def create_model(args, ernie_config):
input_names = ("src_ids", "sent_ids", "pos_ids", "task_ids", "input_mask")
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1]]
dtypes=[
'int64', 'int64', 'int64', 'int64', 'float32'
]
inputs = [fluid.data(name, shape, dtype=dtype) for name, shape, dtype in zip(input_names, shapes, dtypes)]
(src_ids, sent_ids, pos_ids, task_ids, input_mask) = inputs
ernie = ErnieModel(
src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
seq_out = ernie.get_sequence_output()
cls_feats = ernie.get_pooled_output()
# dummy layers to name the latent layers. the save_inf_model produce uncomprehensible names
# like 'save_infer_model/scale_1'
seq_out = fluid.layers.scale(seq_out, scale=1.0, name='ernie_sequence_latent')
cls_feats = fluid.layers.scale(cls_feats, scale=1.0, name='ernie_classification')
for i, inp in enumerate(inputs):
print(f'input[{i}]:', inp.name, inp.shape, inp.dtype)
print('sequence_output :', seq_out.name, seq_out.shape, seq_out.dtype)
print('classifier_output:', cls_feats.name, cls_feats.shape, cls_feats.dtype)
return inputs, [seq_out, cls_feats]
def create_model(args, pyreader_name, ernie_config):
pyreader = fluid.layers.py_reader(
capacity=50,
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1], [-1, 1]],
dtypes=['int64', 'int64', 'int64', 'float', 'int64'],
lod_levels=[0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, input_mask,
seq_lens) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
input_mask=input_mask,
config=ernie_config)
enc_out = ernie.get_sequence_output()
unpad_enc_out = fluid.layers.sequence_unpad(enc_out, length=seq_lens)
cls_feats = ernie.get_pooled_output()
# set persistable = True to avoid memory opimizing
enc_out.persistable = True
unpad_enc_out.persistable = True
cls_feats.persistable = True
graph_vars = {
"cls_embeddings": cls_feats,
"top_layer_embeddings": unpad_enc_out,
}
return pyreader, graph_vars
def create_model(args, pyreader_name, ernie_config, is_prediction=False):
pyreader = fluid.layers.py_reader(
capacity=50,
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, args.max_seq_len],
[-1, args.max_seq_len, 1], [-1, 1]],
dtypes=['int64', 'int64', 'int64', 'float', 'int64', 'int64'],
lod_levels=[0, 0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, self_attn_mask, labels,
seq_lens) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
self_attn_mask=self_attn_mask,
config=ernie_config,
use_fp16=args.use_fp16)
enc_out = ernie.get_sequence_output()
logits = fluid.layers.fc(
input=enc_out,
size=args.num_labels,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_seq_label_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_seq_label_out_b",
initializer=fluid.initializer.Constant(0.)))
ret_labels = fluid.layers.reshape(x=labels, shape=[-1, 1])
ret_infers = fluid.layers.reshape(x=fluid.layers.argmax(logits, axis=2),
shape=[-1, 1])
labels = fluid.layers.flatten(labels, axis=2)
ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
logits=fluid.layers.flatten(logits, axis=2),
label=labels,
return_softmax=True)
loss = fluid.layers.mean(x=ce_loss)
if args.use_fp16 and args.loss_scaling > 1.0:
loss *= args.loss_scaling
graph_vars = {
"loss": loss,
"probs": probs,
"labels": ret_labels,
"infers": ret_infers,
"seq_lens": seq_lens
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def forward(self, features):
src_ids, sent_ids, input_seqlen = features
zero = L.fill_constant([1], dtype='int64', value=0)
input_mask = L.cast(L.equal(src_ids, zero),
'float32') # 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
model = 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['use_fp16'])
enc_out = model.get_sequence_output()
logits = L.fc(
input=enc_out,
size=self.num_label,
num_flatten_dims=2,
param_attr=F.ParamAttr(
name="cls_seq_label_out_w",
initializer=F.initializer.TruncatedNormal(scale=0.02)),
bias_attr=F.ParamAttr(name="cls_seq_label_out_b",
initializer=F.initializer.Constant(0.)))
