def __init__(self, model, config, **kwargs):
super().__init__(kwargs['dataset'],
model,
kwargs['embedding'],
kwargs['data_loader'],
batch_size=config['batch_size'],
device=config['device'])
if config['model'] in {
'BERT-Base', 'BERT-Large', 'HBERT-Base', 'HBERT-Large'
}:
variant = 'bert-large-uncased' if config[
'model'] == 'BERT-Large' else 'bert-base-uncased'
self.tokenizer = BertTokenizer.from_pretrained(
variant, is_lowercase=config['is_lowercase'])
self.processor = kwargs['processor']
if config['split'] == 'test':
self.eval_examples = self.processor.get_test_examples(
config['data_dir'], topic=config['topic'])
else:
self.eval_examples = self.processor.get_dev_examples(
config['data_dir'], topic=config['topic'])
self.config = config
self.ignore_lengths = config['ignore_lengths']
self.y_target = None
self.y_pred = None
self.docid = None
def __init__(self, model, config, **kwargs):
super().__init__(model, kwargs['embedding'], kwargs['train_loader'],
config, None, kwargs['test_evaluator'],
kwargs['dev_evaluator'])
if config['model'] in {
'BERT-Base', 'BERT-Large', 'HBERT-Base', 'HBERT-Large'
}:
variant = 'bert-large-uncased' if config[
'model'] == 'BERT-Large' else 'bert-base-uncased'
self.tokenizer = BertTokenizer.from_pretrained(
variant, is_lowercase=config['is_lowercase'])
self.processor = kwargs['processor']
self.optimizer = config['optimizer']
self.train_examples = self.processor.get_train_examples(
config['data_dir'], topic=config['topic'])
self.num_train_optimization_steps = int(
len(self.train_examples) / config['batch_size'] /
config['gradient_accumulation_steps']) * config['epochs']
self.config = config
self.early_stop = False
self.best_dev_ap = 0
self.iterations = 0
self.unimproved_iters = 0
self.log_header = 'Epoch Iteration Progress Dev/Acc. Dev/Pr. Dev/AP. Dev/F1 Dev/Loss'
self.log_template = ' '.join(
'{:>5.0f},{:>9.0f},{:>6.0f}/{:<5.0f} {:>6.4f},{:>8.4f},{:8.4f},{:8.4f},{:10.4f}'
.split(','))
timestamp = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
self.snapshot_path = os.path.join(self.model_outfile,
config['dataset'].NAME,
'%s.pt' % timestamp)
def __init__(self, model, optimizer, processor, args):
self.args = args
self.model = model
self.optimizer = optimizer
self.processor = processor
self.train_examples = self.processor.get_train_examples(args.data_dir)
self.tokenizer = BertTokenizer.from_pretrained(
args.model, is_lowercase=args.is_lowercase)
timestamp = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
self.snapshot_path = os.path.join(self.args.save_path,
self.processor.NAME,
'%s.pt' % timestamp)
self.num_train_optimization_steps = int(
len(self.train_examples) / args.batch_size /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
self.num_train_optimization_steps = args.num_train_optimization_steps // torch.distributed.get_world_size(
)
self.log_header = 'Epoch Iteration Progress Dev/Acc. Dev/Hamm. Dev/Jacc. Dev/Prec Dev/Rec Dev/micro-F1 Dev/F1 Dev/Loss'
self.log_template = ' '.join(
