def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file,
pytorch_dump_path):
config_path = os.path.abspath(bert_config_file)
tf_path = os.path.abspath(tf_checkpoint_path)
print("Converting TensorFlow checkpoint from {} with config at {}".format(
tf_path, config_path))
# Load weights from TF model
init_vars = tf.train.list_variables(tf_path)
names = []
arrays = []
for name, shape in init_vars:
print("Loading TF weight {} with shape {}".format(name, shape))
array = tf.train.load_variable(tf_path, name)
names.append(name)
arrays.append(array)
# Initialise PyTorch model
config = BertConfig.from_json_file(bert_config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
model = BertForPreTraining(config)
for name, array in zip(names, arrays):
name = name.split('/')
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
# which are not required for using pretrained model
if any(n in ["adam_v", "adam_m"] for n in name):
print("Skipping {}".format("/".join(name)))
continue
pointer = model
for m_name in name:
if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
l = re.split(r'_(\d+)', m_name)
else:
l = [m_name]
if l[0] == 'kernel' or l[0] == 'gamma':
pointer = getattr(pointer, 'weight')
elif l[0] == 'output_bias' or l[0] == 'beta':
pointer = getattr(pointer, 'bias')
elif l[0] == 'output_weights':
pointer = getattr(pointer, 'weight')
else:
pointer = getattr(pointer, l[0])
if len(l) >= 2:
num = int(l[1])
pointer = pointer[num]
if m_name[-11:] == '_embeddings':
pointer = getattr(pointer, 'weight')
elif m_name == 'kernel':
array = np.transpose(array)
try:
assert pointer.shape == array.shape
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
print("Initialize PyTorch weight {}".format(name))
pointer.data = torch.from_numpy(array)
# Save pytorch-model
print("Save PyTorch model to {}".format(pytorch_dump_path))
torch.save(model.state_dict(), pytorch_dump_path)
def eval_all():
# output_model_file = "../../output/best_model"
output_model_file = MODEL_PATH
output_config_file = os.path.join('../model_dir/', args.config_name)
config = BertConfig(output_config_file)
model = BertForQuestionAnswering(config)
if not args.no_pai:
try:
model.load_state_dict(torch.load(output_model_file))#, map_location='cpu'))
except:
model = nn.DataParallel(model)
model.load_state_dict(torch.load(output_model_file))#, map_location='cpu'))
else:
try:
model.load_state_dict(torch.load(output_model_file, map_location='cpu'))
except:
model = nn.DataParallel(model)
model.load_state_dict(torch.load(output_model_file, map_location='cpu'))
result_file_path = os.path.join('../metric', args.result_file_name)
evaluate(model, result_file=result_file_path)
if not args.no_pai:
print(os.getcwd())
pai_file_output = "/Container/thsi_yicui/dureader-bert/Dureader/output"
client.upload(pai_file_output, result_file_path, overwrite=True)
def __init__(self, output_dir: str, bert_config: str, max_seq_length: int):
bert_config = BertConfig.from_json_file(bert_config)
bert_config.experimental_gelu = FLAGS.experimental_gelu
if FLAGS.precision:
bert_config.precision = FLAGS.precision
self.session = tf.compat.v1.Session()
placeholder = tf.compat.v1.placeholder
input_shape = [None, max_seq_length]
self.input_ids = placeholder(tf.int32, input_shape, name='input_ids')
self.input_mask = placeholder(tf.int32, input_shape, name='input_mask')
self.segment_ids = placeholder(tf.int32,
input_shape,
name='segment_ids')
(self.start_logits, self.end_logits) = \
create_model_top(bert_config, False, # is training
self.input_ids, self.input_mask, self.segment_ids,
False, # use_one_hot_embeddings
None) # frozen graph path
latest_model = tf.train.latest_checkpoint(FLAGS.output_dir)
saver = tf.compat.v1.train.Saver()
saver.restore(self.session, latest_model)
self.output_dir = output_dir
self.dest_dir = os.path.join(self.output_dir, "frozen")