propeller.summary.histogram('pred', logits)
return logits, input_seqlen
def create_model(args, pyreader_name, ernie_config, is_prediction=False):
src_ids = fluid.layers.data(name='1',
shape=[-1, args.max_seq_len, 1],
dtype='int64')
sent_ids = fluid.layers.data(name='2',
shape=[-1, args.max_seq_len, 1],
dtype='int64')
pos_ids = fluid.layers.data(name='3',
shape=[-1, args.max_seq_len, 1],
dtype='int64')
task_ids = fluid.layers.data(name='4',
shape=[-1, args.max_seq_len, 1],
dtype='int64')
input_mask = fluid.layers.data(name='5',
shape=[-1, args.max_seq_len, 1],
dtype='float32')
labels = fluid.layers.data(name='7',
shape=[-1, args.max_seq_len, 1],
dtype='int64')
seq_lens = fluid.layers.data(name='8', shape=[-1], dtype='int64')
pyreader = fluid.io.DataLoader.from_generator(feed_list=[
src_ids, sent_ids, pos_ids, task_ids, input_mask, labels, seq_lens
],
capacity=70,
iterable=False)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
enc_out = ernie.get_sequence_output()
enc_out = fluid.layers.dropout(x=enc_out,
dropout_prob=0.1,
dropout_implementation="upscale_in_train")
logits = fluid.layers.fc(
input=enc_out,
size=args.num_labels,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_seq_label_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_seq_label_out_b",
initializer=fluid.initializer.Constant(0.)))
infers = fluid.layers.argmax(logits, axis=2)
ret_infers = fluid.layers.reshape(x=infers, shape=[-1, 1])
lod_labels = fluid.layers.sequence_unpad(labels, seq_lens)
lod_infers = fluid.layers.sequence_unpad(infers, seq_lens)
(_, _, _, num_infer, num_label, num_correct) = fluid.layers.chunk_eval(
input=lod_infers,
label=lod_labels,
chunk_scheme=args.chunk_scheme,
num_chunk_types=((args.num_labels - 1) //
(len(args.chunk_scheme) - 1)))
labels = fluid.layers.flatten(labels, axis=2)
ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
logits=fluid.layers.flatten(logits, axis=2),
label=labels,
return_softmax=True)
input_mask = fluid.layers.flatten(input_mask, axis=2)
ce_loss = ce_loss * input_mask
loss = fluid.layers.mean(x=ce_loss)
graph_vars = {
"inputs": src_ids,
"loss": loss,
"probs": probs,
"seqlen": seq_lens,
"num_infer": num_infer,
"num_label": num_label,
"num_correct": num_correct,
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def create_model(args, pyreader_name, ernie_config, is_training):
pyreader = fluid.layers.py_reader(capacity=50,
shapes=[[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, 1], [-1, 1], [-1, 1]],
dtypes=[
'int64', 'int64', 'int64', 'int64',
'float32', 'int64', 'int64', 'int64'
],
lod_levels=[0, 0, 0, 0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, task_ids, input_mask, start_positions,
end_positions, unique_id) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
enc_out = ernie.get_sequence_output()
enc_out = fluid.layers.dropout(x=enc_out,
dropout_prob=0.1,
dropout_implementation="upscale_in_train")
logits = fluid.layers.fc(
input=enc_out,
size=2,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_mrc_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_mrc_out_b",
initializer=fluid.initializer.Constant(0.)))
