def get_variable_info(var):
if not isinstance(var, fluid.framework.Variable):
raise TypeError("var shoule be an instance of fluid.framework.Variable")
var_info = {
'name': var.name,
'stop_gradient': var.stop_gradient,
'is_data': var.is_data,
'error_clip': var.error_clip,
'type': var.type
}
try:
var_info['dtype'] = convert_dtype_to_string(var.dtype)
var_info['lod_level'] = var.lod_level
var_info['shape'] = var.shape
except:
pass
if isinstance(var, fluid.framework.Parameter):
var_info['trainable'] = var.trainable
var_info['optimize_attr'] = var.optimize_attr
var_info['regularizer'] = var.regularizer
if not version_compare(paddle.__version__, '1.8'):
var_info['gradient_clip_attr'] = var.gradient_clip_attr
var_info['do_model_average'] = var.do_model_average
else:
var_info['persistable'] = var.persistable
return var_info
def _check_paddle_version(self):
if version_compare(self.paddle_version, paddle.__version__):
logger.warning(
"This Module is generated by the PaddlePaddle with version %s, and the local PaddlePaddle version is %s, which may cause serious incompatible bug. Please upgrade PaddlePaddle to the latest version."
% (self.paddle_version, paddle.__version__))
return False
return True
def forward(self, input_ids, position_ids, segment_ids, input_mask):
if version_compare(paddle.__version__, '1.8'):
pooled_output, sequence_output = self.model_runner(
input_ids, position_ids, segment_ids, input_mask)
return {
'pooled_output': pooled_output,
'sequence_output': sequence_output
}
else:
raise RuntimeError(
'{} only support dynamic graph mode in paddle >= 1.8'.format(
self.name))
def _build_net(self):
if version_compare(paddle.__version__, "1.6"):
self.seq_len = fluid.layers.data(name="seq_len",
shape=[-1],
dtype='int64')
else:
self.seq_len = fluid.layers.data(name="seq_len",
shape=[1],
dtype='int64')
seq_len = fluid.layers.assign(self.seq_len)
if self.add_crf:
unpad_feature = fluid.layers.sequence_unpad(self.feature,
length=self.seq_len)
self.emission = fluid.layers.fc(
size=self.num_classes,
input=unpad_feature,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-0.1, high=0.1),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
size = self.emission.shape[1]
fluid.layers.create_parameter(shape=[size + 2, size],
dtype=self.emission.dtype,
name='crfw')
self.ret_infers = fluid.layers.crf_decoding(
input=self.emission, param_attr=fluid.ParamAttr(name='crfw'))
ret_infers = fluid.layers.assign(self.ret_infers)
return [ret_infers]
else:
self.logits = fluid.layers.fc(
input=self.feature,
size=self.num_classes,
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.)))
self.ret_infers = fluid.layers.reshape(x=fluid.layers.argmax(
self.logits, axis=2),
shape=[-1, 1])
ret_infers = fluid.layers.assign(self.ret_infers)
logits = self.logits
logits = fluid.layers.flatten(logits, axis=2)
logits = fluid.layers.softmax(logits)
self.num_labels = logits.shape[1]
return [logits]
def predict(self,
data,
load_best_model=True,
return_result=False,
accelerate_mode=True):
"""
make prediction for the input data.
Args:
data (list): the data will be predicted.
load_best_model (bool): load the best model or not
return_result (bool): return a readable result or just the raw run result
accelerate_mode (bool): use high-performance predictor or not
Returns:
RunState: the running result of predict phase
"""
if not version_compare(paddle.__version__,
"1.6.2") and accelerate_mode:
logger.warning(
"Fail to open predict accelerate mode as it does not support paddle < 1.6.2. Please update PaddlePaddle."
