def forward(self, features):
def FC(inputs, name, i, act):
return L.fc(inputs,
self.hidden_size,
act=act,
param_attr=F.ParamAttr(
name='%s.fc.w_%d' % (name, i),
initializer=F.initializer.XavierInitializer(
fan_in=self.hidden_size,
fan_out=self.hidden_size)),
bias_attr=F.ParamAttr(
name='%s.fc.b_%d' % (name, i),
initializer=F.initializer.Constant(0.)))
title_ids, comment_ids = features
embedding_attr = F.ParamAttr(
name='emb',
initializer=F.initializer.XavierInitializer(
fan_in=self.vocab_size, fan_out=self.embedding_size))
title_encoded = L.embedding(title_ids,
[self.vocab_size, self.embedding_size],
param_attr=embedding_attr)
comment_encoded = L.embedding(comment_ids,
[self.vocab_size, self.embedding_size],
param_attr=embedding_attr)
# Vsum
zero = L.fill_constant(shape=[1], dtype='int64', value=0)
title_pad = L.cast(L.logical_not(L.equal(title_ids, zero)), 'float32')
comment_pad = L.cast(L.logical_not(L.equal(comment_ids, zero)),
'float32')
title_encoded = L.reduce_sum(title_encoded * title_pad, dim=1)
title_encoded = L.softsign(title_encoded)
comment_encoded = L.reduce_sum(comment_encoded * comment_pad, dim=1)
comment_encoded = L.softsign(comment_encoded)
for i in range(self.num_layers):
title_encoded = FC(title_encoded, 'title', i, 'tanh')
for i in range(self.num_layers):
comment_encoded = FC(comment_encoded, 'comment', i, 'tanh')
score = L.reduce_sum(title_encoded * comment_encoded,
dim=1,
keep_dim=True) / np.sqrt(self.hidden_size)
if self.mode is propeller.RunMode.PREDICT:
probs = L.sigmoid(score)
return probs
else:
return score
def _grammar_step(self, logits, next_cell_states, decode_states, actions, gmr_mask):
"""跟进文法约束完成一步解码逻辑
Args:
logits (Variable): shape = [batch_size, beam_size, vocab_size]
next_cell_states (Variable): NULL
decode_states (StateWrapper): NULL
Returns: TODO
Raises: NULL
"""
# 解码出符合语法规则的 token logits
logits, valid_table_mask = self._output_layer(logits, actions, gmr_mask, decode_states.valid_table_mask)
# 初始化 vocab size
self._vocab_size = logits.shape[-1]
self._vocab_size_tensor = layers.fill_constant(shape=[1], dtype='int64', value=logits.shape[-1])
# 计算 log probs，并 mask 掉 finished 部分
step_log_probs = layers.log(layers.softmax(logits))
step_log_probs = self._mask_finished_probs(step_log_probs, decode_states.finished)
scores = layers.reshape(step_log_probs, [-1, self._beam_size * self._vocab_size])
topk_scores, topk_indices = layers.topk(input=scores, k=self._beam_size)
topk_scores = layers.reshape(topk_scores, shape=[-1])
topk_indices = layers.reshape(topk_indices, shape=[-1])
# top-k 对应的 beam
beam_indices = layers.elementwise_floordiv(topk_indices, self._vocab_size_tensor)
# top-k 对应的 token id
token_indices = layers.elementwise_mod(topk_indices, self._vocab_size_tensor)
# 根据 top k 的来源，重新组织 step_log_probs
next_log_probs = nn_utils.batch_gather(
layers.reshape(step_log_probs, [-1, self._beam_size * self._vocab_size]),
topk_indices)
def _beam_gather(x, beam_indices):
"""reshape x to beam dim, and gather each beam_indices
Args:
x (TYPE): NULL
Returns: Variable
"""
x = self.split_batch_beams(x)
return nn_utils.batch_gather(x, beam_indices)
next_cell_states = layers.utils.map_structure(lambda x: _beam_gather(x, beam_indices),
next_cell_states)
next_finished = _beam_gather(decode_states.finished, beam_indices)
next_lens = _beam_gather(decode_states.lengths, beam_indices)
next_lens = layers.elementwise_add(next_lens,
layers.cast(layers.logical_not(next_finished), next_lens.dtype))
next_finished = layers.logical_or(next_finished,
layers.equal(token_indices, self._end_token_tensor))
decode_output = OutputWrapper(topk_scores, token_indices, beam_indices)
decode_states = StateWrapper(next_cell_states, next_log_probs, next_finished, next_lens, valid_table_mask)
return decode_output, decode_states
def is_finished(self, step_idx, source_length, alive_log_probs, finished_scores, finished_in_finished):
"""
is_finished
"""
base_1 = layers.cast(source_length, 'float32') + 55.0
base_1 /= 6.0
max_length_penalty = layers.pow(base_1, self.alpha)
flat_alive_log_probs = layers.reshape(alive_log_probs, [-1])
lower_bound_alive_scores_1 = layers.gather(flat_alive_log_probs, [self.get_alive_index])
lower_bound_alive_scores = lower_bound_alive_scores_1 / max_length_penalty
lowest_score_of_finished_in_finish = layers.reduce_min(finished_scores * finished_in_finished, dim=1)
finished_in_finished = layers.cast(finished_in_finished, 'bool')
lowest_score_of_finished_in_finish += \
((1.0 - layers.cast(layers.reduce_any(finished_in_finished, 1), 'float32')) * -INF)
#print lowest_score_of_finished_in_finish
bound_is_met = layers.reduce_all(layers.greater_than(lowest_score_of_finished_in_finish,
lower_bound_alive_scores))
decode_length = source_length + 50
length_cond = layers.less_than(x=step_idx, y=decode_length)
return layers.logical_and(x=layers.logical_not(bound_is_met), y=length_cond)
def mask_fill(input, mask, value):
"""Fill value to input according to mask
Args:
input: input matrix
mask: mask matrix
value: Fill value
Returns:
output
>>> input
[
[1, 2, 3],
[4, 5, 6]
]
>>> mask
[
[True, True, False],
[True, False, False]
]
>>> mask_fill(input, mask, 0)
[
[1, 2, 0],
[4, 0, 0]
]
"""
return input * layers.cast(layers.logical_not(
mask), input.dtype) + layers.cast(mask, input.dtype) * value
def _build_input_mask(self, src_ids):
zero = L.fill_constant([1], dtype='int64', value=0)
input_mask = L.logical_not(L.equal(src_ids,
zero)) # assume pad id == 0
input_mask = L.cast(input_mask, 'float')
input_mask.stop_gradient = True
return input_mask
def forward(self, features):
src_ids, sent_ids = features
dtype = 'float16' if self.hparam['fp16'] else 'float32'
zero = L.fill_constant([1], dtype='int64', value=0)
input_mask = L.cast(L.logical_not(L.equal(src_ids, zero)), dtype) # assume pad id == 0
#input_mask = L.unsqueeze(input_mask, axes=[2])
d_shape = L.shape(src_ids)
seqlen = d_shape[1]
batch_size = d_shape[0]
pos_ids = L.unsqueeze(L.range(0, seqlen, 1, dtype='int32'), axes=[0])
pos_ids = L.expand(pos_ids, [batch_size, 1])
pos_ids = L.unsqueeze(pos_ids, axes=[2])
pos_ids = L.cast(pos_ids, 'int64')
pos_ids.stop_gradient = True
input_mask.stop_gradient = True
task_ids = L.zeros_like(src_ids) + self.hparam.task_id #this shit wont use at the moment
task_ids.stop_gradient = True
bert = ErnieModel(
src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=self.hparam,
use_fp16=self.hparam['fp16']
)
cls_feats = bert.get_pooled_output()
cls_feats = L.dropout(
x=cls_feats,
dropout_prob=0.1,
dropout_implementation="upscale_in_train"
)
logits = L.fc(
input=cls_feats,
size=self.hparam['num_label'],
param_attr=F.ParamAttr(
name="cls_out_w",
initializer=F.initializer.TruncatedNormal(scale=0.02)),
bias_attr=F.ParamAttr(
name="cls_out_b", initializer=F.initializer.Constant(0.))