logits = fluid.layers.transpose(x=logits, perm=[2, 0, 1])
start_logits, end_logits = fluid.layers.unstack(x=logits, axis=0)
batch_ones = fluid.layers.fill_constant_batch_size_like(input=start_logits,
dtype='int64',
shape=[1],
value=1)
num_seqs = fluid.layers.reduce_sum(input=batch_ones)
def compute_loss(logits, positions):
loss = fluid.layers.softmax_with_cross_entropy(logits=logits,
label=positions)
loss = fluid.layers.mean(x=loss)
return loss
start_loss = compute_loss(start_logits, start_positions)
end_loss = compute_loss(end_logits, end_positions)
loss = (start_loss + end_loss) / 2.0
if args.use_fp16 and args.loss_scaling > 1.0:
loss *= args.loss_scaling
graph_vars = {
"loss": loss,
"num_seqs": num_seqs,
"unique_id": unique_id,
"start_logits": start_logits,
"end_logits": end_logits
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def create_model(ernie_config, is_training=False):
if is_training:
input_fields = {
'names': [
'src_ids', 'pos_ids', 'sent_ids', 'input_mask',
'start_positions', 'end_positions'
],
'shapes': [[None, None], [None, None], [None, None],
[None, None, 1], [None, 1], [None, 1]],
'dtypes': ['int64', 'int64', 'int64', 'float32', 'int64', 'int64'],
'lod_levels': [0, 0, 0, 0, 0, 0],
}
else:
input_fields = {
'names':
['src_ids', 'pos_ids', 'sent_ids', 'input_mask', 'unique_id'],
'shapes': [[None, None], [None, None], [None, None],
[None, None, 1], [None, 1]],
'dtypes': ['int64', 'int64', 'int64', 'float32', 'int64'],
'lod_levels': [0, 0, 0, 0, 0],
}
inputs = [
fluid.data(name=input_fields['names'][i],
shape=input_fields['shapes'][i],
dtype=input_fields['dtypes'][i],
lod_level=input_fields['lod_levels'][i])
for i in range(len(input_fields['names']))
]
data_loader = fluid.io.DataLoader.from_generator(feed_list=inputs,
capacity=50,
iterable=False)
if is_training:
(src_ids, pos_ids, sent_ids, input_mask, start_positions,
end_positions) = inputs
else:
(src_ids, pos_ids, sent_ids, input_mask, unique_id) = inputs
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
enc_out = ernie.get_sequence_output()
logits = fluid.layers.fc(
input=enc_out,
size=2,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_squad_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_squad_out_b",
initializer=fluid.initializer.Constant(0.)))
logits = fluid.layers.transpose(x=logits, perm=[2, 0, 1])
start_logits, end_logits = fluid.layers.unstack(x=logits, axis=0)
batch_ones = fluid.layers.fill_constant_batch_size_like(input=start_logits,
dtype='int64',
shape=[1],
value=1)
num_seqs = fluid.layers.reduce_sum(input=batch_ones)
if is_training:
def compute_loss(logits, positions):
loss = fluid.layers.softmax_with_cross_entropy(logits=logits,
label=positions)
loss = fluid.layers.mean(x=loss)
return loss
start_loss = compute_loss(start_logits, start_positions)
end_loss = compute_loss(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2.0
return data_loader, total_loss, num_seqs
else:
return data_loader, unique_id, start_logits, end_logits, num_seqs
def create_model(args, pyreader_name, ernie_config, is_prediction=False):
pyreader = fluid.layers.py_reader(capacity=50,
shapes=[[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, 1]],
dtypes=[
'int64', 'int64', 'int64', 'int64',
'float32', 'int64', 'int64'
],
lod_levels=[0, 0, 0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, task_ids, input_mask, labels,
seq_lens) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
enc_out = ernie.get_sequence_output()
enc_out = fluid.layers.dropout(x=enc_out,
dropout_prob=0.1,
dropout_implementation="upscale_in_train")
logits = fluid.layers.fc(
input=enc_out,
size=args.num_labels,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_seq_label_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_seq_label_out_b",
initializer=fluid.initializer.Constant(0.)))