if not os.path.exists(self.dest_dir):
os.mkdir(self.dest_dir)
def eval_all():
# output_model_file = "../../output/best_model"
output_model_file = MODEL_PATH
output_config_file = CONFIG_PATH
config = BertConfig(output_config_file)
model = BertForQuestionAnswering(config)
if next(model.parameters()).is_cuda:
try:
model.load_state_dict(torch.load(output_model_file))
except:
model = nn.DataParallel(model)
model.load_state_dict(torch.load(output_model_file))
else:
try:
model.load_state_dict(
torch.load(output_model_file, map_location='cpu'))
except:
model = nn.DataParallel(model)
model.load_state_dict(
torch.load(output_model_file, map_location='cpu'))
result_file_path = os.path.join('../metric', args.result_file_name)
evaluate(model, result_file=result_file_path)
def prepare_model(args, device):
# Prepare model
config = BertConfig.from_json_file(args.bert_config_path)
# Padding for divisibility by 8
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
print('padded vocab size to: {}'.format(config.vocab_size))
# Set some options that the config file is expected to have (but don't need to be set properly
# at this point)
config.pad = False
config.unpad = False
config.dense_seq_output = False
config.fused_mha = False
config.fused_gelu_bias = False
config.fuse_qkv = False
config.fuse_scale = False
config.fuse_mask = False
config.fuse_dropout = False
config.apex_softmax = False
config.enable_stream = False
if config.fuse_mask == True: config.apex_softmax = True
if config.pad == False: config.enable_stream = True
if config.unpad == True: config.fused_mha = False
#Load from TF checkpoint
model = BertForPreTraining.from_pretrained(args.tf_checkpoint,
from_tf=True,
config=config)
return model
def __init__(self):
self.bert_config = BertConfig.from_json_file(
os.path.join(path, 'uncased_L-12_H-768_A-12/bert_config.json'))
self.max_sequence_length = 128
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
if self.max_sequence_length > self.bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length {} because the BERT model was only trained up to sequence length {}"
.format(self.max_sequence_length,
self.bert_config.max_position_embeddings))
self.processor = LogicProcessor()
self.label_list = self.processor.get_labels()
self.tokenizer = tokenization.FullTokenizer(vocab_file=os.path.join(
path, 'uncased_L-12_H-768_A-12/vocab.txt'),
do_lower_case=False)
self.model = BertForSequenceClassification(self.bert_config,
len(self.label_list))
init_checkpoint = os.path.join(path, 'model/logic_model_500.bin')
#Future save model Load code
if init_checkpoint is not None:
self.model.load_state_dict(
torch.load(init_checkpoint, map_location='cpu'))
self.model.to(self.device)
def load_stock_model(model_dir, max_seq_len):
from modeling import BertModel, BertConfig, get_assignment_map_from_checkpoint
tf.compat.v1.reset_default_graph(
) # to scope naming for checkpoint loading (if executed more than once)
bert_config_file = os.path.join(model_dir, "bert_config.json")
bert_ckpt_file = os.path.join(model_dir, "bert_model.ckpt")
pl_input_ids = tf.compat.v1.placeholder(tf.int32,
shape=(1, max_seq_len))
pl_mask = tf.compat.v1.placeholder(tf.int32, shape=(1, max_seq_len))
pl_token_type_ids = tf.compat.v1.placeholder(tf.int32,
shape=(1, max_seq_len))
bert_config = BertConfig.from_json_file(bert_config_file)
s_model = BertModel(config=bert_config,
is_training=False,
input_ids=pl_input_ids,
input_mask=pl_mask,
token_type_ids=pl_token_type_ids,
use_one_hot_embeddings=False)
tvars = tf.compat.v1.trainable_variables()
(assignment_map,
initialized_var_names) = get_assignment_map_from_checkpoint(
tvars, bert_ckpt_file)
tf.compat.v1.train.init_from_checkpoint(bert_ckpt_file, assignment_map)
return s_model, pl_input_ids, pl_token_type_ids, pl_mask
def get_model_from_args(args):