'{:>5.0f},{:>9.0f},{:>6.0f}/{:<5.0f} {:>6.4f},{:>8.4f},{:8.4f},{:8.4f},{:>8.4f},{:8.4f},{:8.4f},{:10.4f}'
.split(','))
self.iterations, self.nb_tr_steps, self.tr_loss = 0, 0, 0
self.best_dev_f1, self.unimproved_iters = 0, 0
self.early_stop = False
def __init__(self, model, processor, args, split='dev'):
self.args = args
self.model = model
self.processor = processor
self.tokenizer = BertTokenizer.from_pretrained(args.model, is_lowercase=args.is_lowercase)
if split == 'test':
self.eval_examples = self.processor.get_test_examples(args.data_dir)
else:
self.eval_examples = self.processor.get_dev_examples(args.data_dir)
def load_text(texts: List[str]):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
max_pos = 512
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased',
do_lower_case=True)
sep_vid = tokenizer.vocab['[SEP]']
cls_vid = tokenizer.vocab['[CLS]']
n_lines = len(texts)
def _process_src(raw):
raw = raw.strip().lower()
src_subtokens = tokenizer.tokenize(raw)
src_subtokens = ['[CLS]'] + src_subtokens + ['[SEP]']
src_subtoken_idxs = tokenizer.convert_tokens_to_ids(src_subtokens)
src_subtoken_idxs = src_subtoken_idxs[:-1][:max_pos]
src_subtoken_idxs[-1] = sep_vid
_segs = [-1] + \
[i for i, t in enumerate(src_subtoken_idxs) if t == sep_vid]
segs = [_segs[i] - _segs[i - 1] for i in range(1, len(_segs))]
segments_ids = []
segs = segs[:max_pos]
for i, s in enumerate(segs):
if (i % 2 == 0):
segments_ids += s * [0]
else:
segments_ids += s * [1]
src = torch.tensor(src_subtoken_idxs)[None, :].to(device)
mask_src = (1 - (src == 0).float()).to(device)
cls_ids = [[
i for i, t in enumerate(src_subtoken_idxs) if t == cls_vid
]]
clss = torch.tensor(cls_ids).to(device)
mask_cls = 1 - (clss == -1).float()
clss[clss == -1] = 0
return src, mask_src, segments_ids, clss, mask_cls
for x in tqdm(texts, total=n_lines):
src, mask_src, segments_ids, clss, mask_cls = _process_src(x)
segs = torch.tensor(segments_ids)[None, :].to(device)
batch = Batch()
batch.src = src
batch.tgt = None
batch.mask_src = mask_src
batch.mask_tgt = None
batch.segs = segs
batch.src_str = [[
sent.replace('[SEP]', '').strip() for sent in x.split('[CLS]')
]]
batch.tgt_str = ['']
batch.clss = clss
batch.mask_cls = mask_cls
batch.batch_size = 1
yield batch
def __init__(self, model, processor, args, split='dev'):
self.args = args
self.model = model
self.processor = processor
self.tokenizer = BertTokenizer.from_pretrained(args.model, is_lowercase=args.is_lowercase)
self.emotioner = Emotion(args.nrc_path, args.max_em_len, args.emotion_filters)
if split == 'test':
self.eval_examples = self.processor.get_test_examples(args.data_dir, args.test_name)
elif split == 'dev':
self.eval_examples = self.processor.get_dev_examples(args.data_dir, args.dev_name)
else:
self.eval_examples = self.processor.get_any_examples(args.data_dir, split)
def pretrain_model():
print("======MLM & NSP Pretraining======")
""" MLM & NSP Pretraining
You can modify this function by yourself.
This function does not affects your final score.