infers = fluid.layers.argmax(logits, axis=2)
ret_labels = fluid.layers.reshape(x=labels, shape=[-1, 1])
ret_infers = fluid.layers.reshape(x=infers, shape=[-1, 1])
lod_labels = fluid.layers.sequence_unpad(labels, seq_lens)
lod_infers = fluid.layers.sequence_unpad(infers, seq_lens)
(_, _, _, num_infer, num_label, num_correct) = fluid.layers.chunk_eval(
input=lod_infers,
label=lod_labels,
chunk_scheme=args.chunk_scheme,
num_chunk_types=((args.num_labels - 1) //
(len(args.chunk_scheme) - 1)))
labels = fluid.layers.flatten(labels, axis=2)
ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
logits=fluid.layers.flatten(logits, axis=2),
label=labels,
return_softmax=True)
input_mask = fluid.layers.flatten(input_mask, axis=2)
ce_loss = ce_loss * input_mask
loss = fluid.layers.mean(x=ce_loss)
if args.use_fp16 and args.loss_scaling > 1.0:
loss *= args.loss_scaling
graph_vars = {
"loss": loss,
"probs": probs,
"labels": ret_labels,
"infers": ret_infers,
"num_infer": num_infer,
"num_label": num_label,
"num_correct": num_correct,
"seq_lens": seq_lens
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def create_model(args, pyreader_name, ernie_config):
pyreader = fluid.layers.py_reader(
capacity=50, #缓冲区的容量数据个数
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, args.num_labels], [-1, 1], [-1, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1]],
dtypes=[
'int64', 'int64', 'int64', 'int64', 'float32', 'float32', 'int64',
'int64', 'int64', 'int64'
],
lod_levels=[0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, task_ids, input_mask, labels, seq_lens,
example_index, tok_to_orig_start_index,
tok_to_orig_end_index) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
#embedding+encoder
enc_out = ernie.get_sequence_output()
#get encoder layer
enc_out = fluid.layers.dropout(x=enc_out,
dropout_prob=0.1,
dropout_implementation="upscale_in_train")
logits = fluid.layers.fc(
input=enc_out,
size=args.num_labels,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_seq_label_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_seq_label_out_b",
initializer=fluid.initializer.Constant(0.)))
logits = fluid.layers.sigmoid(logits)
lod_labels = fluid.layers.sequence_unpad(labels, seq_lens)
lod_logits = fluid.layers.sequence_unpad(logits, seq_lens)
lod_tok_to_orig_start_index = fluid.layers.sequence_unpad(
tok_to_orig_start_index, seq_lens)
lod_tok_to_orig_end_index = fluid.layers.sequence_unpad(
tok_to_orig_end_index, seq_lens)
labels = fluid.layers.flatten(labels, axis=2)
logits = fluid.layers.flatten(logits, axis=2)
input_mask = fluid.layers.flatten(input_mask, axis=2)
# calculate loss
log_logits = fluid.layers.log(logits)
log_logits_neg = fluid.layers.log(1 - logits)
ce_loss = 0. - labels * log_logits - (1 - labels) * log_logits_neg
ce_loss = fluid.layers.reduce_mean(ce_loss, dim=1, keep_dim=True)
ce_loss = ce_loss * input_mask
loss = fluid.layers.mean(x=ce_loss)
graph_vars = {
"inputs": src_ids,
"loss": loss,
"seqlen": seq_lens,
"lod_logit": lod_logits,
"lod_label": lod_labels,
"example_index": example_index,
"tok_to_orig_start_index": lod_tok_to_orig_start_index,
"tok_to_orig_end_index": lod_tok_to_orig_end_index
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def create_model(args, pyreader_name, ernie_config, is_prediction=False):
"""func"""
pyreader = fluid.layers.py_reader(capacity=50,
shapes=[[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1]],
dtypes=[
'int64', 'int64', 'int64', 'int64',
'float32', 'int64', 'int64'
],
lod_levels=[0, 0, 0, 0, 0, 0, 0],
name=pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, task_ids, input_mask, labels,
seq_lens) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
enc_out = ernie.get_sequence_output()
emission = fluid.layers.fc(
input=enc_out,
size=args.num_labels,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-0.1, high=0.1),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)),
num_flatten_dims=2)
crf_cost = fluid.layers.linear_chain_crf(
input=emission,
label=labels,
param_attr=fluid.ParamAttr(name='crfw',
learning_rate=args.crf_learning_rate),
length=seq_lens)
loss = fluid.layers.mean(x=crf_cost)
crf_decode = fluid.layers.crf_decoding(
input=emission,
param_attr=fluid.ParamAttr(name='crfw'),
length=seq_lens)
lod_labels = fluid.layers.squeeze(labels, axes=[-1])
num_chunk_types = (
(args.num_labels - 1) // (len(args.chunk_scheme) - 1)) # IOB配置
(_, _, _, num_infer, num_label,
num_correct) = fluid.layers.chunk_eval(input=crf_decode,
label=lod_labels,
chunk_scheme=args.chunk_scheme,
num_chunk_types=num_chunk_types,
seq_length=seq_lens)
"""
enc_out = fluid.layers.dropout(x=enc_out,
dropout_prob=0.1,
dropout_implementation="upscale_in_train")
logits = fluid.layers.fc(
input=enc_out,
size=args.num_labels,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(
name="cls_seq_label_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name="cls_seq_label_out_b",
initializer=fluid.initializer.Constant(0.)))