config = BertConfig.from_json_file(args.config_file)
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
class BertForQuestionAnswering_int32_inputs(BertForQuestionAnswering):
def forward(self, input_ids, segment_ids, attention_mask):
input_ids, segment_ids, attention_mask = input_ids.long(
), segment_ids.long(), attention_mask.long()
return super().forward(input_ids, segment_ids, attention_mask)
model = BertForQuestionAnswering_int32_inputs(config)
model.enable_apex(False)
if os.path.isfile(args.checkpoint):
state_dict = torch.load(args.checkpoint, map_location="cpu")
state_dict = state_dict["model"] if "model" in state_dict.keys(
) else state_dict
model.load_state_dict(state_dict, strict=False)
if args.precision == "fp16":
model = model.half()
device = "cuda:0" if not args.cpu else "cpu"
model = model.to(device)
model.eval()
model.bermuda_batch_axis = 0 if not args.fixed_batch_dim else None
return model
def load_query_encoder(self, device, args):
# Configure paths for query encoder serving
vocab_path = os.path.join(args.metadata_dir, args.vocab_name)
bert_config_path = os.path.join(
args.metadata_dir, args.bert_config_name.replace(".json", "") + "_" + args.bert_model_option + ".json"
)
# Load pretrained QueryEncoder
bert_config = BertConfig.from_json_file(bert_config_path)
model = DenSPI(bert_config)
if args.parallel:
model = torch.nn.DataParallel(model)
state = torch.load(args.query_encoder_path, map_location='cpu')
try:
model.load_state_dict(state['model'])
logger.info('load okay')
except:
model.load_state_dict(state, strict=False)
check_diff(model.state_dict(), state['model'])
logger.info('Model loaded from %s' % args.query_encoder_path)
model.to(device)
tokenizer = tokenization.FullTokenizer(vocab_file=vocab_path, do_lower_case=not args.do_case)
logger.info('Model loaded from %s' % args.query_encoder_path)
logger.info('Number of model parameters: {:,}'.format(sum(p.numel() for p in model.parameters())))
return model, tokenizer
def convert():
# Initialise PyTorch model
config = BertConfig.from_json_file(args.bert_config_file)
model = BertModel(config)
# Load weights from TF model
path = args.tf_checkpoint_path
print("Converting TensorFlow checkpoint from {}".format(path))
init_vars = tf.train.list_variables(path)
names = []
arrays = []
for name, shape in init_vars:
print("Loading {} with shape {}".format(name, shape))
array = tf.train.load_variable(path, name)
print("Numpy array shape {}".format(array.shape))
names.append(name)
arrays.append(array)
for name, array in zip(names, arrays):
if not name.startswith("bert"):
print("Skipping {}".format(name))
continue
else:
name = name.replace("bert/", "") # skip "bert/"
print("Loading {}".format(name))
name = name.split('/')
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
# which are not required for using pretrained model
if name[0] in ['redictions', 'eq_relationship'
] or name[-1] == "adam_v" or name[-1] == "adam_m":
print("Skipping {}".format("/".join(name)))
continue
pointer = model
for m_name in name:
if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
l = re.split(r'_(\d+)', m_name)
else:
l = [m_name]
if l[0] == 'kernel':
pointer = getattr(pointer, 'weight')
else:
pointer = getattr(pointer, l[0])
if len(l) >= 2:
num = int(l[1])
pointer = pointer[num]
if m_name[-11:] == '_embeddings':
pointer = getattr(pointer, 'weight')
elif m_name == 'kernel':
array = np.transpose(array)
try:
assert pointer.shape == array.shape
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
pointer.data = torch.from_numpy(array)
# Save pytorch-model
torch.save(model.state_dict(), args.pytorch_dump_path)
def convert_tmp_to_pytorch(bert_config_file, pytorch_dump_path):
import torch
from modeling import BertConfig, BertForPreTraining
import pickle
with open("tmp_names", "rb") as fp: # Unpickling