"""
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
comment_train_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'unlabeled'))
title_train_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'news_title'))
train_dataset = PretrainDataset(args.max_seq_length, comment_train_dataset,
title_train_dataset, tokenizer)
comment_val_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'unlabeled', 'dev'))
title_val_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'news_title', 'dev'))
val_dataset = PretrainDataset(args.max_seq_length, comment_val_dataset,
title_val_dataset, tokenizer)
config = Config(args.config_file)
device = torch.device('cpu')
model = MlmNspModel(config).to(torch.device(device))
model.bert.load_state_dict(torch.load(args.checkpoint,
map_location=device))
model_name = 'pretrained'
MLM_train_losses, MLM_val_losses, NSP_train_losses, NSP_val_losses \
= pretraining(model, model_name, train_dataset, val_dataset)
torch.save(model.state_dict(), model_name + '_final.pth')
with open(model_name + '_result.pkl', 'wb') as f:
pickle.dump((MLM_train_losses, MLM_val_losses, NSP_train_losses,
NSP_val_losses), f)
utils_pretrain.plot_values(MLM_train_losses,
MLM_val_losses,
title=model_name + "_mlm")
utils_pretrain.plot_values(NSP_train_losses,
NSP_val_losses,
title=model_name + "_nsp")
print("Final MLM training loss: {:06.4f}".format(MLM_train_losses[-1]))
print("Final MLM validation loss: {:06.4f}".format(MLM_val_losses[-1]))
print("Final NSP training loss: {:06.4f}".format(NSP_train_losses[-1]))
print("Final NSP validation loss: {:06.4f}".format(NSP_val_losses[-1]))
def decode(hp):
tokenizer = BertTokenizer.from_pretrained(hp.bert_model_dir, do_lower_case=False)
if hp.dataset == "lm_raw_data_finance":
dict_file = "../data/dataset_finance/annual_report_entity_list"
elif hp.dataset == "lm_raw_data_novel":
dict_file = "../data/dataset_book9/entity_book9"
entity_dict = EntityDict(hp.dataset, dict_file)
entity_dict.load(os.path.join(hp.bert_model_dir, 'entity.dict'))
device = 'cuda' if torch.cuda.is_available() else 'cpu'
ner_label = NerLabel([hp.decodeset])
if os.path.exists(os.path.join(hp.logdir, 'dict.pt')):
ner_label.load(os.path.join(hp.logdir, 'dict.pt'))
else:
print('dict.pt is not exit')
exit()
decode_dataset = NerDataset(hp.decodeset, ner_label, tokenizer, entity_dict)
model = Net(hp.bert_model_dir, hp.top_rnns, len(ner_label.VOCAB), entity_dict.entity_num, device, hp.finetuning).to(device)
model = nn.DataParallel(model)
## Load the model parameters
if os.path.exists(os.path.join(hp.logdir, 'model.pt')):
model.load_state_dict(torch.load(os.path.join(hp.logdir, 'model.pt')))
else:
print("the pretrianed model path does not exist! ")
exit()
decode_iter = data.DataLoader(dataset=decode_dataset,
batch_size=hp.batch_size,
shuffle=True,
num_workers=4,
collate_fn=pad)
fname = os.path.join(hp.logdir, '_')
precision, recall, f1 = evaluate(model, decode_iter, fname, ner_label, verbose=False)
def main():
parser = argparse.ArgumentParser()
# Parameters
parser.add_argument(
"--output_dir",
default='output',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--checkpoint",
default='large_v1/large_v1_model.bin',
type=str,
help="pre-trained model checkpoint")
parser.add_argument("--config_file",
default='large_v1/large_config.json',
type=str,
help="model configuration file")
parser.add_argument("--vocab_file",
default='large_v1/large_v1_32k_vocab.txt',
type=str,
help="tokenizer vocab file")
parser.add_argument("--train_file",
default='data/korquad/KorQuAD_v1.0_train.json',
type=str,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument("--test",
default=False,
type=bool,
help="Test by sample")
parser.add_argument("--train_data_limit",
default=-1,
type=int,
help="For Check EDA performance limit train data")
######ADD for EDA#######
parser.add_argument(
"--same_random_index",
default=True,
type=bool,
help="for eda data limit, if True every eda pick same random index")
parser.add_argument("--eda_type",
default="no_eda",
type=str,
help="eda type e.g) sr, ri, rs, rd")
parser.add_argument("--alpha", default=0.1, type=float, help="For eda")
parser.add_argument("--num_aug", default=2, type=int, help="For eda")
parser.add_argument("--eda_model_name",
default="word2vec",
type=str,
help="For eda")
parser.add_argument("--min_score", default=0.7, type=float, help="For eda")
parser.add_argument("--dict_file_name",
default="data/syn/synonyms_file.pickle",
type=str)
########################
parser.add_argument(
"--max_seq_length",
default=512,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help=
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument("--per_gpu_train_batch_size",
default=4,
type=int,
help="Total batch size for training.")
parser.add_argument(
'--gradient_accumulation_steps',
default=1,
type=int,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="Weight deay if we apply some.")
parser.add_argument("--num_train_epochs",
default=4.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--adam_epsilon",
default=1e-6,
type=float,
help="Epsilon for Adam optimizer.")