)
propeller.summary.histogram('pred', logits)
if self.mode is propeller.RunMode.PREDICT:
probs = L.softmax(logits)
return probs
else:
return logits
def bow(ids):
embed = L.embedding(
input=ids,
size=[self.config.vocab_size, self.config.emb_size],
dtype=self._emb_dtype,
param_attr=F.ParamAttr(name=self._word_emb_name,
initializer=self._param_initializer),
is_sparse=False)
zero = L.fill_constant(shape=[1], dtype='int64', value=0)
pad = L.cast(L.logical_not(L.equal(ids, zero)), 'float32')
sumed = L.reduce_sum(embed * pad, dim=1)
sumed = L.softsign(sumed)
return sumed
def _process_type_leaf(condition, decoder, grammar_stack, next_inputs,
finished):
"""Process when output type is LEAF
Args:
condition (TYPE): NULL
decoder (TYPE): NULL
grammar_stack (StackData): (gmr_stack_data, gmr_stack_pos)
next_inputs (DecoderInputsWrapper): (input_var, action, grammar_mask)
finished (TYPE): NULL
Returns: None
Raises: NULL
"""
## pop stack
next_output, valid_pos, gmr_stack_tmp = data_structure.Stack.pop(
grammar_stack, mask=True, in_place=False)
valid_pos = fluider.squeeze(valid_pos, [1])
## update next grammar mask
next_actions = layers.elementwise_mul(decoder.grammar_action(next_output),
layers.cast(
valid_pos,
dtype=next_inputs.action.dtype),
axis=0)
next_gmr_mask = layers.elementwise_mul(
decoder.grammar_mask(next_output),
layers.cast(valid_pos, dtype=next_inputs.gmr_mask.dtype),
axis=0)
## save result, while condition is True
new_gmr_stack_data, new_gmr_stack_pos, new_actions, new_gmr_mask = nn_utils.ifelse(
condition,
[gmr_stack_tmp.data, gmr_stack_tmp.pos, next_actions, next_gmr_mask], [
grammar_stack.data, grammar_stack.pos, next_inputs.action,
next_inputs.gmr_mask
])
layers.utils.map_structure(
layers.assign,
[new_gmr_stack_data, new_gmr_stack_pos, next_actions, new_gmr_mask], [
grammar_stack.data, grammar_stack.pos, next_inputs.action,
next_inputs.gmr_mask
])
layers.logical_or(finished,
layers.logical_and(condition,
layers.logical_not(valid_pos)),
out=finished)
def pop(cls, stack_data, mask=True, in_place=True):
"""pop data in stack_data
Args:
stack_data (StackData): (data, pos) with shape ([batch_size, stack_len], [batch_size, 1])
mask (bool): 是否 mask 空栈的返回值。默认为 True
in_place (bool): 默认为 True
Returns: (Variable1, Variable2)
Variable1: pop 得到的值
dtype=stack_data.data.dtype
shape=[-1]
Variable2: 对应位置的值是否合法。入参已经为空的栈，此处为 False。
dtype=bool
shape=[-1]
Raises: NULL
"""
data = stack_data.data
pos = stack_data.pos
# 只有非空的栈才能pop（才合法）
valid_pos = layers.logical_not(cls.empty(stack_data))
new_pos_delta = layers.cast(valid_pos, dtype=pos.dtype)
new_pos = layers.elementwise_sub(pos, new_pos_delta)
# shape = [batch_size]
output = nn_utils.batch_gather(data, new_pos)
# mask 空栈的返回值
if mask:
# shape = [batch_size, 1]
mask_tag = layers.cast(
new_pos_delta,
dtype=data.dtype) if data.dtype != pos.dtype else new_pos_delta
mask_tag = layers.squeeze(mask_tag, [1])
output = layers.elementwise_mul(output, mask_tag)
# 出栈后原位置置为0
updates = layers.zeros_like(output)
new_data = nn_utils.batch_scatter(data,
new_pos,
updates,
overwrite=True,
in_place=in_place)
if in_place:
layers.assign(new_pos, pos)
return output, valid_pos, stack_data
else:
return output, valid_pos, StackData(new_data, new_pos)
def forward(self, words, feats):
"""Forward network"""
batch_size, seq_len = words.shape
# get the mask and lengths of given batch
mask = words != self.pad_index
ext_words = words
# set the indices larger than num_embeddings to unk_index
if hasattr(self, 'pretrained'):
ext_mask = words >= self.word_embed.weight.shape[0]
ext_words = nn.mask_fill(words, ext_mask, self.unk_index)
# get outputs from embedding layers
word_embed = self.word_embed(ext_words)
if hasattr(self, 'pretrained'):
word_embed += self.pretrained(words)
feat_embed = self.feat_embed(feats)
word_embed, feat_embed = self.embed_dropout(word_embed, feat_embed)
# concatenate the word and feat representations
# embed.size = (batch, seq_len, n_embed * 2)
embed = layers.concat((word_embed, feat_embed), axis=-1)
if self.args.encoding_model == "lstm":
x, _ = self.lstm(embed, mask, self.pad_index)
x = self.lstm_dropout(x)
else:
_, x = self.transformer(words, word_emb=embed)
# apply MLPs to the BiLSTM output states
arc_h = self.mlp_arc_h(x)
arc_d = self.mlp_arc_d(x)
rel_h = self.mlp_rel_h(x)
rel_d = self.mlp_rel_d(x)
# get arc and rel scores from the bilinear attention
# [batch_size, seq_len, seq_len]
s_arc = self.arc_attn(arc_d, arc_h)
# [batch_size, seq_len, seq_len, n_rels]
s_rel = layers.transpose(self.rel_attn(rel_d, rel_h),
perm=(0, 2, 3, 1))
# set the scores that exceed the length of each sentence to -1e5
s_arc_mask = nn.unsqueeze(layers.logical_not(mask), 1)
s_arc = nn.mask_fill(s_arc, s_arc_mask, -1e5)
return s_arc, s_rel
def _check_finished(decoder, next_inputs, finished, outputs_array):
"""check finished instance by next_inputs.action, and
update finished tag and write END to outputs
Args:
decoder (TYPE): NULL
next_inputs (TYPE): NULL
finished (TYPE): NULL
outputs_array (TYPE): NULL
Returns: TODO
Raises: NULL
"""
act_stop = tensor.fill_constant_batch_size_like(
next_inputs.action,
shape=next_inputs.action.shape,
value=decoder._grammar.ACTION_STOP,
dtype='int64')
new_finished = layers.logical_and(
layers.equal(next_inputs.action, act_stop),
layers.logical_not(finished))
end_token_id = tensor.fill_constant_batch_size_like(
outputs_array.data,
shape=[-1],
value=decoder._grammar.END,
dtype=outputs_array.data.dtype)
out_data_tmp, out_pos_tmp = data_structure.Array.push(outputs_array,
end_token_id,
in_place=False)
new_data, new_pos = nn_utils.ifelse(
new_finished, [out_data_tmp, out_pos_tmp],
[outputs_array.data, outputs_array.pos])
layers.assign(new_data, outputs_array.data)
layers.assign(new_pos, outputs_array.pos)
layers.logical_or(finished, new_finished, out=finished)
def training_network(self, img, caption):
# build caption and mask
target = caption[:, 1:]
source = caption[:, :-1]
padding_filled = layers.fill_constant_batch_size_like(target, shape=[-1, decoder_config['sentence_length'] - 1],
dtype='int64', value=config.dc['padding_idx'])
mask = layers.equal(target, padding_filled)
mask = layers.cast(layers.logical_not(mask), 'float32')
scale_factor = layers.reduce_sum(mask)
mask.stop_gradient = True
scale_factor.stop_gradient = True
# mdl
decoder = Decoder(decoder_config['hidden_dim'], rnn_layer=1)
image_embed, global_image_feat = self._img2feature(img) # [batch, k+1, hidden], [batch, hidden]
# 这里要改，要么在rnn里面做embedding，要么在外面做！
seq_out = decoder.call(global_image_feat, image_embed, embedding_function, words=source)
loss = layers.squeeze(ImageCaptionModel.loss(target, seq_out), axes=[2])
loss = layers.elementwise_mul(loss, mask)
output_loss = layers.elementwise_div(layers.reduce_sum(loss), scale_factor, name='loss')
return output_loss
def cond_func(step_idx, selected_ids, selected_scores, gather_idx,
caches, trg_src_attn_bias):
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
return layers.logical_and(x=length_cond, y=finish_cond)
def infilling_decode(self):
if self.task_type == "dialog":
emb_num = 4
else:
emb_num = 3
input_shapes = [[-1, self.max_seq_len, 1]] * emb_num + \
[[-1, self.max_seq_len, self.max_seq_len]]
input_dtypes = ['int64'] * emb_num + ['float32']
input_lod_levels = [0] * emb_num + [0]
shapes = input_shapes + [[-1, self.max_seq_len, 1],
[-1, self.max_seq_len, 1], [-1, 1], [-1],
[-1, 1, self.max_seq_len], [-1, 1]]
dtypes = input_dtypes + [
'int64', 'int64', 'float32', 'int32', 'float32', 'int64'
]
lod_levels = input_lod_levels + [2, 2, 2, 0, 0, 0]
inputs = self.to_ternsor(shapes, dtypes, lod_levels)
pyreader = fluid.io.DataLoader.from_generator(feed_list=inputs,
capacity=50,
iterable=False)
emb_ids = {}
for key, value in zip(self.emb_keys, inputs[:emb_num]):