# names = pickle.load(fp, encoding='iso-8859-1')
names = pickle.load(fp)
with open("tmp_arrays", "rb") as fp: # Unpickling
# arrays = pickle.load(fp, encoding='iso-8859-1')
arrays = pickle.load(fp)
# Initialise PyTorch model
config = BertConfig.from_json_file(bert_config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
model = BertForPreTraining(config)
for name, array in zip(names, arrays):
name = name.split('/')
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
# which are not required for using pretrained model
if name[-1] in ["adam_v", "adam_m", 'global_step']:
print("Skipping {}".format("/".join(name)))
continue
pointer = model
for m_name in name:
if fullmatch(r'[A-Za-z]+_\d+', m_name):
# if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
l = re.split(r'_(\d+)', m_name)
else:
l = [m_name]
if l[0] == 'kernel':
pointer = getattr(pointer, 'weight')
elif l[0] == 'output_bias':
pointer = getattr(pointer, 'bias')
elif l[0] == 'output_weights':
pointer = getattr(pointer, 'weight')
else:
pointer = getattr(pointer, l[0])
if len(l) >= 2:
num = int(l[1])
pointer = pointer[num]
if m_name[-11:] == '_embeddings':
pointer = getattr(pointer, 'weight')
elif m_name == 'kernel':
array = np.transpose(array)
try:
assert pointer.shape == array.shape
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
print("Initialize PyTorch weight {}".format(name))
pointer.data = torch.from_numpy(array)
# Save pytorch-model
print("Save PyTorch model to {}".format(pytorch_dump_path))
torch.save(model.state_dict(), pytorch_dump_path)
def convert():
args = parser.parse_args()
args.tf_checkpoint_path = "chinese_L-12_H-768_A-12\\bert_model.ckpt"
args.bert_config_file = "chinese_L-12_H-768_A-12\\bert_config.json"
args.pytorch_dump_path = "chinese_L-12_H-768_A-12\pytorch_model.bin"
# Initialise PyTorch model
config = BertConfig.from_json_file(args.bert_config_file)
model = BertModel(config)
# Load weights from TF model
path = args.tf_checkpoint_path
print("Converting TensorFlow checkpoint from {}".format(path))
init_vars = tf.train.list_variables(path)
names = []
arrays = []
for name, shape in init_vars:
print("Loading {} with shape {}".format(name, shape))
array = tf.train.load_variable(path, name)
print("Numpy array shape {}".format(array.shape))
names.append(name)
arrays.append(array)
for name, array in zip(names, arrays):
name = name[5:] # skip "bert/"
print("Loading {}".format(name))
name = name.split('/')
if name[0] in ['redictions', 'eq_relationship']:
print("Skipping")
continue
pointer = model
for m_name in name:
if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
l = re.split(r'_(\d+)', m_name)
else:
l = [m_name]
if l[0] == 'kernel':
pointer = getattr(pointer, 'weight')
else:
pointer = getattr(pointer, l[0])
if len(l) >= 2:
num = int(l[1])
pointer = pointer[num]
if m_name[-11:] == '_embeddings':
pointer = getattr(pointer, 'weight')
elif m_name == 'kernel':
array = np.transpose(array)
try:
assert pointer.shape == array.shape
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
pointer.data = torch.from_numpy(array)
# Save pytorch-model
torch.save(model.state_dict(), args.pytorch_dump_path)
def get_predictor_model(cls):
config = BertConfig.from_json_file(config_file)
model = BertForQuestionAnswering(config)
model.load_state_dict(
torch.load(MODEL_PATH, map_location='cpu')["model"])
model.to(device)
cls.model = model
return cls.model
def eval_all():
output_model_file = "../model_dir/best_model"
output_config_file = "../model_dir/bert_config.json"
config = BertConfig(output_config_file)
model = BertForQuestionAnswering(config)
model.load_state_dict(
torch.load(output_model_file)) #, map_location='cpu'))
evaluate(model.cpu(), result_file="../metric/predicts.json")
def prepare_model_and_optimizer(self):
# Prepare model
self.config = BertConfig.from_json_file(self.args.config_file)