)
accelerate_mode = False
self.accelerate_mode = accelerate_mode
with self.phase_guard(phase="predict"):
self._predict_data = data
self._predict_start_event()
if load_best_model:
self.init_if_load_best_model()
else:
self.init_if_necessary()
if not self.accelerate_mode:
run_states = self._run()
else:
if not self._predictor:
self._predictor = self._create_predictor()
run_states = self._run_with_predictor()
self._predict_end_event(run_states)
self._predict_data = None
if return_result:
return self._postprocessing(run_states)
return run_states
def __init__(self,
name=None,
directory=None,
module_dir=None,
version=None,
max_seq_len=128,
**kwargs):
if not directory:
return
super(TransformerModule, self).__init__(
name=name,
directory=directory,
module_dir=module_dir,
version=version,
**kwargs)
self.max_seq_len = max_seq_len
if version_compare(paddle.__version__, '1.8'):
with tmp_dir() as _dir:
input_dict, output_dict, program = self.context(
max_seq_len=max_seq_len)
fluid.io.save_inference_model(
dirname=_dir,
main_program=program,
feeded_var_names=[
input_dict['input_ids'].name,
input_dict['position_ids'].name,
input_dict['segment_ids'].name,
input_dict['input_mask'].name
],
target_vars=[
output_dict["pooled_output"],
output_dict["sequence_output"]
],
executor=fluid.Executor(fluid.CPUPlace()))
with fluid.dygraph.guard():
self.model_runner = fluid.dygraph.StaticModelRunner(_dir)
def _build_net(self):
self.seq_len = fluid.layers.data(name="seq_len",
shape=[1],
dtype='int64',
lod_level=0)
if version_compare(paddle.__version__, "1.6"):
self.seq_len_used = fluid.layers.squeeze(self.seq_len, axes=[1])
else:
self.seq_len_used = self.seq_len
#增加gru层相关的代码
grnn_hidden_dim = 256 # 768
crf_lr = 0.2
bigru_num = 2
init_bound = 0.1
def _bigru_layer(input_feature):
"""define the bidirectional gru layer
"""
pre_gru = fluid.layers.fc(
input=input_feature,
size=grnn_hidden_dim * 3,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-init_bound,
high=init_bound),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
gru = fluid.layers.dynamic_gru(
input=pre_gru,
size=grnn_hidden_dim,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-init_bound,
high=init_bound),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
pre_gru_r = fluid.layers.fc(
input=input_feature,
size=grnn_hidden_dim * 3,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-init_bound,
high=init_bound),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
gru_r = fluid.layers.dynamic_gru(
input=pre_gru_r,
size=grnn_hidden_dim,
is_reverse=True,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-init_bound,
high=init_bound),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
bi_merge = fluid.layers.concat(input=[gru, gru_r], axis=1)
return bi_merge
if self.add_crf:
unpad_feature = fluid.layers.sequence_unpad(
self.feature, length=self.seq_len_used)
#增加gru层相关的代码
input_feature = unpad_feature
for i in range(bigru_num):
bigru_output = _bigru_layer(input_feature)
input_feature = bigru_output
unpad_feature = input_feature
self.emission = fluid.layers.fc(
size=self.num_classes,
input=unpad_feature,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(low=-0.1, high=0.1),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
size = self.emission.shape[1]
fluid.layers.create_parameter(shape=[size + 2, size],
dtype=self.emission.dtype,
name='crfw')
self.ret_infers = fluid.layers.crf_decoding(
input=self.emission, param_attr=fluid.ParamAttr(name='crfw'))
ret_infers = fluid.layers.assign(self.ret_infers)
return [ret_infers]
else:
self.logits = fluid.layers.fc(
input=self.feature,
size=self.num_classes,
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.)))
self.ret_infers = fluid.layers.reshape(x=fluid.layers.argmax(
self.logits, axis=2),
shape=[-1, 1])
logits = self.logits
logits = fluid.layers.flatten(logits, axis=2)
logits = fluid.layers.softmax(logits)
self.num_labels = logits.shape[1]
return [logits]