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(
"--n_best_size",
default=20,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions.json "
"output file.")
parser.add_argument(
"--verbose_logging",
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
default=42,
type=int,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--fp16_opt_level',
default='O2',
type=str,
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument("--local_rank",
default=-1,
type=int,
help="For distributed training: local_rank")
parser.add_argument(
'--null_score_diff_threshold',
type=float,
default=0.0,
help=
"If null_score - best_non_null is greater than the threshold predict null."
)
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
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")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(args.vocab_file,
max_len=args.max_seq_length,
do_basic_tokenize=True)
# Prepare model
config = Config.from_json_file(args.config_file)
model = QuestionAnswering(config)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
logger.info("Training hyper-parameters %s", args)
# Training
train_dataset = load_and_cache_examples(args, tokenizer)
train(args, train_dataset, model)
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--config_file",
type=str,
required=True,
help="model configuration file")
parser.add_argument("--vocab_file",
type=str,
required=True,
help="tokenizer vocab file")
parser.add_argument("--checkpoint",
default=None,
type=str,
required=True,
help="fine-tuned model checkpoint")
# Other parameters
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("--eval_batch_size",
default=32,
type=int,
help="Total batch size for eval.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
args.n_gpu = n_gpu
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
level = logging.INFO)
logger.info("device: {} n_gpu: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
set_seed(args)
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]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
test_examples = processor.get_test_examples(args.data_dir)
texts = [test_example.text_a for test_example in test_examples]
# Prepare model
config = Config(args.config_file)
model = SequenceClassificationMultitask(config)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
if args.fp16:
model = model.half()
test_features = convert_examples_to_features(
test_examples, label_list, args.max_seq_length, tokenizer, output_mode)
logger.info("***** Running Evaluation *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in test_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in test_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features], dtype=torch.long)
if output_mode == "classification_multitask":
all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.float)
test_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=args.eval_batch_size)
model.eval()
preds = []
preds0 = []
preds1 = []
preds2 = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(test_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, labels=None)
if output_mode == "classification_multitask":
if len(preds0) == 0:
preds0.append(logits[0].detach().cpu().numpy())
preds1.append(logits[1].detach().cpu().numpy())
preds2.append(logits[2].detach().cpu().numpy())
else:
preds0[0] = np.append(preds0[0], logits[0].detach().cpu().numpy(), axis=0)
preds1[0] = np.append(preds1[0], logits[1].detach().cpu().numpy(), axis=0)
preds2[0] = np.append(preds2[0], logits[2].detach().cpu().numpy(), axis=0)
else:
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(
preds[0], logits.detach().cpu().numpy(), axis=0)
if output_mode == "classification_multitask":
preds0 = preds0[0]
preds1 = preds1[0]
preds2 = preds2[0]
preds0 = np.argmax(preds0, axis=1)
preds1 = np.argmax(preds1, axis=1)
preds2 = np.argmax(preds2, axis=1)
preds = [preds0, preds1, preds2]
label=[]
for p in preds0:
if p==0:
label.append('True')
elif p==1:
label.append('False')
elif output_mode == "classification":
preds = preds[0]
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = preds[0]
preds = np.squeeze(preds)
#TODO: write the csv
#result = compute_metrics(task_name, preds, all_label_ids.numpy())
data = {'comments':texts,
'label':label}
data_df = DataFrame(data)
csv_write_path='/mnt/sdc1/korLM/by_me/out/by_me_9_gender_0608.csv'
# data_df.to_csv(csv_write_path, sep=',',index=False)
with open(csv_write_path, 'w', encoding='utf-8') as f:
wr = csv.writer(f)
wr.writerow(['comments','label'])
for text, lab in zip(texts, label):
wr.writerow([text, lab])
logger.info("***** Eval results *****")
def do_main():
# Set default configuration in args.py
args = get_args()
if args.local_rank == -1 or not args.cuda:
device = torch.device(
"cuda" if torch.cuda.is_available() and args.cuda else "cpu")
