, 'Fuse mechanism only applied with cross-attention'
if cache and use_relative_pos:
assert decode_step is not None, 'Decode Step must provide when use relative position encoding'
if memory is None:
# suppose self-attention from queries alone
h = linear(query, hidden_size * 3, ln=ln, scope="qkv_map")
q, k, v = tf.split(h, 3, -1)
if cache is not None:
k = tf.concat([cache['k'], k], axis=1)
v = tf.concat([cache['v'], v], axis=1)
cache = {
'k': k,
'v': v,
}
else:
q = linear(query, hidden_size, ln=ln, scope="q_map")
if cache is not None and ('mk' in cache and 'mv' in cache):
k, v = cache['mk'], cache['mv']
else:
k = linear(memory, hidden_size, ln=ln, scope="k_map")
v = linear(memory, hidden_size, ln=ln, scope="v_map")
if cache is not None:
cache['mk'] = k
cache['mv'] = v
q = split_heads(q, num_heads)
k = split_heads(k, num_heads)
v = split_heads(v, num_heads)
q *= (hidden_size // num_heads) ** (-0.5)
q_shp = util.shape_list(q)
k_shp = util.shape_list(k)
v_shp = util.shape_list(v)
q_len = q_shp[2] if decode_step is None else decode_step + 1
r_lst = None if decode_step is None else 1
# q * k => attention weights
if use_relative_pos:
r = rpr.get_relative_positions_embeddings(
q_len, k_shp[2], k_shp[3],
max_relative_position, name="rpr_keys", last=r_lst)
logits = rpr.relative_attention_inner(q, k, r, transpose=True)
else:
logits = tf.matmul(q, k, transpose_b=True)
if mem_mask is not None:
logits += mem_mask
weights = tf.nn.softmax(logits)
dweights = util.valid_apply_dropout(weights, dropout)
# weights * v => attention vectors
if use_relative_pos:
r = rpr.get_relative_positions_embeddings(
q_len, k_shp[2], v_shp[3],
max_relative_position, name="rpr_values", last=r_lst)
o = rpr.relative_attention_inner(dweights, v, r, transpose=False)
else:
o = tf.matmul(dweights, v)
o = combine_heads(o)
if fuse_mask is not None:
# This is for AAN, the important part is sharing v_map
v_q = linear(query, hidden_size, ln=ln, scope="v_map")
if cache is not None and 'aan' in cache:
aan_o = (v_q + cache['aan']) / dtype.tf_to_float(fuse_mask + 1)
else:
# Simplified Average Attention Network
aan_o = tf.matmul(fuse_mask, v_q)
if cache is not None:
if 'aan' not in cache:
cache['aan'] = v_q
else:
cache['aan'] = v_q + cache['aan']
# Directly sum both self-attention and cross attention
o = o + aan_o
if out_map:
o = linear(o, hidden_size, ln=ln, scope="o_map")
results = {
'weights': weights,
'output': o,
'cache': cache
}
return results
def decoder(target, state, params):
mask = dtype.tf_to_float(tf.cast(target, tf.bool))
hidden_size = params.hidden_size
initializer = tf.random_normal_initializer(0.0, hidden_size**-0.5)
is_training = ('decoder' not in state)
if is_training:
target, mask = util.remove_invalid_seq(target, mask)
embed_name = "embedding" if params.shared_source_target_embedding \
else "tgt_embedding"
tgt_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size],
initializer=initializer)
tgt_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(tgt_emb, target) * (hidden_size**0.5)
inputs = tf.nn.bias_add(inputs, tgt_bias)
# shift
if is_training:
inputs = tf.pad(inputs, [[0, 0], [1, 0], [0, 0]])
inputs = inputs[:, :-1, :]
inputs = func.add_timing_signal(inputs)
else:
inputs = tf.cond(
tf.reduce_all(tf.equal(target, params.tgt_vocab.pad())),
lambda: tf.zeros_like(inputs), lambda: inputs)
mask = tf.ones_like(mask)
inputs = func.add_timing_signal(inputs,
time=dtype.tf_to_float(state['time']))
inputs = util.valid_apply_dropout(inputs, params.dropout)
with tf.variable_scope("decoder"):