emb_ids[key] = value
input_mask = inputs[emb_num]
tgt_ids, tgt_pos, init_scores, parent_idx, tgt_input_mask, data_ids = inputs[
-6:]
ernie = ErnieModel(emb_ids=emb_ids,
input_mask=input_mask,
config=self.ernie_config,
use_fp16=self.use_fp16,
task_type=self.task_type,
decoding=True,
gather_idx=parent_idx)
max_len = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=self.max_dec_len,
force_cpu=True)
step_idx = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=0,
force_cpu=True)
pos_idx = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=1,
force_cpu=True)
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
ids = layers.array_write(layers.reshape(tgt_ids, (-1, 1)), step_idx)
pos_biases = layers.array_write(layers.reshape(tgt_pos, (-1, 1)),
step_idx)
scores = layers.array_write(init_scores, step_idx)
tgt_masks = layers.array_write(tgt_input_mask, step_idx)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
pos_bias = layers.array_read(array=pos_biases, i=step_idx)
pos_bias = layers.gather(input=pos_bias, index=parent_idx)
tmp_mask = layers.array_read(tgt_masks, i=step_idx)
def gen_batch_like(value,
dtype="int64",
shape=[-1, 1, 1],
is_scalar=True):
if is_scalar:
return layers.fill_constant_batch_size_like(
input=parent_idx,
value=value,
shape=shape,
dtype=dtype)
else:
return layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=parent_idx,
value=1,
shape=shape,
dtype=dtype),
y=value,
axis=0)
tmp_mask = layers.gather(input=tmp_mask, index=parent_idx)
append_0_mask = gen_batch_like(0.0, dtype=tmp_mask.dtype)
append_1_mask = gen_batch_like(1.0, dtype=tmp_mask.dtype)
tmp_mask = layers.concat([tmp_mask, append_1_mask], axis=2)
pre_mask = layers.concat([tmp_mask, append_0_mask], axis=2)
cur_mask = layers.concat([tmp_mask, append_1_mask], axis=2)
cur_ids = gen_batch_like(self.attn_id)
pre_pos = gen_batch_like(step_idx, is_scalar=False)
cur_pos = gen_batch_like(pos_idx, is_scalar=False)
if self.continuous_position:
pre_pos = pre_pos + pos_bias
cur_pos = cur_pos + pos_bias
dec_emb_ids = {
"word_embedding": layers.concat([pre_ids, cur_ids], axis=1),
"pos_embedding": layers.concat([pre_pos, cur_pos], axis=1)
}
if self.task_type == "dialog":
role_ids = gen_batch_like(0)
turn_ids = gen_batch_like(0)
dec_emb_ids["role_embedding"] = layers.concat(
[role_ids, role_ids], axis=1)
dec_emb_ids["turn_embedding"] = layers.concat(
[turn_ids, turn_ids], axis=1)
else:
sent_ids = gen_batch_like(self.tgt_type_id)
dec_emb_ids["sent_embedding"] = layers.concat(
[sent_ids, sent_ids], axis=1)
dec_mask = layers.concat([pre_mask, cur_mask], axis=1)
dec_out = ernie.encode(dec_emb_ids,
dec_mask,
parent_idx,
remove_query=True)
fc_out = self.cal_logit(dec_out[:, 1:, :], None)
topk_scores, topk_indices = layers.topk(
input=layers.softmax(fc_out), k=self.beam_size)
pre_lenpen = layers.pow(
(5.0 + layers.cast(step_idx, pre_scores.dtype)) / 6.0,
self.length_penalty)
cur_lenpen = layers.pow(
(5.0 + layers.cast(pos_idx, pre_scores.dtype)) / 6.0,
self.length_penalty)
accu_scores = layers.elementwise_add(x=layers.log(topk_scores),
y=pre_scores * pre_lenpen,
axis=0) / cur_lenpen
topk_indices = layers.lod_reset(topk_indices, pre_ids)
accu_scores = layers.lod_reset(accu_scores, pre_ids)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=self.beam_size,
end_id=self.eos_idx,
return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
layers.increment(x=pos_idx, value=1.0, in_place=True)
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.array_write(tmp_mask, i=step_idx, array=tgt_masks)
layers.array_write(pos_bias, i=step_idx, array=pos_biases)
layers.assign(gather_idx, parent_idx)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=self.beam_size, end_id=self.eos_idx)
graph_vars = {
"finished_ids": finished_ids,
"finished_scores": finished_scores,
"data_ids": data_ids
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def _run_paddle_logical_not(x):
x = cast_bool_if_necessary(x)
return logical_not(x)
def beam_search():
max_len = layers.fill_constant(shape=[1],
dtype=start_tokens.dtype,
value=max_out_len,
force_cpu=True)
step_idx = layers.fill_constant(shape=[1],
dtype=start_tokens.dtype,
value=0,
force_cpu=True)
cond = layers.less_than(x=step_idx,
y=max_len) # default force_cpu=True
while_op = layers.While(cond)
# array states will be stored for each step.
ids = layers.array_write(layers.reshape(start_tokens, (-1, 1)),
step_idx)
scores = layers.array_write(init_scores, step_idx)
# cell states will be overwrited at each step.
# caches contains states of history steps in decoder self-attention
# and static encoder output projections in encoder-decoder attention
# to reduce redundant computation.
caches = [
{
"k": # for self attention
layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, n_head, 0, d_key],
dtype=enc_output.dtype,
value=0),
"v": # for self attention
layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, n_head, 0, d_value],
dtype=enc_output.dtype,
value=0),
"static_k": # for encoder-decoder attention
layers.create_tensor(dtype=enc_output.dtype),
"static_v": # for encoder-decoder attention
layers.create_tensor(dtype=enc_output.dtype)
} for i in range(n_layer)
]
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
# Since beam_search_op dosen't enforce pre_ids' shape, we can do
# inplace reshape here which actually change the shape of pre_ids.
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
# gather cell states corresponding to selected parent
pre_src_attn_bias = layers.gather(trg_src_attn_bias,
index=parent_idx)
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_src_attn_bias, # cann't use lod tensor here
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype),
y=step_idx,
axis=0)
logits = wrap_decoder(trg_vocab_size,
max_in_len,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
prepostprocess_dropout,
attention_dropout,
relu_dropout,
preprocess_cmd,
postprocess_cmd,
weight_sharing,
dec_inputs=(pre_ids, pre_pos, None,
pre_src_attn_bias),
enc_output=enc_output,
caches=caches,
gather_idx=parent_idx,
bos_idx=bos_idx)
# intra-beam topK
topk_scores, topk_indices = layers.topk(
input=layers.softmax(logits), k=beam_size)
accu_scores = layers.elementwise_add(x=layers.log(topk_scores),
y=pre_scores,
axis=0)
# beam_search op uses lod to differentiate branches.
accu_scores = layers.lod_reset(accu_scores, pre_ids)
# topK reduction across beams, also contain special handle of
# end beams and end sentences(batch reduction)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
end_id=eos_idx,
return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
# cell states(caches) have been updated in wrap_decoder,
# only need to update beam search states here.
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.assign(gather_idx, parent_idx)
layers.assign(pre_src_attn_bias, trg_src_attn_bias)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=beam_size, end_id=eos_idx)
return finished_ids, finished_scores
def beam_search():
"""Beam search function"""
max_len = layers.fill_constant(shape=[1],
dtype=start_tokens.dtype,
value=self.max_out_len,
force_cpu=True)
min_len = layers.fill_constant(shape=[1],
dtype=start_tokens.dtype,
value=self.min_out_len)
neg_inf = layers.fill_constant(shape=[1],
dtype='float32',
value=-INF)
step_idx = layers.fill_constant(shape=[1],
dtype=start_tokens.dtype,
value=0,
force_cpu=True)
step_next_idx = layers.fill_constant(shape=[1],
dtype=start_tokens.dtype,
value=1,
force_cpu=True)
cond = layers.less_than(x=step_idx,
y=max_len) # default force_cpu=True
while_op = layers.While(cond)
# array states will be stored for each step.
ids = layers.array_write(layers.reshape(start_tokens, (-1, 1)),
step_idx)
scores = layers.array_write(init_scores, step_idx)