# Padding for divisibility by 8
if self.config.vocab_size % 8 != 0:
self.config.vocab_size += 8 - (self.config.vocab_size % 8)
self.model = BertForPreTraining(self.config)
self.another_model = BertForPreTraining(self.config)
self.model.to(self.device)
self.another_model.to(self.device)
param_optimizer = list(self.model.named_parameters())
no_decay = ['bias', 'gamma', 'beta', 'LayerNorm']
optimizer_grouped_parameters = []
names = []
for n, p in param_optimizer:
if not any(nd in n for nd in no_decay):
optimizer_grouped_parameters.append({
'params': [p],
'weight_decay': 0.01,
'name': n
})
names.append({'params': [n], 'weight_decay': 0.01})
if any(nd in n for nd in no_decay):
optimizer_grouped_parameters.append({
'params': [p],
'weight_decay': 0.00,
'name': n
})
names.append({'params': [n], 'weight_decay': 0.00})
if self.args.phase2:
max_steps = self.args.max_steps
tmp = max_steps * 10
r = self.args.phase1_end_step / tmp
lr = self.args.learning_rate * (1 - r)
else:
max_steps = int(self.args.max_steps / 9 * 10)
lr = self.args.learning_rate
if self.args.optimizer == "lamb":
self.optimizer = BertLAMB(optimizer_grouped_parameters,
lr=lr,
warmup=self.args.warmup_proportion
if not self.args.phase2 else -1,
t_total=max_steps)
elif self.args.optimizer == "adam":
self.optimizer = BertAdam(optimizer_grouped_parameters,
lr=lr,
warmup=self.args.warmup_proportion
if not self.args.phase2 else -1,
t_total=max_steps)
def initialize_model(args):
''' return model, ready to trace '''
config = BertConfig.from_json_file(args.config_file)
if config.vocab_size % 8 != 0:
config.vocab_size += 8 - (config.vocab_size % 8)
model = BertForQuestionAnswering(config)
model.enable_apex(False)
state_dict = torch.load(args.checkpoint, map_location='cpu')["model"]
model.load_state_dict(state_dict)
if args.fp16:
model.half()
return model
def __init__(self,
model_name: str,
models_dir='_models',
device='/device:GPU:0',
is_training=False,
use_one_hot_embeddings=False,
verb=0):
self.model_name = model_name
self.models_dir = models_dir
if verb > 0:
print('\n*** BertMC *** initializing (folder: %s, model: %s)' %
(self.models_dir, self.model_name))
self.graph = tf.Graph()
with self.graph.as_default():
device = device_TF(devices=device, verb=verb)[0]
if verb > 1: print(' > building graph on cuda: %s' % device)
with tf.device(device):
self.features = {
'input_ids':
tf.placeholder(shape=[None, None], dtype=tf.int32),
'input_mask':
tf.placeholder(shape=[None, None], dtype=tf.int32),
'input_type_ids':
tf.placeholder(shape=[None, None], dtype=tf.int32)
}
super(BertMC, self).__init__(
config=BertConfig.from_json_file(self.models_dir + '/' +
self.model_name +
'/bert_config.json'),
is_training=is_training,
input_ids=self.features['input_ids'],
input_mask=self.features['input_mask'],
token_type_ids=self.features['input_type_ids'],
use_one_hot_embeddings=use_one_hot_embeddings)
self.tvars = tf.trainable_variables()
checkpoint = self.models_dir + '/' + self.model_name + '/bert_model.ckpt'
(assignment_map,
initialized_variable_names) = get_assignment_map_from_checkpoint(
self.tvars, checkpoint)
tf.train.init_from_checkpoint(checkpoint, assignment_map)
init = tf.global_variables_initializer()
self.sess = tf.Session(
graph=self.graph, config=tf.ConfigProto(allow_soft_placement=True))
self.sess.run(init)
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file,
pytorch_dump_path):
# Initialise PyTorch model
config = BertConfig.from_json_file(bert_config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
model = BertForPreTraining(config)
# Load weights from tf checkpoint
load_tf_weights_in_bert(model, tf_checkpoint_path)
# Save pytorch-model
print("Save PyTorch model to {}".format(pytorch_dump_path))
torch.save(model.state_dict(), pytorch_dump_path)
def bert_module_fn(is_training):
"""Spec function for a token embedding module."""