x = inputs
for layer in range(params.num_decoder_layer):
if params.deep_transformer_init:
layer_initializer = tf.variance_scaling_initializer(
params.initializer_gain * (layer + 1)**-0.5,
mode="fan_avg",
distribution="uniform")
else:
layer_initializer = None
with tf.variable_scope("layer_{}".format(layer),
initializer=layer_initializer):
with tf.variable_scope("average_attention"):
x_fwds = []
for strategy in params.strategies:
with tf.variable_scope(strategy):
x_fwd = average_attention_strategy(
strategy, x, mask, state, layer, params)
x_fwds.append(x_fwd)
x_fwd = tf.add_n(x_fwds) / len(x_fwds)
# FFN activation
if params.use_ffn:
y = func.ffn_layer(
x_fwd,
params.filter_size,
hidden_size,
dropout=params.relu_dropout,
)
else:
y = x_fwd
# Gating layer
z = func.linear(tf.concat([x, y], axis=-1),
hidden_size * 2,
scope="z_project")
i, f = tf.split(z, 2, axis=-1)
y = tf.sigmoid(i) * x + tf.sigmoid(f) * y
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
with tf.variable_scope("cross_attention"):
y = func.dot_attention(
x,
state['encodes'],
func.attention_bias(state['mask'], "masking"),
hidden_size,
num_heads=params.num_heads,
dropout=params.attention_dropout,
cache=None if is_training else
state['decoder']['state']['layer_{}'.format(layer)])
if not is_training:
# mk, mv
state['decoder']['state']['layer_{}'.format(layer)]\
.update(y['cache'])
y = y['output']
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
with tf.variable_scope("feed_forward"):
y = func.ffn_layer(
x,
params.filter_size,
hidden_size,
dropout=params.relu_dropout,
)
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
feature = x
if 'dev_decode' in state:
feature = x[:, -1, :]
embed_name = "tgt_embedding" if params.shared_target_softmax_embedding \
else "softmax_embedding"
embed_name = "embedding" if params.shared_source_target_embedding \
else embed_name
softmax_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size],
initializer=initializer)
feature = tf.reshape(feature, [-1, params.embed_size])
logits = tf.matmul(feature, softmax_emb, False, True)
logits = tf.cast(logits, tf.float32)
soft_label, normalizer = util.label_smooth(target,
util.shape_list(logits)[-1],
factor=params.label_smooth)
centropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,
labels=soft_label)
centropy -= normalizer
centropy = tf.reshape(centropy, tf.shape(target))
mask = tf.cast(mask, tf.float32)
per_sample_loss = tf.reduce_sum(centropy * mask, -1) / tf.reduce_sum(
mask, -1)
loss = tf.reduce_mean(per_sample_loss)
# these mask tricks mainly used to deal with zero shapes, such as [0, 1]
loss = tf.cond(tf.equal(tf.shape(target)[0], 0),
lambda: tf.constant(0, dtype=tf.float32), lambda: loss)
return loss, logits, state, per_sample_loss
def encoder(source, params):
mask = dtype.tf_to_float(tf.cast(source, tf.bool))
hidden_size = params.hidden_size
initializer = tf.random_normal_initializer(0.0, hidden_size**-0.5)
source, mask = util.remove_invalid_seq(source, mask)
embed_name = "embedding" if params.shared_source_target_embedding \
else "src_embedding"
src_emb = tf.get_variable(embed_name,
[params.src_vocab.size(), params.embed_size],
initializer=initializer)
src_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(src_emb, source) * (hidden_size**0.5)
inputs = tf.nn.bias_add(inputs, src_bias)
inputs = func.add_timing_signal(inputs)
inputs = util.valid_apply_dropout(inputs, params.dropout)
with tf.variable_scope("encoder"):
x = inputs
for layer in range(params.num_encoder_layer):
if params.deep_transformer_init:
layer_initializer = tf.variance_scaling_initializer(
params.initializer_gain * (layer + 1)**-0.5,
mode="fan_avg",
distribution="uniform")
else:
layer_initializer = None
with tf.variable_scope("layer_{}".format(layer),
initializer=layer_initializer):
with tf.variable_scope("self_attention"):
y = func.dot_attention(x,
None,
func.attention_bias(
mask, "masking"),
hidden_size,
num_heads=params.num_heads,
dropout=params.attention_dropout)
y = y['output']
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
with tf.variable_scope("feed_forward"):
y = func.ffn_layer(
x,
params.filter_size,
hidden_size,
dropout=params.relu_dropout,
)
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
source_encodes = x
x_shp = util.shape_list(x)
return {
"encodes": source_encodes,
"decoder_initializer": {
"layer_{}".format(l): {
# plan aan
"aan": dtype.tf_to_float(tf.zeros([x_shp[0], 1, hidden_size])),
}
for l in range(params.num_decoder_layer)
},
"mask": mask
}
def decoder(target, state, params):
mask = dtype.tf_to_float(tf.cast(target, tf.bool))
hidden_size = params.hidden_size
initializer = tf.random_normal_initializer(0.0, hidden_size**-0.5)
is_training = ('decoder' not in state)
if is_training:
target, mask = util.remove_invalid_seq(target, mask)
embed_name = "embedding" if params.shared_source_target_embedding \
else "tgt_embedding"
tgt_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size],
initializer=initializer)
tgt_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(tgt_emb, target) * (hidden_size**0.5)
inputs = tf.nn.bias_add(inputs, tgt_bias)
# shift
if is_training:
inputs = tf.pad(inputs, [[0, 0], [1, 0], [0, 0]])
inputs = inputs[:, :-1, :]
inputs = func.add_timing_signal(inputs)
else:
inputs = tf.cond(
tf.reduce_all(tf.equal(target, params.tgt_vocab.pad())),
lambda: tf.zeros_like(inputs), lambda: inputs)
mask = tf.ones_like(mask)
inputs = func.add_timing_signal(inputs,
time=dtype.tf_to_float(state['time']))
inputs = util.valid_apply_dropout(inputs, params.dropout)
# Applying L0Drop
# --------
source_memory = state["encodes"]
source_mask = state["mask"]
# source_pruning: log alpha_i = x_i w^T
source_pruning = func.linear(source_memory, 1, scope="source_pruning")
if is_training: # training
source_memory, l0_mask = l0norm.var_train(
(source_memory, source_pruning))
l0_norm_loss = tf.squeeze(l0norm.l0_norm(source_pruning), -1)
l0_norm_loss = tf.reduce_sum(l0_norm_loss * source_mask,
-1) / tf.reduce_sum(source_mask, -1)
l0_norm_loss = tf.reduce_mean(l0_norm_loss)
l0_norm_loss = l0norm.l0_regularization_loss(
l0_norm_loss,
reg_scalar=params.l0_norm_reg_scalar,
start_reg_ramp_up=params.l0_norm_start_reg_ramp_up,
end_reg_ramp_up=params.l0_norm_end_reg_ramp_up,
warm_up=params.l0_norm_warm_up,
)
# force the model to only attend to unmasked position
source_mask = dtype.tf_to_float(
tf.cast(tf.squeeze(l0_mask, -1), tf.bool)) * source_mask
else: # evaluation
source_memory, l0_mask = l0norm.var_eval(
(source_memory, source_pruning))
l0_norm_loss = 0.0
source_memory, source_mask, count_mask = extract_encodes(
source_memory, source_mask, l0_mask)
count_mask = tf.expand_dims(tf.expand_dims(count_mask, 1), 1)
# --------
with tf.variable_scope("decoder"):
x = inputs
for layer in range(params.num_decoder_layer):
if params.deep_transformer_init:
layer_initializer = tf.variance_scaling_initializer(
params.initializer_gain * (layer + 1)**-0.5,
mode="fan_avg",
distribution="uniform")
else:
layer_initializer = None
with tf.variable_scope("layer_{}".format(layer),