# cell states will be overwrited at each step.
# caches contains states of history steps in decoder self-attention
# and static encoder output projections in encoder-decoder attention
# to reduce redundant computation.
caches = [
{
"k": # for self attention
layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, self._n_head, 0, self._emb_size // self._n_head],
dtype=enc_words_output.dtype,
value=0),
"v": # for self attention
layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, self._n_head, 0, self._emb_size // self._n_head],
dtype=enc_words_output.dtype,
value=0),
"static_k_word": # for encoder-decoder attention
layers.create_tensor(dtype=enc_words_output.dtype),
"static_v_word": # for encoder-decoder attention
layers.create_tensor(dtype=enc_words_output.dtype),
"static_k_sent": # for encoder-decoder attention
layers.create_tensor(dtype=enc_sents_output.dtype),
"static_v_sent": # for encoder-decoder attention
layers.create_tensor(dtype=enc_sents_output.dtype)
} for i in range(self._dec_n_layer)
]
trigram_blocking = TrigramBlocking(start_tokens,
self.tokenizer,
use_fp16=self._use_fp16,
beam_size=self.beam_size)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
# Since beam_search_op dosen't enforce pre_ids' shape, we can do
# inplace reshape here which actually change the shape of pre_ids.
# pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
# gather cell states corresponding to selected parent
pre_src_words_attn_bias = layers.gather(
tgt_src_words_attn_bias, index=parent_idx)
pre_src_sents_attn_bias = layers.gather(
tgt_src_sents_attn_bias, index=parent_idx)
pre_graph_attn_bias = layers.gather(graph_attn_bias,
index=parent_idx)
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=
pre_src_sents_attn_bias, # cann't use lod tensor here
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype),
y=step_idx,
axis=0)
logits = self.decode(
dec_input=(pre_ids, pre_pos, None, pre_src_words_attn_bias,
pre_src_sents_attn_bias, pre_graph_attn_bias),
enc_words_output=enc_words_output,
enc_sents_output=enc_sents_output,
caches=caches,
gather_idx=parent_idx)
# prevent generating end token if length less than min_out_len
eos_index = layers.fill_constant(
shape=[layers.shape(logits)[0]],
dtype='int64',
value=self.eos_idx)
eos_index = fluid.one_hot(eos_index, depth=self.voc_size)
less_cond = layers.cast(layers.less_than(x=step_idx,
y=min_len),
dtype='float32')
less_val = layers.elementwise_mul(less_cond, neg_inf)
eos_val = layers.elementwise_mul(eos_index, less_val, axis=0)
revised_logits = layers.elementwise_add(logits,
eos_val,
axis=0)
# topK reduction across beams, also contain special handle of
# end beams and end sentences(batch reduction)
topk_scores, topk_indices = layers.topk(
input=layers.softmax(revised_logits), k=self.beam_size)
# Roll-Back previous-scores for length-penalty
# previous-scores has been length-penaltied, before this timestep length-penalty, need roll-back
# because of doing this, we need store the length-penaltied score in `scores`
# while calculating use the un-penaltied score
# -> safe for step_idx == 0 (initialization state), because previous-score == 0
pre_timestep_length_penalty = fluid.layers.pow(
((5.0 + fluid.layers.cast(step_idx, pre_scores.dtype)) /
6.0), self.len_penalty)
pre_scores_wo_len_penalty = fluid.layers.elementwise_mul(
pre_scores, pre_timestep_length_penalty)
# calc trigram-blocking delta scores for current alive sequence
if self.block_trigram:
trigram_blocking.update_seq(pre_ids, parent_idx)
trigram_blocking.expand_cand_seq(topk_indices)
fluid.layers.py_func(
func=trigram_blocking.blocking_forward,
x=[
trigram_blocking.cand_seq,
trigram_blocking.id2is_full_token
],
out=trigram_blocking.delta_score_out,
backward_func=None)
layers.Print(trigram_blocking.delta_score_out,
summarize=100,
message="trigram_blocking.delta_score_out")
pre_scores_wo_len_penalty = fluid.layers.elementwise_add(
x=trigram_blocking.delta_score_out,
y=pre_scores_wo_len_penalty,
axis=0)
# => [N, topk]
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores),
y=pre_scores_wo_len_penalty,
axis=0)
cur_timestep_length_penalty = layers.pow(
((5.0 + layers.cast(step_next_idx, accu_scores.dtype)) /
6.0), self.len_penalty)
curr_scores = layers.elementwise_div(
accu_scores, cur_timestep_length_penalty)
# beam_search op uses lod to differentiate branches.
curr_scores = layers.lod_reset(curr_scores, pre_ids)
topk_indices = layers.lod_reset(topk_indices, pre_ids)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=curr_scores,
beam_size=self.beam_size,
end_id=self.eos_idx,
return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
layers.increment(x=step_next_idx, value=1.0, in_place=True)
# cell states(caches) have been updated in wrap_decoder,
# only need to update beam search states here.
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.assign(gather_idx, parent_idx)
layers.assign(pre_src_words_attn_bias, tgt_src_words_attn_bias)
layers.assign(pre_src_sents_attn_bias, tgt_src_sents_attn_bias)
layers.assign(pre_graph_attn_bias, graph_attn_bias)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(
layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=self.beam_size, end_id=self.eos_idx)
return finished_ids, finished_scores
def decoder(self, init_state):
"""
implement decoder in inference mode
"""
# pd.Print(init_state)
# define counter variable in the decoding
array_len = pd.fill_constant(shape=[1],
dtype='int64',
value=self.max_length)
counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
static_count = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
# define tensor array to save content at each time step, and write initial id, score and state
state_h_array = pd.create_array('float32')
pd.array_write(self.h, array=state_h_array, i=counter)
state_c_array = pd.create_array('float32')
pd.array_write(self.c, array=state_c_array, i=counter)
src_indexes = fluid.layers.data(name='source_index',
shape=[1],
dtype='int64',
lod_level=1)
src_index_array = pd.create_array('int64')
pd.array_write(src_indexes, array=src_index_array, i=counter)
ids_array = pd.create_array('int64')
scores_array = pd.create_array('float32')
init_ids = fluid.layers.data(name="init_ids",
shape=[1],
dtype="int64",
lod_level=2)
init_scores = fluid.layers.data(name="init_scores",
shape=[1],
dtype="float32",
lod_level=2)
pd.array_write(init_ids, array=ids_array, i=counter)
pd.array_write(init_scores, array=scores_array, i=counter)
encoder_vec_array = pd.create_array('float32')
pd.array_write(self.encoder_vec,
array=encoder_vec_array,
i=static_count)
encoder_vec_full_array = pd.create_array('float32')
pd.array_write(self.encoder_vec_full,
array=encoder_vec_full_array,
i=static_count)
encoder_proj_array = pd.create_array('float32')
pd.array_write(self.encoder_proj,
array=encoder_proj_array,
i=static_count)
event_embedding_array = pd.create_array('float32')
pd.array_write(self.event_embedding,
array=event_embedding_array,
i=static_count)
# define conditional variable to stop loop
cond = pd.less_than(x=counter, y=array_len)
# define while_op
while_op = pd.While(cond=cond)
with while_op.block(): # define the computing of each step
# pd.Print(counter)
# obtain input at present step of decoder, including id chosen at previous step, corresponding score and state at previous step.