input_ids = tf.placeholder(shape=[None, None],
dtype=tf.int32,
name="input_ids")
input_mask = tf.placeholder(shape=[None, None],
dtype=tf.int32,
name="input_mask")
token_type = tf.placeholder(shape=[None, None],
dtype=tf.int32,
name="segment_ids")
config = BertConfig.from_json_file(config_path)
model = BertModel(config=config,
is_training=is_training,
input_ids=input_ids,
input_mask=input_mask,
token_type_ids=token_type)
model.input_to_output()
seq_output = model.get_all_encoder_layers()[-1]
config_file = tf.constant(value=config_path,
dtype=tf.string,
name="config_file")
vocab_file = tf.constant(value=vocab_path,
dtype=tf.string,
name="vocab_file")
lower_case = tf.constant(do_lower_case)
tf.add_to_collection(tf.GraphKeys.ASSET_FILEPATHS, config_file)
tf.add_to_collection(tf.GraphKeys.ASSET_FILEPATHS, vocab_file)
input_map = {
"input_ids": input_ids,
"input_mask": input_mask,
"segment_ids": token_type
}
output_map = {"sequence_output": seq_output}
output_info_map = {
"vocab_file": vocab_file,
"do_lower_case": lower_case
}
hub.add_signature(name="tokens", inputs=input_map, outputs=output_map)
hub.add_signature(name="tokenization_info",
inputs={},
outputs=output_info_map)
def __init__(self, output_dir: str, task_name: str, bert_config: str,
max_seq_length: int):
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"xnli": XnliProcessor
}
task_name = task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
num_labels = len(label_list)
# create model for CPU/dGPU, not TPU
use_one_hot_embeddings = False
bert_config = BertConfig.from_json_file(bert_config)
bert_config.experimental_gelu = FLAGS.experimental_gelu
if FLAGS.precision:
bert_config.precision = FLAGS.precision
self.session = tf.compat.v1.Session()
placeholder = tf.compat.v1.placeholder
input_shape = [None, max_seq_length]
self.label_ids = placeholder(tf.int32, [None], name='label_ids')
self.input_ids = placeholder(tf.int32, input_shape, name='input_ids')
self.input_mask = placeholder(tf.int32, input_shape, name='input_mask')
self.segment_ids = placeholder(tf.int32,
input_shape,
name='segment_ids')
self.loss, self.per_example_loss, self.logits, self.probabilities = \
create_model_top(bert_config, False, # is training
self.input_ids, self.input_mask, self.segment_ids,
self.label_ids, num_labels, use_one_hot_embeddings,
None) # frozen graph path
latest_model = tf.train.latest_checkpoint(FLAGS.output_dir)
saver = tf.compat.v1.train.Saver()
saver.restore(self.session, latest_model)
self.output_dir = output_dir
self.dest_dir = os.path.join(self.output_dir, "frozen")
if not os.path.exists(self.dest_dir):
os.mkdir(self.dest_dir)
def init_models():
models = []
for model_path, model_num, config_path in zip(MODEL_PATHS, MODEL_NUMS, CONFIG_PATHS):
config = BertConfig(config_path)
model = BertForQuestionAnswerings[model_num](config)
try:
model.load_state_dict(torch.load(model_path))
except:
model = nn.DataParallel(model)
model.load_state_dict(torch.load(model_path))
models.append(model)
return models
def __init__(self):
self.input_ids = tf.placeholder(tf.int32, [None, None])
self.input_mask = tf.placeholder(tf.int32, [None, None])
self.model = BertModel(config=BertConfig.from_json_file(bert_config),
is_training=True,
input_ids=self.input_ids,
input_mask=self.input_mask)
self.is_training = tf.placeholder(tf.bool, [])
self.predictions = self.construct_model(self.model)
self.id_predictions = tf.argmax(self.predictions, axis=2)
self.Y = tf.placeholder(tf.float32, [None, None, len(entity_types)])
self.tokenizer = tokenization.FullTokenizer(vocab_file,
do_lower_case=False)
self.model_path = "/scratch/sanjay/bert/bert_ner_model.ckpt"
self.saver = tf.train.Saver()
def init_modle(self, model):
print(f'starting to init model')
vocab_path = os.path.join(self.model_path, 'vocab.txt')
bert_config_file = os.path.join(self.model_path, 'bert_config.json')
self.bert_config = BertConfig.from_json_file(bert_config_file)
self.model = model(self.bert_config, 2)
weight_path = os.path.join(self.model_path, 'pytorch_model.bin')
new_state_dict = torch.load(weight_path)
new_state_dict = dict([(k[7:], v) if k.startswith('module') else (k, v)
for k, v in new_state_dict.items()])
self.model.load_state_dict(new_state_dict)
self.tokenizer = tokenization.FullTokenizer(vocab_file=vocab_path)