initializer=layer_initializer):
with tf.variable_scope("self_attention"):
y = func.dot_attention(
x,
None,
func.attention_bias(tf.shape(mask)[1], "causal"),
hidden_size,
num_heads=params.num_heads,
dropout=params.attention_dropout,
cache=None if is_training else
state['decoder']['state']['layer_{}'.format(layer)])
if not is_training:
# k, v
state['decoder']['state']['layer_{}'.format(layer)] \
.update(y['cache'])
y = y['output']
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
with tf.variable_scope("cross_attention"):
if is_training:
y = func.dot_attention(
x,
source_memory,
func.attention_bias(source_mask, "masking"),
hidden_size,
num_heads=params.num_heads,
dropout=params.attention_dropout,
)
else:
y = dot_attention(x,
source_memory,
func.attention_bias(
source_mask, "masking"),
hidden_size,
count_mask=count_mask,
num_heads=params.num_heads,
dropout=params.attention_dropout,
cache=state['decoder']['state'][
'layer_{}'.format(layer)])
# mk, mv
state['decoder']['state']['layer_{}'.format(layer)] \
.update(y['cache'])
y = y['output']
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
with tf.variable_scope("feed_forward"):
y = func.ffn_layer(
x,
params.filter_size,
hidden_size,
dropout=params.relu_dropout,
)
x = func.residual_fn(x, y, dropout=params.residual_dropout)
x = func.layer_norm(x)
feature = x
if 'dev_decode' in state:
feature = x[:, -1, :]
embed_name = "tgt_embedding" if params.shared_target_softmax_embedding \
else "softmax_embedding"
embed_name = "embedding" if params.shared_source_target_embedding \
else embed_name
softmax_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size],
initializer=initializer)
feature = tf.reshape(feature, [-1, params.embed_size])
logits = tf.matmul(feature, softmax_emb, False, True)
logits = tf.cast(logits, tf.float32)
soft_label, normalizer = util.label_smooth(target,
util.shape_list(logits)[-1],
factor=params.label_smooth)
centropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,
labels=soft_label)
centropy -= normalizer
centropy = tf.reshape(centropy, tf.shape(target))
mask = tf.cast(mask, tf.float32)
per_sample_loss = tf.reduce_sum(centropy * mask, -1) / tf.reduce_sum(
mask, -1)
loss = tf.reduce_mean(per_sample_loss)
loss = loss + l0_norm_loss
# these mask tricks mainly used to deal with zero shapes, such as [0, 1]
loss = tf.cond(tf.equal(tf.shape(target)[0], 0),
lambda: tf.constant(0, tf.float32), lambda: loss)
return loss, logits, state, per_sample_loss
def deep_att_dec_rnn(cell_name,
x,
memory,
d,
init_state=None,
mask=None,
mem_mask=None,
ln=False,
sm=True,
depth=1,
num_heads=1):
"""Self implemented conditional-RNN procedure, supporting mask trick"""
# cell_name: gru, lstm or atr
# x: input sequence embedding matrix, [batch, seq_len, dim]
# memory: the conditional part
# d: hidden dimension for rnn
# mask: mask matrix, [batch, seq_len]
# mem_mask: memory mask matrix, [batch, mem_seq_len]
# ln: whether use layer normalization
# init_state: the initial hidden states, for cache purpose
# sm: whether apply swap memory during rnn scan
# depth: depth for the decoder in deep attention
# num_heads: number of attention heads, multi-head attention
# dp: variational dropout
in_shape = util.shape_list(x)
batch_size, time_steps = in_shape[:2]
mem_shape = util.shape_list(memory)
cell_lower = rnn.get_cell(cell_name,
d,
ln=ln,
scope="{}_lower".format(cell_name))
cells_higher = []
for layer in range(depth):
cell_higher = rnn.get_cell(cell_name,
d,
ln=ln,
scope="{}_higher_{}".format(
cell_name, layer))