pre_ids = pd.array_read(array=ids_array, i=counter)
pre_h_state = pd.array_read(array=state_h_array, i=counter)
pre_c_state = pd.array_read(array=state_c_array, i=counter)
# pre_score = pd.array_read(array=scores_array, i=counter)
pre_score = pd.array_read(array=scores_array, i=static_count)
_encoder_input_ids = pd.array_read(array=src_index_array,
i=static_count)
event_embedding = pd.array_read(array=event_embedding_array,
i=static_count)
# print("pre_h_state", pre_h_state)
encoder_vec = pd.array_read(array=encoder_vec_array,
i=static_count)
encoder_vec_full = pd.array_read(array=encoder_vec_full_array,
i=static_count)
encoder_proj = pd.array_read(array=encoder_proj_array,
i=static_count)
# # update input state as state correspondent with id chosen at previous step
# pre_h_state_expanded = pd.sequence_expand(pre_h_state, pre_score)
# pre_c_state_expanded = pd.sequence_expand(pre_c_state, pre_score)
# computing logic of decoder under the same train mode, including input vector and computing unit of decoder
# compute predicting probability of normalized word
pre_ids_emb = pd.embedding(
input=pre_ids,
size=[self.target_dict_dim, self.embedding_dim],
dtype='float32',
param_attr=fluid.ParamAttr(name="trg_embedding"))
# pd.Print(pre_ids_emb)
att_context = self.simple_attention(encoder_vec, encoder_proj,
pre_h_state)
# print("att_context", att_context)
# print("pre_ids_emb", pre_ids_emb)
# pd.Print(att_context)
prob_c = fluid.layers.sequence_expand_as(pre_score, encoder_vec)
# pd.Print(prob_c)
current_score, current_h, current_c, this_prob_c = self.copy_decoder(
pre_ids_emb, encoder_vec, encoder_vec_full, encoder_proj,
_encoder_input_ids, pre_ids, prob_c, att_context, pre_h_state,
pre_c_state, event_embedding)
# decoder_inputs = fluid.layers.concat(
# input=[att_context, pre_ids_emb], axis=1)
# current_h, current_c = self.lstm_step(
# decoder_inputs, pre_h_state, pre_c_state, self.decoder_size)
# # compute predicting probability of nomarlized word
# current_score = fluid.layers.fc(input=current_h,
# size=self.target_dict_dim,
# act='softmax',
# param_attr=fluid.ParamAttr(name="out_softmax_w"),
# bias_attr=fluid.ParamAttr(name="out_softmax_b"))
# # current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb],
# # size=decoder_size,
# # act='tanh')
# current_state_with_lod = pd.lod_reset(x=current_h, y=pre_score)
# current_score = pd.fc(input=current_state_with_lod,
# size=self.target_dict_dim,
# act='softmax',
# param_attr=fluid.ParamAttr(name="out_softmax_w"),
# bias_attr=fluid.ParamAttr(name="out_softmax_b"))
# print(current_score)
topk_scores, topk_indices = pd.topk(current_score,
k=self.beam_size)
# pd.Print(topk_indices)
# pd.Print(topk_scores)
selected_ids, selected_scores = topk_indices, topk_scores
# # compute accumulated score and perform beam search
# accu_scores = pd.elementwise_add(
# x=pd.log(topk_scores), y=pd.reshape(pre_score, shape=[-1]), axis=0)
# selected_ids, selected_scores = pd.beam_search(
# pre_ids,
# pre_score,
# topk_indices,
# accu_scores,
# self.beam_size,
# # end_id=self.end_id,
# end_id=999999,
# level=0)
# pd.Print(selected_ids)
# pd.Print(selected_scores)
pd.increment(x=counter, value=1, in_place=True)
# write search result and corresponding hidden layer into tensor array
pd.array_write(current_h, array=state_h_array, i=counter)
pd.array_write(current_c, array=state_c_array, i=counter)
pd.array_write(selected_ids, array=ids_array, i=counter)
pd.array_write(selected_scores, array=scores_array, i=counter)
# pd.Print(selected_ids)
# pd.Print(selected_scores)
# update condition to stop loop
length_cond = pd.less_than(x=counter, y=array_len)
finish_cond = pd.logical_not(pd.is_empty(x=selected_ids))
pd.logical_and(x=length_cond, y=finish_cond, out=cond)
# pd.Print(array_len)
# translation_ids, translation_scores = pd.beam_search_decode(
# ids=ids_array, scores=scores_array, beam_size=self.beam_size, end_id=self.end_id)
# pd.Print(translation_ids)
translation_ids, translation_ids_index = pd.tensor_array_to_tensor(
ids_array, axis=1)
translation_scores, translation_scores_index = pd.tensor_array_to_tensor(
scores_array, axis=1)
return translation_ids, translation_scores
def _do_beam_search(trg_vocab_size, max_in_len, n_layer, n_head, d_key,
d_value, d_model, d_inner_hid, prepostprocess_dropout,
attention_dropout, relu_dropout, preprocess_cmd,
postprocess_cmd, weight_sharing, beam_size, max_len,
bos_idx, eos_idx, ids, scores, parent_idx,
trg_src_attn_bias, caches, enc_output, step_idx):
"""
do beam search
"""
cond = layers.less_than(x=step_idx, y=max_len) # default force_cpu=True
while_op = layers.While(cond)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
# Since beam_search_op dosen't enforce pre_ids' shape, we can do
# inplace reshape here which actually change the shape of pre_ids.
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
# gather cell states corresponding to selected parent
pre_src_attn_bias = layers.gather(trg_src_attn_bias, index=parent_idx)
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_src_attn_bias, # cann't use lod tensor here
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype),
y=step_idx,
axis=0)
logits = wrap_decoder(trg_vocab_size,
max_in_len,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
prepostprocess_dropout,
attention_dropout,
relu_dropout,
preprocess_cmd,
postprocess_cmd,
weight_sharing,
dec_inputs=(pre_ids, pre_pos, None,
pre_src_attn_bias),
enc_output=enc_output,
caches=caches,
gather_idx=parent_idx,
bos_idx=bos_idx)
# intra-beam topK
topk_scores, topk_indices = layers.topk(input=layers.softmax(logits),
k=beam_size)
accu_scores = layers.elementwise_add(x=layers.log(topk_scores),
y=pre_scores,
axis=0)
# beam_search op uses lod to differentiate branches.
accu_scores = layers.lod_reset(accu_scores, pre_ids)
# topK reduction across beams, also contain special handle of
# end beams and end sentences(batch reduction)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
end_id=eos_idx,
return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
# cell states(caches) have been updated in wrap_decoder,
# only need to update beam search states here.
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.assign(gather_idx, parent_idx)
layers.assign(pre_src_attn_bias, trg_src_attn_bias)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
def inference(self, model, inputs, outputs):
"""
Run inference.
Args:
inputs(dict): Its key is input name(str) and its value is a Variable.
model(object): A generate model. Need to implement `_generation_network` and `_calc_logits`.
Returns:
dict(str:Variable): Its key is output name(str) and its value is a Variable.
"""
# prepare while loop
max_len = layers.fill_constant(
shape=[1], dtype="int64", value=self.max_dec_len, force_cpu=True)
min_len = layers.fill_constant(
shape=[1], dtype="int64", value=self.min_dec_len, force_cpu=True)
step_idx = layers.fill_constant(
shape=[1], dtype="int64", value=0, force_cpu=True)
ids = layers.array_write(layers.reshape(inputs["tgt_ids"], (-1, 1)), step_idx)
pos_biases = layers.array_write(layers.reshape(inputs["tgt_pos"], (-1, 1)), step_idx)
scores = layers.array_write(inputs["init_score"], step_idx)
tgt_generation_mask = layers.array_write(inputs["tgt_generation_mask"], step_idx)
parent_idx = inputs["parent_idx"]
if self.decoding_strategy == "beam_search":
beam_size = self.beam_size
else:
beam_size = 1
eos_penalty = np.zeros(self.vocab_size, dtype="float32")
eos_penalty[self.eos_id] = -1e9
eos_penalty = layers.assign(eos_penalty)
token_penalty = np.zeros(self.vocab_size, dtype="float32")
token_penalty[self.unk_id] = -1e9
if self.mask_id >= 0:
token_penalty[self.mask_id] = -1e9
token_penalty = layers.assign(token_penalty)
# start while loop
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
pos_bias = layers.array_read(array=pos_biases, i=step_idx)
pos_bias = layers.gather(input=pos_bias, index=parent_idx)
tmp_tgt_generation_mask = layers.array_read(tgt_generation_mask, i=step_idx)
dtype = tmp_tgt_generation_mask.dtype
append_mask = layers.fill_constant_batch_size_like(
input=pre_ids,
value=1.0,
shape=[-1, 1, 1],
dtype=dtype)
tmp_tgt_generation_mask = layers.concat([tmp_tgt_generation_mask, append_mask], axis=2)
pre_mask = tmp_tgt_generation_mask = layers.gather(input=tmp_tgt_generation_mask, index=parent_idx)
pre_sent = layers.fill_constant_batch_size_like(
input=pre_mask,
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype)
if self.continuous_position:
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_mask,
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype), y=step_idx, axis=0) + pos_bias
else:
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_mask,
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype), y=step_idx, axis=0)
if self.use_role:
pre_role = layers.fill_constant_batch_size_like(
input=pre_mask,
value=0,
shape=[-1, 1, 1],
dtype=pre_ids.dtype)
else:
pre_role = None
dec_out, _ = model._generation_network(
token_ids=pre_ids,
type_ids=pre_sent,
pos_ids=pre_pos,
role_ids=pre_role,
generation_mask=tmp_tgt_generation_mask,
gather_idx=parent_idx)
logits = model._calc_logits(dec_out)
# ignore unk and mask token
if self.ignore_unk:
logits = layers.elementwise_add(logits, token_penalty, axis=1)
# min dec length
min_len_cond = layers.less_than(x=step_idx, y=min_len)
def min_len_penalty():
"""Plus minimum length penalty."""
return layers.elementwise_add(logits, eos_penalty, axis=1)
def no_penalty():
"""No penalty."""
