def load_model(self, model_dir: str, model_config: str = "model_config.json"):
model_config = os.path.join(model_dir,model_config)
model_config = json.load(open(model_config))
output_config_file = os.path.join(model_dir, CONFIG_NAME)
output_model_file = os.path.join(model_dir, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=model_config["num_labels"])
model.load_state_dict(torch.load(output_model_file))
tokenizer = FullTokenizer(model_file='cased_bert_base_pytorch/mn_cased.model', vocab_file='cased_bert_base_pytorch/mn_cased.vocab', do_lower_case=False)
return model, tokenizer, model_config
def load_model(self, model_dir: str, model_config: str = "model_config.json"):
model_config = os.path.join(model_dir,model_config)
model_config = json.load(open(model_config))
output_config_file = os.path.join(model_dir, CONFIG_NAME)
output_model_file = os.path.join(model_dir, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=model_config["num_labels"])
model.load_state_dict(torch.load(output_model_file))
tokenizer = BertTokenizer.from_pretrained(model_config["bert_model"],do_lower_case=False)
return model, tokenizer, model_config
def test_BertForTokenClassification():
input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])
config = BertConfig(vocab_size_or_config_json_file=32000,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072)
num_labels = 2
model = BertForTokenClassification(config, num_labels)
print(model(input_ids, token_type_ids, input_mask))
def __init__(self,
model_dir,
batch_size,
epoch,
max_seq_length=128,
local_rank=-1,
no_cuda=False):
self._batch_size = batch_size
self._local_rank = local_rank
self._max_seq_length = max_seq_length
self._device, self._n_gpu = get_device(no_cuda=no_cuda)
self._model_config = json.load(
open(os.path.join(model_dir, "model_config.json"), "r"))
self._label_to_id = self._model_config['label_map']
self._label_map = {
v: k
for k, v in self._model_config['label_map'].items()
}
self._bert_tokenizer = \
BertTokenizer.from_pretrained(model_dir,
do_lower_case=self._model_config['do_lower'])
output_config_file = os.path.join(model_dir, CONFIG_NAME)
output_model_file = os.path.join(
model_dir, "pytorch_model_ep{}.bin".format(epoch))
config = BertConfig(output_config_file)
self._model = BertForTokenClassification(config,
num_labels=len(
self._label_map))
self._model.load_state_dict(
torch.load(output_model_file,
map_location=lambda storage, loc: storage
if no_cuda else None))
self._model.to(self._device)
self._model.eval()
return
def load_model(self,
model_dir: str,
model_config: str = "model_config.json"):
model_config = os.path.join(model_dir, model_config)
model_config = json.load(open(model_config))
output_config_file = os.path.join(model_dir, CONFIG_NAME)
output_model_file = os.path.join(model_dir, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(
config, num_labels=model_config["num_labels"])
if torch.cuda.is_available() and not self.no_cuda:
model.load_state_dict(torch.load(output_model_file))
else:
model.load_state_dict(
torch.load(output_model_file, map_location='cpu'))
return model, model_config
def load_model(self,
model_dir: str,
model_config: str = "model_config.json"):
model_config = os.path.join(model_dir, model_config)
model_config = json.load(open(model_config))
output_config_file = os.path.join(model_dir, CONFIG_NAME)
output_model_file = os.path.join(model_dir, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(
config, num_labels=model_config["num_labels"])
model.load_state_dict(
torch.load(output_model_file, map_location=self.device))
if self.docker:
fn = os.path.join('/root/.pytorch_pretrained_bert', TMF)
tokenizer = BertTokenizer.from_pretrained(fn,
cache_dir=None,
do_lower_case=False)
else:
tokenizer = BertTokenizer.from_pretrained(
model_config["bert_model"], do_lower_case=False)
return model, tokenizer, model_config
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_file",
default='../../data/eng-2015.conll',
type=str,
required=True,
help="train file path")
parser.add_argument("--dev_file",
default='../../data/eng-2016.conll',
type=str,
required=True,
help="dev file path")
parser.add_argument("--bert_model", default='bert-base-cased', type=str, required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--finetune_dir",
default='NER_BERT',
type=str,
required=False,
help="The output")
parser.add_argument("--output_dir",
default='NER_BERT',
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.")