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.model.to(self.device)
self.model.eval()
print(f'init {model} model finished')
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file,
pytorch_dump_path):
# 加载模型参数
config = BertConfig.from_json_file(bert_config_file)
print("Building PyTorch model from configuration: {}".format(str(config)))
# 加载模型
model = BertForPreTraining(config)
# 加载检查点参数到模型中，进行处理
# 但是有一个问题，为什么加不加返回model都能返回？？？猜测其为内部已进行处理
load_tf_weights_in_bert(model, tf_checkpoint_path)
print("Save PyTorch model to {}".format(pytorch_dump_path))
# 保存pytorch的检查点
torch.save(model.state_dict(), pytorch_dump_path)
def test_encode_context(self):
x_context_value = [
[ 151, 12553, 8997, 8792, 10086, 8168, 10481, 9356, 8174, 10404, 9066, 10003, 10610, 10879]
+ [0 for i in range(312 - 14)],
[8670, 11136, 8997, 10564, 8303, 8228, 8373, 10003, 8307, 119, 151, 12553, 8233, 8815]
+ [0 for i in range(312 - 14)]
]
bert_scope = tf.VariableScope(name="bert", reuse=tf.AUTO_REUSE)
bert_config = BertConfig.from_json_file(self._poly_encoder_config.bert_config)
x_context = tf.convert_to_tensor(value=x_context_value, dtype=tf.int32)
context_vecs, poly_code_mask = self._encoder_inst.encode_context(x_context=x_context,
bert_config=bert_config,
bert_scope=bert_scope)
print(context_vecs)
print(poly_code_mask)
def test_encode_candidate(self):
x_response_value = [
[10378, 119, 119, 151, 8815, 8281, 8211, 10425, 8154, 0, 0, 0, 0, 0]
+ [0 for i in range(512 - 14)],
[165, 8991, 8181, 8184, 131, 120, 120, 8134, 11300, 10540, 8735, 8207, 0, 0]
+ [0 for i in range(512 - 14)]
]
bert_scope = tf.VariableScope(name="bert", reuse=tf.AUTO_REUSE)
bert_config = BertConfig.from_json_file(self._poly_encoder_config.bert_config)
x_response = tf.convert_to_tensor(value=x_response_value, dtype=tf.int32)
x_response_emb, x_response_mask = self._encoder_inst.encode_candidate(x_response=x_response,
bert_config=bert_config,
bert_scope=bert_scope)
print(x_response_emb)
print(x_response_mask)
def make_global_options(task_specific_parsers=[]):
# Parse command-line arguments
command_line_parser = create_command_line_parser()
all_options_parser = create_all_options_parser()
for task_parser in task_specific_parsers:
all_options_parser = task_parser(all_options_parser)
known_command_line_args, unknown_command_line_args = command_line_parser.parse_known_args(
)
if known_command_line_args.help or known_command_line_args.config is None:
all_options_parser.print_help()
sys.exit(os.EX_OK)
# Parse options specified in the configuration file into
config_file_path = known_command_line_args.config
opts_from_config_file = BertConfig.from_json_file(config_file_path)
# Build the global options structure from the default options
current_options = vars(all_options_parser.parse_args())
unknown_options = [
opt for opt in opts_from_config_file.keys()
if opt not in current_options.keys()
]
if unknown_options:
logging.error(f"Unonwn options: {unknown_options}")
sys.exit(os.EX_USAGE)
# Overwrite global options by those specified in the config file.
current_options.update(opts_from_config_file)
options_namespace = argparse.Namespace(**current_options)
# Overwrite with command-line arguments
all_options_namespace = all_options_parser.parse_args(
unknown_command_line_args, options_namespace)
logging.info(
f"Overwrite configuration parameters: {', '.join(unknown_command_line_args)}"
)
# argparse.Namespace -> dict()
opts = vars(all_options_namespace)
return opts
def create_stock_bert_graph(bert_config_file, max_seq_len):
from modeling import BertModel, BertConfig
tf_placeholder = tf.compat.v1.placeholder
pl_input_ids = tf_placeholder(tf.int32, shape=(1, max_seq_len))
pl_mask = tf_placeholder(tf.int32, shape=(1, max_seq_len))
pl_token_type_ids = tf_placeholder(tf.int32, shape=(1, max_seq_len))
bert_config = BertConfig.from_json_file(bert_config_file)
s_model = BertModel(config=bert_config,
is_training=False,
input_ids=pl_input_ids,
input_mask=pl_mask,
token_type_ids=pl_token_type_ids,
use_one_hot_embeddings=False)
return s_model, pl_input_ids, pl_mask, pl_token_type_ids
def feature_extactor(self, dummy=False):
if not dummy:
model_pre_trained = PreTrainedBertModel(self.model_dir,
Verbose=True)
else:
print("Using randomly initialized model...")