return logits
logits = layers.case([(min_len_cond, min_len_penalty)], default=no_penalty)
# get probs
probs = layers.softmax(logits / self.temperature)
if self.decoding_strategy == "beam_search":
topk_scores, topk_indices = layers.topk(
input=probs, k=beam_size)
else:
if self.decoding_strategy.startswith("sampling"):
sampling_ids = layers.sampling_id(probs, dtype="int")
elif self.decoding_strategy.startswith("topk_sampling"):
topk_probs, _ = layers.topk(input=probs, k=self.topk)
ge_cond = layers.cast(
layers.greater_equal(
probs,
layers.unsqueeze(topk_probs[:, -1], [1])),
"float32")
old_probs = probs
probs = probs * ge_cond / layers.reduce_sum(topk_probs, dim=-1, keep_dim=True)
sampling_ids = layers.sampling_id(probs, dtype="int")
probs = old_probs
else:
raise ValueError(self.decoding_strategy)
sampling_scores = layers.one_hot(
layers.unsqueeze(sampling_ids, [1]), probs.shape[1]
)
sampling_scores = sampling_scores * probs - (1 - sampling_scores) * 1e3
topk_scores, topk_indices = layers.topk(
input=sampling_scores, k=1)
pre_len = layers.cast(step_idx, "float32")
layers.increment(x=step_idx, value=1.0, in_place=True)
cur_len = layers.cast(step_idx, "float32")
# update scores
if self.length_average:
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores), y=pre_scores * pre_len, axis=0) / cur_len
elif self.length_penalty > 0:
pre_lp = layers.pow((5 + pre_len) / 6, self.length_penalty)
cur_lp = layers.pow((5 + cur_len) / 6, self.length_penalty)
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores), y=pre_scores * pre_lp, axis=0) / cur_lp
else:
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores), y=pre_scores, axis=0)
topk_indices = layers.lod_reset(topk_indices, pre_ids)
accu_scores = layers.lod_reset(accu_scores, pre_ids)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
end_id=self.eos_id,
return_parent_idx=True)
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.array_write(pre_mask, i=step_idx, array=tgt_generation_mask)
layers.array_write(pos_bias, i=step_idx, array=pos_biases)
layers.assign(gather_idx, parent_idx)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=beam_size, end_id=self.eos_id)
predictions = {
"finished_ids": finished_ids,
"finished_scores": finished_scores,
"token_ids": inputs["token_ids"],
"data_id": inputs["data_id"]
}
return predictions
def decode_with_grammar(decoder, inits, decode_vocab, max_step_num, **kwargs):
"""A modification of paddle.fluid.layers.dynamic_decode(...).
Dynamic decoding performs :code:`decoder.step()` repeatedly until the returned
Tensor indicating finished status contains all True values or the number of
decoding step reachs to :attr:`max_step_num`.
:code:`decoder.initialize()` would be called once before the decoding loop.
If the `decoder` has implemented `finalize` method, :code:`decoder.finalize()`
would be called once after the decoding loop.
Args:
decoder(Decoder): An instance of `Decoder`.
inits(tuple): Argument passed to `decoder.initialize`.
decode_vocab(DecoderDynamicVocab): namedtuple(table table_len column column_len value value_len)
max_step_num(int): The maximum number of steps.
**kwargs: Additional keyword arguments. Arguments passed to `decoder.step`.
Returns:
tuple: A tuple( :code:`(final_outputs, final_states)` ) including the final \
outputs and states, both are Tensor or nested structure of Tensor. \
`final_outputs` has the same structure and data types as \
:code:`decoder.output_dtype` , and each Tenser in `final_outputs` \
is the stacked of all decoding steps' outputs, which might be revised \
by :code:`decoder.finalize` . `final_states` is the counterpart \
at last time step of initial states returned by :code:`decoder.initialize` , \
thus has the same structure with it and has tensors with same shapes \
and data types.
"""
step_cnt = tensor.fill_constant(shape=[1], dtype="int64", value=1)
max_step_num_tensor = tensor.fill_constant(shape=[1],
dtype="int64",
value=max_step_num - 2)
# shape = [batch_size, beam_size, ...]
initial_inputs, initial_states, initial_finished = decoder.initialize(
inits, decode_vocab)
global_inputs, global_states, global_finished = (initial_inputs,
initial_states,
initial_finished)
inputs = initial_inputs
states = initial_states
# 保存输出结果
outputs_arr_data = tensor.fill_constant_batch_size_like(
inputs.input,
shape=[-1, decoder.beam_size, max_step_num],
dtype=decoder.output_dtype.predicted_ids,
value=0)
outputs_arr_pos = tensor.fill_constant_batch_size_like(
inputs.input, shape=[-1, decoder.beam_size, 1], dtype='int64', value=0)
outputs_array = data_structure.ArrayData(
decoder.merge_batch_beams(outputs_arr_data),
decoder.merge_batch_beams(outputs_arr_pos))
sequence_lengths = tensor.cast(tensor.zeros_like(initial_finished),
"int64")
# 按语法解码的相关约束数据结构
grammar_stack_dat = tensor.fill_constant_batch_size_like(
inputs.input,
shape=[-1, decoder.beam_size, max_step_num * STACK_EXPAND_TIMES],
dtype='int64',
value=0)
grammar_stack_pos = tensor.fill_constant_batch_size_like(
inputs.input, shape=[-1, decoder.beam_size, 1], dtype='int64', value=0)
grammar_stack = data_structure.StackData(
decoder.merge_batch_beams(grammar_stack_dat),
decoder.merge_batch_beams(grammar_stack_pos))
############ 循环解码，直到全部为 finish 状态 ############
# finish 的判断：通过 global_finished/next_finished && max_step_num 判断
cond = layers.logical_not((layers.reduce_all(initial_finished)))
while_op = layers.While(cond)
with while_op.block():
# step_outputs --> OutputWrapper
# next_states --> StateWrapper
# next_inputs --> DecoderInputsWrapper
step_outputs, next_states, next_inputs = decoder.step(
inputs, states, **kwargs)
predicted_ids = step_outputs.predicted_ids
_save_predict_output(outputs_array, predicted_ids,
next_states.finished)
pred_gmr_type = decoder.grammar_type(predicted_ids)
cond_type_leaf = layers.equal(pred_gmr_type, decoder.GMR_TYPE.LEAF)
cond_type_midd = layers.equal(pred_gmr_type, decoder.GMR_TYPE.MID)
_process_type_leaf(cond_type_leaf, decoder, grammar_stack, next_inputs,
next_states.finished)
_process_type_midd(cond_type_midd, decoder, grammar_stack, next_inputs,
predicted_ids)
##next_sequence_lengths = layers.elementwise_add(sequence_lengths,
## tensor.cast(layers.logical_not(global_finished), sequence_lengths.dtype))
_check_finished(decoder, next_inputs, next_states.finished,
outputs_array)
layers.utils.map_structure(tensor.assign, next_inputs, global_inputs)
layers.utils.map_structure(tensor.assign, next_states, global_states)
tensor.assign(next_states.finished, global_finished)
##tensor.assign(next_sequence_lengths, sequence_lengths)
# 更新循环条件
layers.increment(x=step_cnt, value=1.0, in_place=True)
layers.logical_and(
layers.logical_not(layers.reduce_all(next_states.finished)),
layers.less_equal(step_cnt, max_step_num_tensor), cond)
final_outputs = outputs_array.data
final_states = global_states
final_outputs, final_states = decoder.finalize(final_outputs,
global_states,
sequence_lengths)
return final_outputs, final_states
def gru_attention_infer(self, decoder_boot, max_length, char_num,
word_vector_dim, encoded_vector, encoded_proj,
decoder_size):
init_state = decoder_boot
beam_size = 1
array_len = layers.fill_constant(
shape=[1], dtype='int64', value=max_length)
counter = layers.zeros(shape=[1], dtype='int64', force_cpu=True)
# fill the first element with init_state
state_array = layers.create_array('float32')
layers.array_write(init_state, array=state_array, i=counter)
# ids, scores as memory
ids_array = layers.create_array('int64')
scores_array = layers.create_array('float32')
rois_shape = layers.shape(init_state)
batch_size = layers.slice(
rois_shape, axes=[0], starts=[0], ends=[1]) + 1
lod_level = layers.range(
start=0, end=batch_size, step=1, dtype=batch_size.dtype)
init_ids = layers.fill_constant_batch_size_like(
input=init_state, shape=[-1, 1], value=0, dtype='int64')
init_ids = layers.lod_reset(init_ids, lod_level)
init_ids = layers.lod_append(init_ids, lod_level)
init_scores = layers.fill_constant_batch_size_like(
input=init_state, shape=[-1, 1], value=1, dtype='float32')
init_scores = layers.lod_reset(init_scores, init_ids)
layers.array_write(init_ids, array=ids_array, i=counter)
layers.array_write(init_scores, array=scores_array, i=counter)
full_ids = fluid.layers.fill_constant_batch_size_like(
input=init_state, shape=[-1, 1], dtype='int64', value=1)
cond = layers.less_than(x=counter, y=array_len)
while_op = layers.While(cond=cond)
with while_op.block():
pre_ids = layers.array_read(array=ids_array, i=counter)
pre_state = layers.array_read(array=state_array, i=counter)
pre_score = layers.array_read(array=scores_array, i=counter)
pre_ids_emb = layers.embedding(
input=pre_ids,
size=[char_num, word_vector_dim],
dtype='float32')
context = self.simple_attention(encoded_vector, encoded_proj,
pre_state, decoder_size)
# expand the recursive_sequence_lengths of pre_state
# to be the same with pre_score
pre_state_expanded = layers.sequence_expand(pre_state, pre_score)
context_expanded = layers.sequence_expand(context, pre_score)
fc_1 = layers.fc(input=context_expanded,
size=decoder_size * 3,
bias_attr=False,
name="rnn_fc1")
fc_2 = layers.fc(input=pre_ids_emb,
size=decoder_size * 3,
bias_attr=False,
name="rnn_fc2")
decoder_inputs = fc_1 + fc_2
current_state, _, _ = layers.gru_unit(
input=decoder_inputs,
hidden=pre_state_expanded,
size=decoder_size * 3)
current_state_with_lod = layers.lod_reset(
x=current_state, y=pre_score)
# use score to do beam search
current_score = layers.fc(input=current_state_with_lod,
size=char_num,
bias_attr=True,
act='softmax',
name="rnn_out_fc")
topk_scores, topk_indices = layers.topk(current_score, k=beam_size)
new_ids = fluid.layers.concat([full_ids, topk_indices], axis=1)
fluid.layers.assign(new_ids, full_ids)
layers.increment(x=counter, value=1, in_place=True)
# update the memories
layers.array_write(current_state, array=state_array, i=counter)
layers.array_write(topk_indices, array=ids_array, i=counter)
layers.array_write(topk_scores, array=scores_array, i=counter)