## Other parameters
parser.add_argument("--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from s3")
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",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_finetune",
action='store_true',
help="Whether to run finetuning.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=32,
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("--no_cuda",
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 accumulate before performing a backward/update pass.")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
args = parser.parse_args()
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:
torch.cuda.set_device(args.local_rank)
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')
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
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 = 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.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train:
raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case, do_basic_tokenize=False)
label_list = get_labels()
num_labels = len(label_list)
train_examples = read_ner_example(args.train_file, args.do_lower_case)
num_train_optimization_steps = None
if args.do_train:
#train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank))
model = BertForTokenClassification.from_pretrained(args.bert_model,
cache_dir=cache_dir,
num_labels = num_labels)
if args.fp16:
model.half()
if args.do_finetune:
if not os.path.exists(args.finetune_dir) and not os.listdir(args.finetune_dir):
raise ValueError("Finetune directory ({}) is empty.".format(args.finetune_dir))
finetune_model_file = os.path.join(args.finetune_dir, WEIGHTS_NAME)
finetune_config_file = os.path.join(args.finetune_dir, CONFIG_NAME)
config = BertConfig(finetune_config_file)
#model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(finetune_model_file))
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 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_optimization_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)
all_label_masks = torch.tensor([f.label_mask for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids, all_label_masks)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
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, label_masks = batch
loss = model(input_ids, segment_ids, input_mask, label_ids, label_masks)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
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:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
else:
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
#model = BertForTokenClassification.from_pretrained(args.bert_model, num_labels=num_labels)
model.to(device)
if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
eval_examples = read_ner_example(args.dev_file, args.do_lower_case)
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)
all_label_masks = torch.tensor([f.label_mask for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids, all_label_masks)
# Run prediction for full data
eval_sampler = SequentialSampler(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
pred_list = []
label_list = []
for input_ids, input_mask, segment_ids, label_ids, label_masks in tqdm(eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
label_masks = label_masks.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids, label_masks)
logits = model(input_ids, segment_ids, input_mask)
active_loss = label_masks.view(-1) == 1
active_logits = logits.view(-1, num_labels)[active_loss]
#print(active_logits.shape)
active_labels = label_ids.view(-1)[active_loss]
active_logits = active_logits.detach().cpu().numpy()
#print(active_logits.shape)
active_labels = active_labels.to('cpu').numpy()
active_preds = np.argmax(active_logits, axis=1)
#print(active_labels.shape, active_preds.shape)
#tmp_eval_accuracy = accuracy(logits, label_ids, label_masks)
#eval_loss += tmp_eval_loss.mean().item()
#eval_accuracy += tmp_eval_accuracy
pred_list.extend(active_preds)
label_list.extend(active_labels)
#print(active_labels.shape)
nb_eval_examples += active_labels.shape[0]
nb_eval_steps += 1
#eval_loss = eval_loss / nb_eval_steps
#eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss/nb_tr_steps if args.do_train else None
eval_f1_micro = f1_score(label_list, pred_list, average='micro')
eval_f1_none = f1_score(label_list, pred_list, average=None)
result = {'eval_f1_micro': eval_f1_micro,
'eval_f1_none': eval_f1_none,
'global_step': global_step,
'loss': loss}
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])))
output_pred_file = os.path.join(args.output_dir, "pred_results.conll")
label_map = get_labels()
print(len(label_list), len(pred_list))
with open(output_pred_file, 'w') as f, open(args.dev_file) as dev_f:
idx = 1
for l, p, dl in zip(label_list, pred_list, dev_f):
if len(dl) == 0:
print(dl)
f.write('\n')
idx = 1
continue
f.write(' '.join((str(idx), label_map[l], label_map[p])) + '\n')
idx += 1
return examples
def get_labels():
label_file = '../data/KBP-19/labels.txt'
with open(label_file) as f:
return [line.strip() for line in f]
return ['B-GPE', 'I-GPE', 'O', 'B-PER', 'I-PER',
'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC', 'B-TTL', 'I-TTL',
'I-FAC', 'B-FAC', 'B-VEH', 'I-VEH', 'B-WEA', 'I-WEA']
num_labels = len(get_labels())
tokenizer = BertTokenizer.from_pretrained('bert-base-cased',
do_lower_case=False, do_basic_tokenize=False)
model_dir = '../KBP_19_bert_ner_5e-5'
output_model_file = os.path.join(model_dir, WEIGHTS_NAME)
output_config_file = os.path.join(model_dir, CONFIG_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
def pred_ner(sent):
eval_examples = read_sent(sent)
label_list = get_labels()
eval_features = convert_examples_to_features(
eval_examples, label_list, 300, tokenizer)
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)
all_label_masks = torch.tensor([f.label_mask for f in eval_features], dtype=torch.long)
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(
"--log_dir",
default=None,
type=str,
required=True,
help="The log dir. Should contain the .txt file (or other data file) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written."