cells_higher.append(cell_higher)
if init_state is None:
init_state = cell_lower.get_init_state(shape=[batch_size])
if mask is None:
mask = dtype.tf_to_float(tf.ones([batch_size, time_steps]))
if mem_mask is None:
mem_mask = dtype.tf_to_float(tf.ones([batch_size, mem_shape[1]]))
# prepare projected encodes and inputs
cache_inputs = cell_lower.fetch_states(x)
cache_inputs = [tf.transpose(v, [1, 0, 2]) for v in list(cache_inputs)]
proj_memories = func.linear(memory,
mem_shape[-1],
bias=False,
ln=ln,
scope="context_att")
mask_ta = tf.transpose(tf.expand_dims(mask, -1), [1, 0, 2])
init_context = dtype.tf_to_float(
tf.zeros([batch_size, depth, mem_shape[-1]]))
init_weight = dtype.tf_to_float(
tf.zeros([batch_size, depth, num_heads, mem_shape[1]]))
mask_pos = len(cache_inputs)
def _step_fn(prev, x):
t, h_, c_, a_ = prev
m, v = x[mask_pos], x[:mask_pos]
# the first decoder rnn subcell, composing previous hidden state with the current word embedding
s_ = cell_lower(h_, v)
s_ = m * s_ + (1. - m) * h_
atts, att_ctxs = [], []
for layer in range(depth):
# perform attention
prev_cell = cell_lower if layer == 0 else cells_higher[layer - 1]
vle = func.additive_attention(
prev_cell.get_hidden(s_),
memory,
mem_mask,
mem_shape[-1],
ln=ln,
num_heads=num_heads,
proj_memory=proj_memories,
scope="deep_attention_{}".format(layer))
a, c = vle['weights'], vle['output']
atts.append(tf.expand_dims(a, 1))
att_ctxs.append(tf.expand_dims(c, 1))
# perform next-level recurrence
c_c = cells_higher[layer].fetch_states(c)
ss_ = cells_higher[layer](s_, c_c)
s_ = m * ss_ + (1. - m) * s_
h = s_
a = tf.concat(atts, axis=1)
c = tf.concat(att_ctxs, axis=1)
return t + 1, h, c, a
time = tf.constant(0, dtype=tf.int32, name="time")
step_states = (time, init_state, init_context, init_weight)
step_vars = cache_inputs + [mask_ta]
outputs = tf.scan(_step_fn,
step_vars,
initializer=step_states,
parallel_iterations=32,
swap_memory=sm)
output_ta = outputs[1]
context_ta = outputs[2]
attention_ta = outputs[3]
outputs = tf.transpose(output_ta, [1, 0, 2])
output_states = outputs[:, -1]
# batch x target length x depth x mem-dimension
contexts = tf.transpose(context_ta, [1, 0, 2, 3])
# batch x num_heads x depth x target length x source length
attentions = tf.transpose(attention_ta, [1, 3, 2, 0, 4])
return (outputs, output_states), \
(cells_higher[-1].get_hidden(outputs), cells_higher[-1].get_hidden(output_states)), \
contexts, attentions
def decoder(target, state, params):
mask = dtype.tf_to_float(tf.cast(target, tf.bool))
hidden_size = params.hidden_size
is_training = ('decoder' not in state)
if is_training:
target, mask = util.remove_invalid_seq(target, mask)
embed_name = "embedding" if params.shared_source_target_embedding \
else "tgt_embedding"
tgt_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size])
tgt_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(tgt_emb, target)
inputs = tf.nn.bias_add(inputs, tgt_bias)
# shift
if is_training:
inputs = tf.pad(inputs, [[0, 0], [1, 0], [0, 0]])
inputs = inputs[:, :-1, :]
else:
inputs = tf.cond(
tf.reduce_all(tf.equal(target, params.tgt_vocab.pad())),
lambda: tf.zeros_like(inputs), lambda: inputs)
mask = tf.ones_like(mask)
inputs = util.valid_apply_dropout(inputs, params.dropout)
with tf.variable_scope("decoder"):
x = inputs
for layer in range(params.num_decoder_layer):
with tf.variable_scope("layer_{}".format(layer)):
init_state = state["decoder_initializer"]["layer_{}".format(
layer)]
if not is_training:
init_state = state["decoder"]["state"]["layer_{}".format(
layer)]