# update the break condition:
# up to the max length or all candidates of
# source sentences have ended.
length_cond = layers.less_than(x=counter, y=array_len)
finish_cond = layers.logical_not(layers.is_empty(x=topk_indices))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
return full_ids
def beam_search():
max_len = layers.fill_constant(
shape=[1], dtype=start_tokens.dtype, value=max_out_len)
step_idx = layers.fill_constant(
shape=[1], dtype=start_tokens.dtype, value=0)
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
# array states will be stored for each step.
ids = layers.array_write(start_tokens, step_idx)
scores = layers.array_write(init_scores, step_idx)
# cell states will be overwrited at each step.
# caches contains states of history steps to reduce redundant
# computation in decoder.
caches = [{
"k": layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, 0, d_model],
dtype=enc_output.dtype,
value=0),
"v": layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, 0, d_model],
dtype=enc_output.dtype,
value=0)
} for i in range(n_layer)]
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_scores = layers.array_read(array=scores, i=step_idx)
# sequence_expand can gather sequences according to lod thus can be
# used in beam search to sift states corresponding to selected ids.
pre_src_attn_bias = layers.sequence_expand(
x=trg_src_attn_bias, y=pre_scores)
pre_enc_output = layers.sequence_expand(x=enc_output, y=pre_scores)
pre_caches = [{
"k": layers.sequence_expand(
x=cache["k"], y=pre_scores),
"v": layers.sequence_expand(
x=cache["v"], y=pre_scores),
} for cache in caches]
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_enc_output, # cann't use pre_ids here since it has lod
value=1,
shape=[-1, 1],
dtype=pre_ids.dtype),
y=layers.increment(
x=step_idx, value=1.0, in_place=False),
axis=0)
logits = wrap_decoder(
trg_vocab_size,
max_in_len,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate,
weight_sharing,
dec_inputs=(
pre_ids, pre_pos, None, pre_src_attn_bias, trg_data_shape,
slf_attn_pre_softmax_shape, slf_attn_post_softmax_shape,
src_attn_pre_softmax_shape, src_attn_post_softmax_shape),
enc_output=pre_enc_output,
caches=pre_caches)
topk_scores, topk_indices = layers.topk(
input=layers.softmax(logits), k=beam_size)
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores),
y=layers.reshape(
pre_scores, shape=[-1]),
axis=0)
# beam_search op uses lod to distinguish branches.
topk_indices = layers.lod_reset(topk_indices, pre_ids)
selected_ids, selected_scores = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
end_id=eos_idx)
layers.increment(x=step_idx, value=1.0, in_place=True)
# update states
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.assign(pre_src_attn_bias, trg_src_attn_bias)
layers.assign(pre_enc_output, enc_output)
for i in range(n_layer):
layers.assign(pre_caches[i]["k"], caches[i]["k"])
layers.assign(pre_caches[i]["v"], caches[i]["v"])
layers.assign(
layers.elementwise_add(
x=slf_attn_pre_softmax_shape,
y=attn_pre_softmax_shape_delta),
slf_attn_pre_softmax_shape)
layers.assign(
layers.elementwise_add(
x=slf_attn_post_softmax_shape,
y=attn_post_softmax_shape_delta),
slf_attn_post_softmax_shape)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=beam_size, end_id=eos_idx)
return finished_ids, finished_scores
def forward(self, x, y):
# x,y误差一帧
u1 = zeros_like(x)
u2 = zeros_like(x)
l_t = self.l * self.t
taut = self.a / self.t
grad2_x = self.conv_img_grad(y)
# grad2_x[:, :, :, 0] = 0.5 * (x[:, :, :, 1] - x[:, :, :, 0])
# grad2_x[:, :, :, -1] = 0.5 * (x[:, :, :, -1] - x[:, :, :, -2])
grad2_y = self.conv_img_grad2(y)
# grad2_y[:, :, 0, :] = 0.5 * (x[:, :, 1, :] - x[:, :, 0, :])
# grad2_y[:, :, -1, :] = 0.5 * (x[:, :, -1, :] - x[:, :, -2, :])
p11 = zeros_like(x)
p12 = zeros_like(x)
p21 = zeros_like(x)
p22 = zeros_like(x)
gsqx = grad2_x**2
gsqy = grad2_y**2
grad = gsqx + gsqy + 1e-12
rho_c = y - grad2_x * u1 - grad2_y * u2 - x
for i in range(self.n_iter):
rho = rho_c + grad2_x * u1 + grad2_y * u2 + 1e-12
v1 = zeros_like(x)
v2 = zeros_like(x)
mask1 = rho < -l_t * grad
mask2 = rho > l_t * grad
mask3 = logical_and(logical_not(logical_or(mask1, mask2)),
(grad > 1e-12))
mask1 = cast(mask1, dtype='float32')
mask2 = cast(mask2, dtype='float32')
mask3 = cast(mask3, dtype='float32')
mask1.stop_gradient = True
mask2.stop_gradient = True
mask3.stop_gradient = True
# v1 = v1 + l_t * grad2_x * mask1 - l_t * grad2_x * mask2 - (rho / grad) * grad2_x * mask3
# v2 = v2 + l_t * grad2_y * mask1 - l_t * grad2_y * mask2 - (rho / grad) * grad2_y * mask3
v1 = elementwise_add(
u1,
elementwise_add(
elementwise_mul(l_t * grad2_x, mask1),
elementwise_add(
elementwise_mul(-l_t * grad2_x, mask2),
elementwise_mul(-elementwise_div(rho, grad),
elementwise_mul(grad2_x, mask3)))))
v2 = elementwise_add(
u2,
elementwise_add(
elementwise_mul(l_t * grad2_y, mask1),
elementwise_add(
elementwise_mul(-l_t * grad2_y, mask2),
elementwise_mul(-elementwise_div(rho, grad),
elementwise_mul(grad2_y, mask3)))))
del rho
del mask1
del mask2
del mask3
v1 += u1
v2 += u2
u1 = v1 + self.t * self.divergence(p11, p12)
u2 = v2 + self.t * self.divergence(p21, p22)
del v1
del v2
u1 = u1
u2 = u2
u1x, u1y = self.forward_grad(u1)
u2x, u2y = self.forward_grad(u2)
p11 = (p11 + taut * u1x) / (1. +
taut * sqrt(u1x**2 + u1y**2 + 1e-12))
p12 = (p12 + taut * u1y) / (1. +
taut * sqrt(u1x**2 + u1y**2 + 1e-12))
p21 = (p21 + taut * u2x) / (1. +
taut * sqrt(u2x**2 + u2y**2 + 1e-12))
p22 = (p22 + taut * u2y) / (1. +
taut * sqrt(u2x**2 + u2y**2 + 1e-12))
del u1x
del u1y
del u2x
del u2y
return u1, u2
def decoder_decode(context, is_sparse):
init_state = context
array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
# fill the first element with init_state
state_array = pd.create_array('float32')
pd.array_write(init_state, array=state_array, i=counter)
# ids, scores as memory
ids_array = pd.create_array('int64')
scores_array = pd.create_array('float32')
init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2)
init_scores = pd.data(
name="init_scores", shape=[1], dtype="float32", lod_level=2)
pd.array_write(init_ids, array=ids_array, i=counter)
pd.array_write(init_scores, array=scores_array, i=counter)
cond = pd.less_than(x=counter, y=array_len)
while_op = pd.While(cond=cond)
with while_op.block():
pre_ids = pd.array_read(array=ids_array, i=counter)
pre_state = pd.array_read(array=state_array, i=counter)
pre_score = pd.array_read(array=scores_array, i=counter)
# expand the recursive_sequence_lengths of pre_state to be the same with pre_score
pre_state_expanded = pd.sequence_expand(pre_state, pre_score)
pre_ids_emb = pd.embedding(
input=pre_ids,
size=[dict_size, word_dim],
dtype='float32',
is_sparse=is_sparse)
# use rnn unit to update rnn
current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb],
size=decoder_size,
act='tanh')
current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score)
# use score to do beam search
current_score = pd.fc(input=current_state_with_lod,
size=target_dict_dim,
act='softmax')
topk_scores, topk_indices = pd.topk(current_score, k=beam_size)
# calculate accumulated scores after topk to reduce computation cost
accu_scores = pd.elementwise_add(
x=pd.log(topk_scores), y=pd.reshape(
pre_score, shape=[-1]), axis=0)
selected_ids, selected_scores = pd.beam_search(
pre_ids,
pre_score,
topk_indices,
accu_scores,
beam_size,
end_id=10,
level=0)
pd.increment(x=counter, value=1, in_place=True)
# update the memories
pd.array_write(current_state, array=state_array, i=counter)
pd.array_write(selected_ids, array=ids_array, i=counter)
pd.array_write(selected_scores, array=scores_array, i=counter)