)
# Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from s3")
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",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
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("--no_cuda",
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 accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
# log setting
handler = logging.FileHandler(os.path.join(args.log_dir, "log.txt"))
handler.setFormatter(logging.DEBUG)
formatter = logging.Formatter(
'%(asctime)s - %(levelname)s - %(name)s - %(message)s')
handler.setFormatter(formatter)
logger.addHandler(handler)
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {
# bbn processor
"bbn": BBNNerProcessor,
}
output_modes = {
"bbn": "classification",
}
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:
torch.cuda.set_device(args.local_rank)
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')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
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 = 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.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
if task_name == 'bbn':
label_list = processor.get_labels(args.data_dir)
else:
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForTokenClassification.from_pretrained(args.bert_model,
cache_dir=cache_dir,
num_labels=num_labels)
if args.fp16:
model.half()
try:
model.to(device)
except Exception:
logger.warning("toGPU failed, failed msg:" + traceback.format_exc())
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
# prepare Data
# train_label_ids, dev_label_ids, test_label_ids = process_data(tokenizer, processor, args.data_dir, args.max_seq_length)
if args.do_train:
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_optimization_steps)
train_data = torch.load(os.path.join(args.data_dir, "train.pt"))
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in range(int(args.num_train_epochs)):
tr_loss = 0
last_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# define a new function to compute loss values for both output_modes
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.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
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:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if abs(loss.item() - last_loss) <= 5e-10:
break
# if abs(loss.item() - last_loss) != 0:
# print("iterate fine")
# print("step: " + str(step))
# print(abs(loss.item() - last_loss))
last_loss = loss.item()
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
else:
logger.info("preparing model")
# model = BertForTokenClassification.from_pretrained(
# args.bert_model, num_labels=num_labels)
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
print("Model's state_dict:")
for param_tensor in model.state_dict():
print(param_tensor, "\t", model.state_dict()[param_tensor].size())
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
model.eval()
# eval_examples = processor.get_dev_examples(args.data_dir)
eval_examples = processor.get_dev_examples(args.data_dir)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_data = torch.load(os.path.join(args.data_dir, "dev.pt"))
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
active_labels_dataset = []
i = 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():
logits, active_loss = model(input_ids, segment_ids, input_mask, labels=None)
active_labels = label_ids.view(-1)[active_loss]
# create eval loss and other metric required by the task
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
# tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
# active_labels.view(-1))
tmp_eval_loss = 0
elif output_mode == "regression":
loss_fct = MSELoss()
tmp_eval_loss = loss_fct(logits.view(-1), active_labels.view(-1))
# eval_loss += tmp_eval_loss.mean().item()
eval_loss += 0
nb_eval_steps += 1
# if len(preds) == 0:
# preds.append(logits.detach().cpu().numpy())
# else:
# preds[0] = np.append(preds[0],
# logits.detach().cpu().numpy(),
# axis=0)
logits = np.argmax(logits.detach().cpu().numpy(), axis=1)
preds.append(logits)
active_labels_dataset.append(active_labels)
eval_loss = eval_loss / nb_eval_steps
# preds = preds[0]
preds_flat = []
labels_flat = []
for s in preds:
for l in s: # l is label
preds_flat.append(l)
for s in active_labels_dataset:
for l in s:
labels_flat.append(l.detach().cpu().numpy())
preds_flat = np.array(preds_flat)
labels_flat = np.array(labels_flat)
for i in range(len(preds_flat)):
if preds_flat[i] == 37:
preds_flat[i] = 7
elif preds_flat[i] == 34:
preds_flat[i] == 12
elif preds_flat[i] == 26:
preds_flat[i] = 36
elif preds_flat[i] == 36:
preds_flat[i] = 37
elif preds_flat[i] == 41:
preds_flat[i] = 34
elif preds_flat[i] == 31:
preds_flat[i] = 39
elif preds_flat[i] == 15:
preds_flat[i] = 38
# label_map = dict()
# for i in range(len(preds_flat)):
# key = str(preds_flat[i]) + '-' + str(labels_flat[i])
# if key in label_map.keys():
# label_map[key] += 1
# else:
# label_map[key] = 0
# for k in label_map.keys():
# if label_map[k] > 1000:
# print(k, ":", label_map[k])
# if output_mode == "classification":
# preds = np.argmax(preds, axis=2)
# elif output_mode == "regression":
# preds = np.squeeze(preds)
result = compute_metrics(task_name, preds_flat, labels_flat)
loss = tr_loss / nb_tr_steps if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
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])))
# hack for MNLI-MM
if task_name == "mnli":
task_name = "mnli-mm"
processor = processors[task_name]()
if os.path.exists(args.output_dir +
'-MM') and os.listdir(args.output_dir +
'-MM') and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".