infers = fluid.layers.argmax(logits, axis=2)
ret_infers = fluid.layers.reshape(x=infers, shape=[-1, 1])
lod_labels = fluid.layers.sequence_unpad(labels, seq_lens)
lod_infers = fluid.layers.sequence_unpad(infers, seq_lens)
num_chunk_types = (
(args.num_labels - 1) // (len(args.chunk_scheme) - 1)) # IOB配置
(_, _, _, num_infer, num_label,
num_correct) = fluid.layers.chunk_eval(input=lod_infers,
label=lod_labels,
chunk_scheme=args.chunk_scheme,
num_chunk_types=num_chunk_types)
labels = fluid.layers.flatten(labels, axis=2)
ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
logits=fluid.layers.flatten(logits, axis=2),
label=labels,
return_softmax=True)
input_mask = fluid.layers.flatten(input_mask, axis=2)
ce_loss = ce_loss * input_mask
loss = fluid.layers.mean(x=ce_loss)
"""
graph_vars = {
"inputs": src_ids,
"loss": loss,
"seqlen": seq_lens,
"crf_decode": crf_decode,
"num_infer": num_infer,
"num_label": num_label,
"num_correct": num_correct,
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def create_model(self, decoding=False):
if decoding:
return self.infilling_decode()
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]]
query_input_shapes = [[-1, self.max_seq_len, 1]] * emb_num + \
[[-1, self.max_seq_len, self.max_seq_len * 2]]
input_dtypes = ['int64'] * emb_num + ['float32']
input_lod_levels = [0] * emb_num + [0]
shapes = input_shapes + query_input_shapes + [[-1, 1], [-1, 1]]
dtypes = input_dtypes * 2 + ['int64', 'int64']
lod_levels = input_lod_levels * 2 + [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[0][key] = value
for key, value in zip(self.emb_keys,
inputs[emb_num + 1:emb_num * 2 + 1]):
emb_ids[1][key] = value
input_mask, input_query_mask = inputs[emb_num], inputs[2 * emb_num + 1]
tgt_labels, tgt_pos = inputs[-2:]
ernie = ErnieModel(emb_ids=emb_ids,
input_mask=[input_mask, input_query_mask],
config=self.ernie_config,
use_fp16=self.use_fp16,
task_type=self.task_type)
enc_out = ernie.get_sequence_output()
fc_out = self.cal_logit(enc_out, tgt_pos)
if self.label_smooth:
out_size = self.ernie_config[
"tgt_vocab_size"] or self.ernie_config['vocab_size']
labels = fluid.layers.label_smooth(label=fluid.layers.one_hot(
input=tgt_labels, depth=out_size),
epsilon=self.label_smooth)
ce_loss = layers.softmax_with_cross_entropy(logits=fc_out,
label=labels,
soft_label=True)
#probs = fluid.layers.log(fluid.layers.softmax(fc_out))
#ce_loss = fluid.layers.kldiv_loss(probs, labels, reduction='batchmean')
else:
ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
logits=fc_out, label=tgt_labels, return_softmax=True)
loss = fluid.layers.mean(x=ce_loss)
graph_vars = {"loss": loss}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def create_model(args,
pyreader_name,
ernie_config,
is_prediction=False,
task_name="",
is_classify=False,
is_regression=False,
ernie_version="1.0"):
if is_classify:
# 增加邻接矩阵和核心词的shape
pyreader = fluid.layers.py_reader(capacity=50,
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, 1], [-1, 1],
[-1, args.max_seq_len, args.max_seq_len], [-1, 2]],
dtypes=[
'int64', 'int64', 'int64', 'int64', 'float32', 'int64', 'int64', 'int64',
'int64'
],
lod_levels=[0, 0, 0, 0, 0, 0, 0, 0, 0],
name=task_name + "_" + pyreader_name,
use_double_buffer=True)
elif is_regression:
pyreader = fluid.layers.py_reader(capacity=50,
shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
[-1, args.max_seq_len, 1], [-1, 1], [-1, 1]],
dtypes=['int64', 'int64', 'int64', 'int64', 'float32', 'float32', 'int64'],
lod_levels=[0, 0, 0, 0, 0, 0, 0],
name=task_name + "_" + pyreader_name,
use_double_buffer=True)
(src_ids, sent_ids, pos_ids, task_ids, input_mask, labels, qids, adj_mat,
head_ids) = fluid.layers.read_file(pyreader)
ernie = ErnieModel(src_ids=src_ids,
position_ids=pos_ids,
sentence_ids=sent_ids,
task_ids=task_ids,
input_mask=input_mask,
config=ernie_config,
use_fp16=args.use_fp16)
erinie_output = ernie.get_sequence_output()
cls_feats = ernie.get_pooled_output()
# 增加GAT网络
gat = gnn.GAT(input_size=768, hidden_size=100, output_size=50, dropout=0.0, alpha=0.1, heads=12, layer=2)
# 将ernie的表示和邻接矩阵输入到gat网络中得到包含句子结构信息的表示
gat_emb = gat.forward(erinie_output, adj_mat)
# 提取核心词的表示
gat_emb = utils.index_sample(gat_emb, head_ids)
# 将[CLS]和核心词的表示拼接，供下游网络使用
cls_feats = fluid.layers.concat([cls_feats, gat_emb], axis=1)
cls_feats = fluid.layers.dropout(x=cls_feats, dropout_prob=0.1, dropout_implementation="upscale_in_train")
logits = fluid.layers.fc(input=cls_feats,
size=args.num_labels,
param_attr=fluid.ParamAttr(name=task_name + "_cls_out_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(name=task_name + "_cls_out_b",
initializer=fluid.initializer.Constant(0.)))
if is_prediction:
probs = fluid.layers.softmax(logits)
feed_targets_name = [src_ids.name, sent_ids.name, pos_ids.name, input_mask.name]
if ernie_version == "2.0":
feed_targets_name += [task_ids.name]
return pyreader, probs, feed_targets_name
assert is_classify != is_regression, 'is_classify or is_regression must be true and only one of them can be true'
num_seqs = fluid.layers.create_tensor(dtype='int64')
if is_classify:
ce_loss, probs = fluid.layers.softmax_with_cross_entropy(logits=logits, label=labels, return_softmax=True)
loss = fluid.layers.mean(x=ce_loss)
accuracy = fluid.layers.accuracy(input=probs, label=labels, total=num_seqs)
graph_vars = {
"loss": loss,
"probs": probs,
"accuracy": accuracy,
"labels": labels,
"num_seqs": num_seqs,
"qids": qids
}
elif is_regression:
cost = fluid.layers.square_error_cost(input=logits, label=labels)
loss = fluid.layers.mean(x=cost)
graph_vars = {"loss": loss, "probs": logits, "labels": labels, "num_seqs": num_seqs, "qids": qids}
else:
raise ValueError('unsupported fine tune mode. only supported classify/regression')
return pyreader, graph_vars