n_gpu = torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
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')
print('Device:', str(device).upper())
print('Number of GPUs:', n_gpu)
print('Distributed training:', bool(args.local_rank != -1))
print('FP16:', args.fp16)
# Set random seed for reproducibility
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)
dataset_map = {
'SST-2': SST2Processor,
'Reuters': ReutersProcessor,
'IMDB': IMDBProcessor,
'AAPD': AAPDProcessor,
'AGNews': AGNewsProcessor,
'Yelp2014': Yelp2014Processor,
'Sogou': SogouProcessor,
'Personality': PersonalityProcessor,
'News_art': News_artProcessor,
'News': News_Processor,
'UCI_yelp': UCI_yelpProcessor,
'Procon': ProconProcessor,
'Style': StyleProcessor,
'ProconDual': ProconDualProcessor,
'Pan15': Pan15_Processor,
'Pan14E': Pan14E_Processor,
'Pan14N': Pan14N_Processor,
'Perspectrum': PerspectrumProcessor
}
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.dataset not in dataset_map:
raise ValueError('Unrecognized dataset')
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path,
dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = False
tokenizer = BertTokenizer.from_pretrained(args.model,
is_lowercase=args.is_lowercase)
train_examples = None
num_train_optimization_steps = None
if not args.trained_model:
train_examples = processor.get_train_examples(args.data_dir,
args.train_name)
num_train_optimization_steps = int(
len(train_examples) / args.batch_size /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.model, cache_dir=cache_dir, num_labels=args.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(
"Install NVIDIA Apex to use distributed and FP16 training.")
model = DDP(model)
'''elif n_gpu > 1: changed by marjan
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 NVIDIA Apex for distributed and FP16 training")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.lr,
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.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
trainer = BertTrainer(model, optimizer, processor, args)
if not args.trained_model:
trainer.train()
model = torch.load(trainer.snapshot_path)
else:
model = BertForSequenceClassification.from_pretrained(
args.model, num_labels=args.num_labels)
model_ = torch.load(args.trained_model,
map_location=lambda storage, loc: storage)
state = {}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
model.load_state_dict(state)
model = model.to(device)
evaluate_split(model, processor, args, split=args.dev_name)
evaluate_split(model, processor, args, split=args.test_name)
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path,
dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = True
tokenizer = BertTokenizer.from_pretrained(args.model,
is_lowercase=args.is_lowercase)
train_examples = None
num_train_optimization_steps = None
if not args.trained_model:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.batch_size /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--output_dir", default='output', type=str,
help="The output directory where the model checkpoints and predictions will be written.")
parser.add_argument("--checkpoint", default='pretrain_ckpt/bert_small_ckpt.bin',
type=str,
help="checkpoint")
parser.add_argument("--model_config", default='data/bert_small.json',
type=str)
# Other parameters
parser.add_argument("--train_file", default='data/KorQuAD_v1.0_train.json', type=str,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument("--max_seq_length", default=512, type=int,
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded.")
parser.add_argument("--doc_stride", default=128, type=int,
help="When splitting up a long document into chunks, how much stride to take between chunks.")
parser.add_argument("--max_query_length", default=96, type=int,
help="The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument("--train_batch_size", default=16, type=int, help="Total batch size for training.")
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs", default=4.0, type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--max_grad_norm", default=1.0, type=float,
help="Max gradient norm.")
parser.add_argument("--adam_epsilon", default=1e-6, type=float,
help="Epsilon for Adam optimizer.")
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("--n_best_size", default=20, type=int,
help="The total number of n-best predictions to generate in the nbest_predictions.json "
"output file.")
parser.add_argument("--max_answer_length", default=30, type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument("--verbose_logging", action='store_true',
help="If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--fp16_opt_level', type=str, default='O2',
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument('--null_score_diff_threshold',
type=float, default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.")