format(args.output_dir))
if not os.path.exists(args.output_dir + '-MM'):
os.makedirs(args.output_dir + '-MM')
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer,
output_mode)[0]
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)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
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():
logits = model(input_ids,
segment_ids,
input_mask,
labels=None)
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
preds = np.argmax(preds, axis=1)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / nb_tr_steps if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir + '-MM',
"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])))
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_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written."
)
# Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from s3")
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",
action='store_true',
help="Whether to run training.")
parser.add_argument(
"--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
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(
"--no_cuda",
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 accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument(
'--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument(
'--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {"ner": NerProcessor}
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:
torch.cuda.set_device(args.local_rank)
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')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
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 = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(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.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.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) + 1
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
print("num_train_optimization_steps: ", num_train_optimization_steps)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForTokenClassification.from_pretrained(args.bert_model,
cache_dir=cache_dir,
num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 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_optimization_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 = 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.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
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:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
label_map = {i: label for i, label in enumerate(label_list, 1)}
model_config = {
"bert_model": args.bert_model,
"do_lower": args.do_lower_case,
"max_seq_length": args.max_seq_length,
"num_labels": len(label_list) + 1,
"label_map": label_map
}
json.dump(
model_config,
open(os.path.join(args.output_dir, "model_config.json"), "w"))
# Load a trained model and config that you have fine-tuned
else:
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
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)
# Run prediction for full data
eval_sampler = SequentialSampler(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
y_true = []
y_pred = []
label_map = {i: label for i, label in enumerate(label_list, 1)}
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
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():
logits = model(input_ids, segment_ids, input_mask)
logits = torch.argmax(F.log_softmax(logits, dim=2), dim=2)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
input_mask = input_mask.to('cpu').numpy()
for i, mask in enumerate(input_mask):
temp_1 = []
temp_2 = []
for j, m in enumerate(mask):
if j == 0:
continue
if m:
if label_map[label_ids[i][j]] != "X":
temp_1.append(label_map[label_ids[i][j]])
temp_2.append(label_map[logits[i][j]])
else:
temp_1.pop()
temp_2.pop()
break
y_true.append(temp_1)
y_pred.append(temp_2)
report = classification_report(y_true, y_pred, digits=4)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
logger.info("\n%s", report)
writer.write(report)
def predict(OUTPUT_DIR, in_sentences):
""" predict a bert model
OUTPUT_DIR :: contains pretrained models
in_sentences :: is a list of sentences on which tagging has to be performed
"""
PRED_BATCH_SIZE = 64
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_config = os.path.join(OUTPUT_DIR, "model_config.json")
model_config = json.load(open(model_config))
output_config_file = os.path.join(OUTPUT_DIR, CONFIG_NAME)
output_model_file = os.path.join(OUTPUT_DIR, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config,
num_labels=model_config["num_labels"])
model.load_state_dict(torch.load(output_model_file))
model.to(device)
tokenizer = BertTokenizer.from_pretrained(
model_config["bert_model"], do_lower_case=model_config["do_lower"])
in_examples = [
InputExample(guid="",
text_a=x,
text_b=None,
label=["O"] * len(x.split(" "))) for x in in_sentences
]
in_features = convert_examples_to_features(in_examples, label_list,
MAX_SEQ_LENGTH, tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in in_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in in_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in in_features],
dtype=torch.long)
pred_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids)