# update the break condition: up to the max length or all candidates of
# source sentences have ended.
length_cond = pd.less_than(x=counter, y=array_len)
finish_cond = pd.logical_not(pd.is_empty(x=selected_ids))
pd.logical_and(x=length_cond, y=finish_cond, out=cond)
translation_ids, translation_scores = pd.beam_search_decode(
ids=ids_array, scores=scores_array, beam_size=beam_size, end_id=10)
# return init_ids, init_scores
return translation_ids, translation_scores
def fast_decode(self):
"""create model for inference"""
if self.task_type == "dialog":
emb_num = 4
else:
emb_num = 3
input_shapes = [[-1, self.max_seq_len, 1]] * emb_num + \
[[-1, self.max_seq_len, self.max_seq_len]]
input_dtypes = ['int64'] * emb_num + ['float32']
input_lod_levels = [0] * emb_num + [0]
shapes = input_shapes + [[-1, 1, 1], [-1, 1, 1], [-1, 1], [-1],
[-1, 1, self.max_seq_len], [-1, 1]]
dtypes = input_dtypes + [
'int64', 'int64', 'float32', 'int32', 'float32', 'int64'
]
lod_levels = input_lod_levels + [2, 2, 2, 0, 0, 0]
inputs = self.to_tensor(shapes, dtypes, lod_levels)
pyreader = fluid.io.DataLoader.from_generator(feed_list=inputs,
capacity=70,
iterable=False)
emb_ids = {}
for key, value in zip(self.emb_keys, inputs[:emb_num]):
emb_ids[key] = value
input_mask = inputs[emb_num]
tgt_ids, tgt_pos, init_scores, parent_idx, tgt_input_mask, data_ids = inputs[
-6:]
unimo = UNIMOModel(emb_ids=emb_ids,
input_mask=input_mask,
config=self.gene_config,
task_type=self.task_type,
decoding=True,
gather_idx=parent_idx)
max_len = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=self.max_out_len,
force_cpu=True)
min_len = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=self.min_out_len,
force_cpu=True)
neg_inf = layers.fill_constant(shape=[1], dtype='float32', value=-1e18)
step_idx = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=0,
force_cpu=True)
step_next_idx = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=1,
force_cpu=True)
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
ids = layers.array_write(layers.reshape(tgt_ids, (-1, 1)), step_idx)
pos_biases = layers.array_write(tgt_pos, step_idx)
scores = layers.array_write(init_scores, step_idx)
tgt_masks = layers.array_write(tgt_input_mask, step_idx)
trigram_blocking = TrigramBlocking(tgt_ids,
self.tokenizer,
beam_size=self.beam_size)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
pos_bias = layers.array_read(array=pos_biases, i=step_idx)
pos_bias = layers.gather(input=pos_bias, index=parent_idx)
def gen_batch_like(value,
dtype="int64",
shape=[-1, 1, 1],
is_scalar=True):
"""generate batch"""
if is_scalar:
return layers.fill_constant_batch_size_like(
input=parent_idx,
value=value,
shape=shape,
dtype=dtype)
else:
return layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=parent_idx,
value=1,
shape=shape,
dtype=dtype),
y=value,
axis=0)
tmp_mask = layers.array_read(tgt_masks, i=step_idx)
tmp_mask = layers.gather(input=tmp_mask, index=parent_idx)
append_1_mask = gen_batch_like(1.0, dtype=tmp_mask.dtype)
pre_mask = layers.concat([tmp_mask, append_1_mask], axis=2)
pre_pos = gen_batch_like(step_idx, is_scalar=False)
pre_pos = pre_pos + pos_bias ####################### pos start from 2
pre_sent = gen_batch_like(self.tgt_type_id, dtype=pre_ids.dtype)
dec_emb_ids = {"word_embedding": pre_ids, "pos_embedding": pre_pos}
if self.task_type == "dialog":
role_ids = gen_batch_like(0)
turn_ids = gen_batch_like(0)
dec_emb_ids["role_embedding"] = role_ids
dec_emb_ids["turn_embedding"] = turn_ids
else:
dec_emb_ids["sent_embedding"] = pre_sent
dec_out = unimo.encode(emb_ids=dec_emb_ids,
input_mask=pre_mask,
gather_idx=parent_idx)
fc_out = self.cal_logit(dec_out, None)
# prevent generating end token if length less than min_out_len
eos_index = layers.fill_constant(shape=[layers.shape(fc_out)[0]],
dtype='int64',
value=self.eos_id)
eos_index = fluid.one_hot(eos_index, depth=self.vocab_size)
less_cond = layers.cast(layers.less_than(x=step_idx, y=min_len),
dtype='float32')
less_val = layers.elementwise_mul(less_cond, neg_inf)
eos_val = layers.elementwise_mul(eos_index, less_val, axis=0)
revised_logits = layers.elementwise_add(fc_out, eos_val, axis=0)
# topK reduction across beams, also contain special handle of
# end beams and end sentences(batch reduction)
topk_scores, topk_indices = layers.topk(
input=layers.softmax(revised_logits), k=self.beam_size)
# Roll-Back previous-scores for length-penalty
# previous-scores has been length-penaltied, before this timestep length-penalty, need roll-back
# because of doing this, we need store the length-penaltied score in `scores`
# while calculating use the un-penaltied score
# -> safe for step_idx == 0 (initialization state), because previous-score == 0
pre_timestep_length_penalty = fluid.layers.pow(
((5.0 + fluid.layers.cast(step_idx, pre_scores.dtype)) / 6.0),
self.length_penalty)
pre_scores_wo_len_penalty = fluid.layers.elementwise_mul(
pre_scores, pre_timestep_length_penalty)
# calc trigram-blocking delta scores for current alive sequence
if self.block_trigram:
trigram_blocking.update_seq(pre_ids, parent_idx)
trigram_blocking.expand_cand_seq(topk_indices)
fluid.layers.py_func(func=trigram_blocking.blocking_forward,
x=[
trigram_blocking.cand_seq,
trigram_blocking.id2is_full_token
],
out=trigram_blocking.delta_score_out,
backward_func=None)
pre_scores_wo_len_penalty = fluid.layers.elementwise_add(
x=trigram_blocking.delta_score_out,
y=pre_scores_wo_len_penalty,
axis=0)
# => [N, topk]
accu_scores = layers.elementwise_add(x=layers.log(topk_scores),
y=pre_scores_wo_len_penalty,
axis=0)
cur_timestep_length_penalty = layers.pow(
((5.0 + layers.cast(step_next_idx, accu_scores.dtype)) / 6.0),
self.length_penalty)
curr_scores = layers.elementwise_div(accu_scores,
cur_timestep_length_penalty)
# beam_search op uses lod to differentiate branches.
curr_scores = layers.lod_reset(curr_scores, pre_ids)
topk_indices = layers.lod_reset(topk_indices, pre_ids)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=curr_scores,
beam_size=self.beam_size,
end_id=self.eos_id,
return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
layers.increment(x=step_next_idx, value=1.0, in_place=True)
# cell states(caches) have been updated in wrap_decoder,
# only need to update beam search states here.
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.array_write(pre_mask, i=step_idx, array=tgt_masks)
layers.array_write(pos_bias, i=step_idx, array=pos_biases)
layers.assign(gather_idx, parent_idx)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=self.beam_size, end_id=self.eos_id)
graph_vars = {
"finished_ids": finished_ids,
"finished_scores": finished_scores,
"data_ids": data_ids
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