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {} n_gpu: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
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 os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer('ko_vocab_32k.txt',
max_len=args.max_seq_length,
do_basic_tokenize=True)
# Prepare model
config = Config.from_json_file(args.model_config)
model = QuestionAnswering(config)
model.bert.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
cached_train_features_file = args.train_file + '_{0}_{1}_{2}'.format(str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length))
train_examples = read_squad_examples(input_file=args.train_file, is_training=True, version_2_with_negative=False)
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
logger.info(" Saving train features into cached file %s", cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
num_train_optimization_steps = int(len(train_features) / args.train_batch_size) * args.num_train_epochs
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['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 = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = WarmupLinearSchedule(optimizer,
warmup_steps=num_train_optimization_steps*0.1,
t_total=num_train_optimization_steps)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
num_train_step = 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_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_start_positions, all_end_positions)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
global_step = 0
epoch = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
iter_bar = tqdm(train_dataloader, desc="Train(XX Epoch) Step(XX/XX) (Mean loss=X.X) (loss=X.X)")
tr_step, total_loss, mean_loss = 0, 0., 0.
for step, batch in enumerate(iter_bar):
if n_gpu == 1:
batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch
loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
tr_step += 1
total_loss += loss
mean_loss = total_loss / tr_step
iter_bar.set_description("Train Step(%d / %d) (Mean loss=%5.5f) (loss=%5.5f)" %
(global_step, num_train_step, mean_loss, loss.item()))
logger.info("** ** * Saving file * ** **")
model_checkpoint = "korquad_%d.bin" % (epoch)
logger.info(model_checkpoint)
output_model_file = os.path.join(args.output_dir,model_checkpoint)
if n_gpu > 1:
torch.save(model.module.state_dict(), output_model_file)
else:
torch.save(model.state_dict(), output_model_file)
epoch += 1
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_corpus",
default=None,
type=str,
required=True,
help="The input train corpus.")
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-base-multilingual, bert-base-chinese.")
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("--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("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--learning_rate",
default=3e-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("--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
parser.add_argument("--do_lower_case",
action='store_true',
help="Whether to lower case the input text. True for uncased models, False for cased models.")
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('--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")
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:
raise ValueError("Training is currently the only implemented execution option. Please set `do_train`.")
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))
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)
#train_examples = None
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_corpus)
train_dataset = BERTDataset(args.train_corpus, tokenizer, seq_len=args.max_seq_length,
corpus_lines=None, on_memory=args.on_memory)
num_train_optimization_steps = int(
len(train_dataset) / 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
model = BertForPreTraining.from_pretrained(args.bert_model)
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)
# Prepare optimizer
if args.do_train:
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)
warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on next(file)
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, 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, lm_label_ids, is_next = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next)
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.get_lr(global_step, 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
logger.info("** ** * Saving fine - tuned model ** ** * ")
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)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
from sklearn.svm import SVC
from sklearn import metrics
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.manifold import TSNE
import math
import matplotlib.pyplot as plt
from sklearn.decomposition import PCA
from sklearn.feature_selection import SelectPercentile, chi2
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
# String templates for logging results
LOG_HEADER = 'Split Dev/Acc. Dev/Pr. Dev/Re. Dev/F1'
LOG_TEMPLATE = ' '.join('{:>5s},{:>9.4f},{:>8.4f},{:8.4f},{:8.4f}'.split(','))
bert_tokenizer = BertTokenizer.from_pretrained('bert-base-uncased',
is_lowercase=True)
def get_feature_vector(evaluators,
use_idf=False,
ngram_range=(1, 1),
max_seq_len=256):
train_ev, dev_ev, test_ev = evaluators
train_text, test_text, dev_text = [], [], []
for i, x in enumerate(train_ev.eval_examples):
#tokens_a = x.text_a.strip().split()
tokens_a = [t for t in x.text_a.strip().split() if t not in ['', ' ']]