# # Run prediction for full data
pred_sampler = SequentialSampler(pred_data)
pred_dataloader = DataLoader(pred_data,
sampler=pred_sampler,
batch_size=PRED_BATCH_SIZE,
drop_last=False)
model.eval()
preds = []
label_map = model_config["label_map"]
for input_ids, input_mask, segment_ids in tqdm(pred_dataloader,
desc="Predicting"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask)
logits = torch.argmax(F.log_softmax(logits, dim=2), dim=2)
logits = logits.detach().cpu().numpy()
pred_batch = []
for i, mask in enumerate(input_mask):
temp_1 = []
for j, m in enumerate(mask):
if j == 0:
continue
if m:
if label_map[str(logits[i][j])] != "X":
temp_1.append(label_map[str(logits[i][j])])
else:
temp_1.pop()
break
pred_batch.append(temp_1)
preds.extend(pred_batch)
return [(sentence, pred) for sentence, pred in zip(in_sentences, preds)]
def train_and_evaluate(OUTPUT_DIR, do_train=True, do_eval=True):
""" Train and evaluate a BERT NER Model"""
BATCH_SIZE = 32
LEARNING_RATE = 2e-5
NUM_TRAIN_EPOCHS = 5.0
#in this steps lr will be low and training will be slow
WARMUP_PROPORTION = 0.1
if os.path.exists(OUTPUT_DIR) and os.listdir(OUTPUT_DIR) and do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
OUTPUT_DIR))
if not os.path.exists(OUTPUT_DIR):
os.makedirs(OUTPUT_DIR)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased",
do_lower_case=True)
if do_train:
train_examples, num_train_examples = create_datasets("AGE/train.txt")
num_train_steps = int(
math.ceil(num_train_examples / BATCH_SIZE * NUM_TRAIN_EPOCHS))
num_warmup_steps = int(num_train_steps * WARMUP_PROPORTION)
model = BertForTokenClassification.from_pretrained(
"bert-base-uncased", num_labels=num_labels)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=LEARNING_RATE,
warmup=WARMUP_PROPORTION,
t_total=num_train_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
train_features = convert_examples_to_features(train_examples,
label_list,
MAX_SEQ_LENGTH,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", num_train_examples)
logger.info(" Batch size = %d", 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)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=BATCH_SIZE)
model.train()
# for name, param in model.named_parameters():
# if param.requires_grad:
# print(name)
# return
for _ in trange(int(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)
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
optimizer.step()
optimizer.zero_grad()
global_step += 1
print(tr_loss)
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(OUTPUT_DIR, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(OUTPUT_DIR, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
label_map = {i: label for i, label in enumerate(label_list, 1)}
model_config = {
"bert_model": "bert-base-uncased",
"do_lower": True,
"max_seq_length": MAX_SEQ_LENGTH,
"num_labels": len(label_list) + 1,
"label_map": label_map
}
json.dump(model_config,
open(os.path.join(OUTPUT_DIR, "model_config.json"), "w"))
else:
output_config_file = os.path.join(OUTPUT_DIR, CONFIG_NAME)
output_model_file = os.path.join(OUTPUT_DIR, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
model.to(device)
if do_eval:
EVAL_BATCH_SIZE = 32
eval_examples, num_eval_examples = create_datasets("AGE/valid.txt")
eval_features = convert_examples_to_features(eval_examples, label_list,
MAX_SEQ_LENGTH, tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", num_eval_examples)
logger.info(" Batch size = %d", 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)
# # Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=EVAL_BATCH_SIZE)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
y_true = []
y_pred = []
label_map = {i: label for i, label in enumerate(label_list, 1)}
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
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():
logits = model(input_ids, segment_ids, input_mask)
logits = torch.argmax(F.log_softmax(logits, dim=2), dim=2)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
input_mask = input_mask.to('cpu').numpy()
for i, mask in enumerate(input_mask):
temp_1 = []
temp_2 = []
for j, m in enumerate(mask):
if j == 0:
continue
if m:
if label_map[label_ids[i][j]] != "X":
temp_1.append(label_map[label_ids[i][j]])
temp_2.append(label_map[logits[i][j]])
else:
temp_1.pop()
temp_2.pop()
break
y_true.append(temp_1)
y_pred.append(temp_2)
report = classification_report(y_true, y_pred)
output_eval_file = os.path.join(OUTPUT_DIR, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
logger.info("\n%s", report)
writer.write(report)
def model_eval(batch_size,
label_map,
processor,
device,
num_train_epochs=1,
output_dir=None,
model=None,
local_rank=-1,
no_cuda=False,
dry_run=False):
output_eval_file = None
if output_dir is not None:
output_eval_file = os.path.join(output_dir,
processor.get_evaluation_file())
logger.info('Write evaluation results to: {}'.format(output_eval_file))
dataloader = processor.get_dev_examples(batch_size, local_rank)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(dataloader))
logger.info(" Batch size = %d", batch_size)
results = list()
output_config_file = None
if output_dir is not None:
output_config_file = os.path.join(output_dir, CONFIG_NAME)
for ep in trange(1, int(num_train_epochs) + 1, desc="Epoch"):
if dry_run and ep > 1:
logger.info("Dry run. Stop.")