if layer == 0 or params.use_deep_att:
returns = rnn.cond_rnn(params.cell,
x,
state["encodes"],
hidden_size,
init_state=init_state,
mask=mask,
num_heads=params.num_heads,
mem_mask=state["mask"],
ln=params.layer_norm,
sm=params.swap_memory,
one2one=False)
(_, hidden_state), (outputs,
_), contexts, attentions = returns
c = contexts
else:
if params.caencoder:
returns = rnn.cond_rnn(params.cell,
x,
c,
hidden_size,
init_state=init_state,
mask=mask,
mem_mask=mask,
ln=params.layer_norm,
sm=params.swap_memory,
num_heads=params.num_heads,
one2one=True)
(_, hidden_state), (outputs,
_), contexts, attentions = returns
else:
outputs = rnn.rnn(params.cell,
tf.concat([x, c], -1),
hidden_size,
mask=mask,
init_state=init_state,
ln=params.layer_norm,
sm=params.swap_memory)
outputs, hidden_state = outputs[1]
if not is_training:
state['decoder']['state']['layer_{}'.format(
layer)] = hidden_state
y = func.linear(outputs,
params.embed_size,
ln=False,
scope="ff")
# short cut via residual connection
if x.get_shape()[-1].value == y.get_shape()[-1].value:
x = func.residual_fn(x, y, dropout=params.dropout)
else:
x = y
if params.layer_norm:
x = func.layer_norm(x, scope="ln")
if params.dl4mt_redict:
feature = func.linear(tf.concat([x, c], -1),
params.embed_size,
ln=params.layer_norm,
scope="ff")
feature = tf.nn.tanh(feature)
feature = util.valid_apply_dropout(feature, params.dropout)
else:
feature = x
if 'dev_decode' in state:
feature = x[:, -1, :]
embed_name = "tgt_embedding" if params.shared_target_softmax_embedding \
else "softmax_embedding"
embed_name = "embedding" if params.shared_source_target_embedding \
else embed_name
softmax_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size])
feature = tf.reshape(feature, [-1, params.embed_size])
logits = tf.matmul(feature, softmax_emb, False, True)
logits = tf.cast(logits, tf.float32)
soft_label, normalizer = util.label_smooth(target,
util.shape_list(logits)[-1],
factor=params.label_smooth)
centropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,
labels=soft_label)
centropy -= normalizer
centropy = tf.reshape(centropy, tf.shape(target))
mask = tf.cast(mask, tf.float32)
per_sample_loss = tf.reduce_sum(centropy * mask, -1) / tf.reduce_sum(
mask, -1)
loss = tf.reduce_mean(per_sample_loss)
# these mask tricks mainly used to deal with zero shapes, such as [0, 1]
loss = tf.cond(tf.equal(tf.shape(target)[0], 0),
lambda: tf.constant(0, dtype=tf.float32), lambda: loss)
return loss, logits, state, per_sample_loss
def encoder(source, params):
mask = dtype.tf_to_float(tf.cast(source, tf.bool))
hidden_size = params.hidden_size
source, mask = util.remove_invalid_seq(source, mask)
# extract source word embedding and apply dropout
embed_name = "embedding" if params.shared_source_target_embedding \
else "src_embedding"
src_emb = tf.get_variable(embed_name,
[params.src_vocab.size(), params.embed_size])
src_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(src_emb, source)
inputs = tf.nn.bias_add(inputs, src_bias)
inputs = util.valid_apply_dropout(inputs, params.dropout)
# the encoder module used in the deep attention paper
with tf.variable_scope("encoder"):
# x: embedding input, h: the hidden state
x = inputs
h = 0
z = 0
for layer in range(params.num_encoder_layer + 1):
with tf.variable_scope("layer_{}".format(layer)):
if layer == 0:
# for the first layer, we perform a normal rnn layer to collect context information
outputs = rnn.rnn(params.cell,
x,
hidden_size,
mask=mask,
ln=params.layer_norm,
sm=params.swap_memory)
h = outputs[1][0]
else:
# for deeper encoder layers, we incorporate both embedding input and previous inversed hidden
# state sequence as input.
# the embedding informs current input while hidden state tells future context
is_reverse = (layer % 2 == 1)
outputs = rnn.cond_rnn(
params.cell,
tf.reverse(x, [1]) if is_reverse else x,
tf.reverse(h, [1]) if is_reverse else h,
hidden_size,
mask=tf.reverse(mask, [1]) if is_reverse else mask,
ln=params.layer_norm,
sm=params.swap_memory,
num_heads=params.num_heads,
one2one=True)
h = outputs[1][0]
h = tf.reverse(h, [1]) if is_reverse else h
# the final hidden state used for decoder state initialization
z = outputs[1][1]
with tf.variable_scope("decoder_initializer"):
decoder_cell = rnn.get_cell(params.cell,
hidden_size,
ln=params.layer_norm)
return {
"encodes": h,
"decoder_initializer": {
'layer': decoder_cell.get_init_state(x=z, scope="dec_init_state")
},
"mask": mask
}
def decoder(target, state, params):
mask = dtype.tf_to_float(tf.cast(target, tf.bool))
hidden_size = params.hidden_size
is_training = ('decoder' not in state)
# handling target-side word embedding, including shift-padding for training
embed_name = "embedding" if params.shared_source_target_embedding \
else "tgt_embedding"
tgt_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size])
tgt_bias = tf.get_variable("bias", [params.embed_size])
inputs = tf.gather(tgt_emb, target)
inputs = tf.nn.bias_add(inputs, tgt_bias)
# shift
if is_training:
inputs = tf.pad(inputs, [[0, 0], [1, 0], [0, 0]])
inputs = inputs[:, :-1, :]
else:
inputs = tf.cond(
tf.reduce_all(tf.equal(target, params.tgt_vocab.pad())),
lambda: tf.zeros_like(inputs), lambda: inputs)
mask = tf.ones_like(mask)
inputs = util.valid_apply_dropout(inputs, params.dropout)
with tf.variable_scope("decoder"):
x = inputs
init_state = state["decoder_initializer"]["layer"]
if not is_training:
init_state = state["decoder"]["state"]["layer"]
returns = deep_att_dec_rnn(params.cell,
x,
state["encodes"],
hidden_size,
init_state=init_state,
mask=mask,
num_heads=params.num_heads,
mem_mask=state["mask"],
ln=params.layer_norm,
sm=params.swap_memory,
depth=params.num_decoder_layer)
(_, hidden_state), (outputs, _), contexts, attentions = returns
if not is_training:
state['decoder']['state']['layer'] = hidden_state
x = outputs
cshp = util.shape_list(contexts)
c = tf.reshape(contexts, [cshp[0], cshp[1], cshp[2] * cshp[3]])
feature = func.linear(tf.concat([x, c, inputs], -1),
params.embed_size,
ln=params.layer_norm,
scope="ff")
feature = tf.nn.tanh(feature)
feature = util.valid_apply_dropout(feature, params.dropout)
if 'dev_decode' in state:
feature = feature[:, -1, :]
embed_name = "tgt_embedding" if params.shared_target_softmax_embedding \
else "softmax_embedding"
embed_name = "embedding" if params.shared_source_target_embedding \
else embed_name
softmax_emb = tf.get_variable(embed_name,
[params.tgt_vocab.size(), params.embed_size])
feature = tf.reshape(feature, [-1, params.embed_size])
logits = tf.matmul(feature, softmax_emb, False, True)
logits = tf.cast(logits, tf.float32)
soft_label, normalizer = util.label_smooth(target,
util.shape_list(logits)[-1],
factor=params.label_smooth)
centropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,
labels=soft_label)
centropy -= normalizer
centropy = tf.reshape(centropy, tf.shape(target))
mask = tf.cast(mask, tf.float32)
per_sample_loss = tf.reduce_sum(centropy * mask, -1) / tf.reduce_sum(
mask, -1)
loss = tf.reduce_mean(per_sample_loss)
# these mask tricks mainly used to deal with zero shapes, such as [0, 1]
loss = tf.cond(tf.equal(tf.shape(target)[0], 0),
lambda: tf.constant(0, dtype=tf.float32), lambda: loss)
return loss, logits, state, per_sample_loss