model_pre_trained = BERTModel(config=BertConfig(),
Verbose=True,
trainable=False)
model_pre_trained.build(input_shape=(self.batch_size,
self.max_seq_length))
print("Computing embeddings...")
self.bert_embeddings = model_pre_trained(K.variable(
self.all_input_ids))
print("Evaluating...")
self.all_encoder_layers = np.array(
[K.eval(emb) for emb in self.bert_embeddings[:-1]])
print("Output BERT shape: ", self.all_encoder_layers.shape)
def __init__(self, model_path):
""" to obtain sentences embeddings model
model path: init model weight path
"""
vocab_path = os.path.join(model_path, 'vocab.txt')
bert_config_file = os.path.join(model_path, 'bert_config.json')
self.bert_config = BertConfig.from_json_file(bert_config_file)
print(f'starting to init model')
self.model = TwoSentenceClassifier(self.bert_config, 2)
weight_path = os.path.join(model_path, 'pytorch_model.bin')
new_state_dict = torch.load(weight_path)
new_state_dict = dict([(k[7:], v) if k.startswith('module') else (k, v)
for k, v in new_state_dict.items()])
self.model.load_state_dict(new_state_dict)
self.tokenizer = tokenization.FullTokenizer(vocab_file=vocab_path)
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.model.to(self.device)
self.model.eval()
print(f'init model finished')
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
parser.add_argument("--bert_config_file",
default=None,
type=str,
required=True,
help="The config json file corresponding to the pre-trained BERT model. \n"
"This specifies the model architecture.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--vocab_file",
default=None,
type=str,
required=True,
help="The vocabulary file that the BERT model was trained on.")
parser.add_argument("--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written.")
## Other parameters
parser.add_argument("--init_checkpoint",
default=None,
type=str,
help="Initial checkpoint (usually from a pre-trained BERT model).")
parser.add_argument("--do_lower_case",
default=False,
action='store_true',
help="Whether to lower case the input text. True for uncased models, False for cased models.")
parser.add_argument("--max_seq_length",
default=128,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--warmup_proportion",
default=0.1,
type=float,
help="Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--save_checkpoints_steps",
default=1000,
type=int,
help="How often to save the model checkpoint.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--gradient_accumulation_steps',
type=int,
default=1,
help="Number of updates steps to accumualte before performing a backward/update pass.")
parser.add_argument('--optimize_on_cpu',
default=False,
action='store_true',
help="Whether to perform optimization and keep the optimizer averages on CPU")
parser.add_argument('--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=128,
help='Loss scaling, positive power of 2 values can improve fp16 convergence.')
args = parser.parse_args()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"news": NewsProcessor,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
# torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info("16-bits training currently not supported in distributed training")
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError("At least one of `do_train` or `do_eval` must be True.")
bert_config = BertConfig.from_json_file(args.bert_config_file)
if args.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length {} because the BERT model was only trained up to sequence length {}".format(
args.max_seq_length, bert_config.max_position_embeddings))
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
tokenizer = tokenization.FullTokenizer(
vocab_file=args.vocab_file, do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
label_list = processor.get_labels()
print("label_list.size:%d\n" %(len(label_list)))
# Prepare model
model = BertForSequenceClassification(bert_config, len(label_list))
if args.init_checkpoint is not None:
model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu'))
if args.fp16:
model.half()
model.to(device)
#if args.local_rank != -1:
#model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
# output_device=args.local_rank)
#elif n_gpu > 1:
# model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if n not in no_decay], 'weight_decay_rate': 0.01},
{'params': [p for n, p in param_optimizer if n in no_decay], 'weight_decay_rate': 0.0}
]
optimizer = BERTAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(
train_examples, label_list, args.max_seq_length, tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = RandomSampler(train_data)
#train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss, _ = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(param_optimizer, model.named_parameters(), test_nan=True)
if is_nan:
logger.info("FP16 TRAINING: Nan in gradients, reducing loss scaling")
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
if args.do_eval:
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
if args.local_rank == -1:
eval_sampler = SequentialSampler(eval_data)
else:
eval_sampler = SequentialSampler(eval_data)
#eval_sampler = DistributedSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
result = {'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': tr_loss/nb_tr_steps}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