break
if output_config_file is not None:
# Load a trained model and config that you have fine-tuned
output_model_file = os.path.join(
output_dir, "pytorch_model_ep{}.bin".format(ep))
if not os.path.exists(output_model_file):
logger.info(
"Stopping at epoch {} since model file is missing.".format(
ep))
break
config = BertConfig(output_config_file)
model = BertForTokenClassification(config,
num_labels=len(label_map))
model.load_state_dict(
torch.load(output_model_file,
map_location=lambda storage, loc: storage
if no_cuda else None))
model.to(device)
if model is None:
raise ValueError('Model required for evaluation.')
model.eval()
y_pred, y_true = model_predict_compare(dataloader, device, label_map,
model, dry_run)
lines = [
'empty ' + 'XXX ' + v + ' ' + p for yt, yp in zip(y_true, y_pred)
for v, p in zip(yt, yp)
]
res = conll_eval(lines)
# print(res)
evals = \
pd.concat([pd.DataFrame.from_dict(res['overall']['evals'], orient='index', columns=['ALL']),
pd.DataFrame.from_dict(res['slots']['LOC']['evals'], orient='index', columns=['LOC']),
pd.DataFrame.from_dict(res['slots']['PER']['evals'], orient='index', columns=['PER']),
pd.DataFrame.from_dict(res['slots']['ORG']['evals'], orient='index', columns=['ORG']),
], axis=1).T
stats = \
pd.concat(
[pd.DataFrame.from_dict(res['overall']['stats'], orient='index', columns=['ALL']),
pd.DataFrame.from_dict(res['slots']['LOC']['stats'], orient='index', columns=['LOC']),
pd.DataFrame.from_dict(res['slots']['PER']['stats'], orient='index', columns=['PER']),
pd.DataFrame.from_dict(res['slots']['ORG']['stats'], orient='index', columns=['ORG'])],
axis=1, sort=True).T
evals['epoch'] = ep
stats['epoch'] = ep
results.append(
pd.concat([
evals.reset_index().set_index(['index', 'epoch']),
stats.reset_index().set_index(['index', 'epoch'])
],
axis=1))
if output_eval_file is not None:
pd.concat(results).to_pickle(output_eval_file)
results = pd.concat(results)
print(results)
return results
def main():
parser = get_parser()
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {"ner": NerProcessor}
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:
torch.cuda.set_device(args.local_rank)
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')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
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 = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(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.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.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) + 1
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForTokenClassification.from_pretrained(args.bert_model,
cache_dir=cache_dir,
num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 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_optimization_steps)
all_input_ids = torch.tensor(
[f.input_ids for f in train_features],
dtype=torch.long) # words converted to ids by tokenizer
all_input_mask = torch.tensor(
[f.input_mask for f in train_features],
dtype=torch.long) # mask keeps everything
all_segment_ids = torch.tensor(
[f.segment_ids for f in train_features],
dtype=torch.long) # segment ids are all 0
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
# labels are ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC", "X", "[CLS]", "[SEP]"]
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 = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train(
) # this just sets the model to be in training mode (is method on pytorch nn.Module)
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.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
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:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
label_map = {i: label for i, label in enumerate(label_list, 1)}
model_config = {
"bert_model": args.bert_model,
"do_lower": args.do_lower_case,
"max_seq_length": args.max_seq_length,
"num_labels": len(label_list) + 1,
"label_map": label_map
}
json.dump(
model_config,
open(os.path.join(args.output_dir, "model_config.json"), "w"))
# Load a trained model and config that you have fine-tuned
else:
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
config = BertConfig(output_config_file)
model = BertForTokenClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
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)
# Run prediction for full data
eval_sampler = SequentialSampler(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
y_true = []
y_pred = []
label_map = {i: label for i, label in enumerate(label_list, 1)}
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
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():
logits = model(input_ids, segment_ids, input_mask)
logits = torch.argmax(F.log_softmax(logits, dim=2), dim=2)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
input_mask = input_mask.to('cpu').numpy()
for i, mask in enumerate(input_mask):
temp_1 = []
temp_2 = []
for j, m in enumerate(mask):
if j == 0:
continue
if m:
if label_map[label_ids[i][j]] != "X":
temp_1.append(label_map[label_ids[i][j]])
temp_2.append(label_map[logits[i][j]])
else:
temp_1.pop()
temp_2.pop()
break
y_true.append(temp_1)
y_pred.append(temp_2)
report = classification_report(y_true, y_pred, digits=4)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
logger.info("\n%s", report)
writer.write(report)