#tokens_a = bert_tokenizer.tokenize(x.text_a)
tokens_a = tokens_a[:max_seq_len]
train_text.append(' '.join(tokens_a))
def main(input, output):
max_seq_length = 512
doc_stride = 64
max_query_length = 64
batch_size = 16
n_best_size = 20
max_answer_length = 30
seed = 42
fp16 = False
device = torch.device("cpu")
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {} n_gpu: {}".format(device, n_gpu))
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(seed)
tokenizer = BertTokenizer(vocab_file=VOCAB_PATH,
max_len=max_seq_length,
do_basic_tokenize=True)
eval_examples = read_squad_examples(input_file=input,
is_training=False,
version_2_with_negative=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
doc_stride=doc_stride,
max_query_length=max_query_length,
is_training=False)
# Prepare model
config = Config.from_json_file(CONFIG_PATH)
model = QuestionAnswering(config)
if fp16 is True:
model.half()
model.load_state_dict(torch.load(CHK_PATH, map_location=device))
model.to(device)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", 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_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_example_index)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=batch_size)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits = model(
input_ids, segment_ids, input_mask)
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist()
end_logits = batch_end_logits[i].detach().cpu().tolist()
eval_feature = eval_features[example_index.item()]
unique_id = int(eval_feature.unique_id)
all_results.append(
RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
output_nbest_file = os.path.join("nbest_predictions.json")
write_predictions(eval_examples, eval_features, all_results, n_best_size,
max_answer_length, False, output, output_nbest_file,
None, False, False, 0.0)
def do_main():
# Set default configuration in args.py
args = get_args()
if args.local_rank == -1 or not args.cuda:
device = torch.device(
"cuda" if torch.cuda.is_available() and args.cuda else "cpu")
n_gpu = torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
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')
print('Device:', str(device).upper())
print('Number of GPUs:', n_gpu)
print('Distributed training:', bool(args.local_rank != -1))
print('FP16:', args.fp16)
# Set random seed for reproducibility
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)
dataset_map = {'News_art': News_artProcessor, 'News': News_Processor}
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.dataset not in dataset_map:
raise ValueError('Unrecognized dataset')
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path,
dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = False
tokenizer = BertTokenizer.from_pretrained(args.model,
is_lowercase=args.is_lowercase)
train_examples = None
num_train_optimization_steps = None
if args.trained_model:
train_examples = processor.get_train_examples(args.data_dir,
args.train_name)
num_train_optimization_steps = int(
math.ceil(len(train_examples) / args.batch_size) /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.model, num_labels=2) # creating news model!
#model = BertForSequenceClassification.from_pretrained(args.model, cache_dir=cache_dir, num_labels=args.num_labels)
if args.fp16:
model.half()
model.to(device)
#model = BertForSequenceClassification.from_pretrained(args.model, num_labels=args.num_labels)
model_ = torch.load(
args.trained_model,
map_location=lambda storage, loc: storage) # load personality model
state = {}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
del state['classifier.weight'] # removing personality classifier!
del state['classifier.bias']
model.load_state_dict(state, strict=False)
model = model.to(device)
# 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
}]
print('t_total :', num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
args.freez_bert = False
trainer = BertTrainer(model, optimizer, processor, args)
trainer.train()
model = torch.load(trainer.snapshot_path)
evaluate_split(model, processor, args, split=args.dev_name)
evaluate_split(model, processor, args, split=args.test_name)
def do_main():
# Set default configuration in args.py
args = get_args()
if args.local_rank == -1 or not args.cuda:
device = torch.device(
"cuda" if torch.cuda.is_available() and args.cuda else "cpu")
n_gpu = torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
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')
print('Device:', str(device).upper())
print('Number of GPUs:', n_gpu)
print('Distributed training:', bool(args.local_rank != -1))
print('FP16:', args.fp16)
# Set random seed for reproducibility
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)
dataset_map = {'News_art': News_artProcessor, 'News': News_Processor}
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if args.dataset not in dataset_map:
raise ValueError('Unrecognized dataset')
args.batch_size = args.batch_size // args.gradient_accumulation_steps
args.device = device
args.n_gpu = n_gpu
args.num_labels = dataset_map[args.dataset].NUM_CLASSES
args.is_multilabel = dataset_map[args.dataset].IS_MULTILABEL
if not args.trained_model:
save_path = os.path.join(args.save_path,
dataset_map[args.dataset].NAME)
os.makedirs(save_path, exist_ok=True)
processor = dataset_map[args.dataset]()
args.is_lowercase = 'uncased' in args.model
args.is_hierarchical = False
tokenizer = BertTokenizer.from_pretrained(args.model,
is_lowercase=args.is_lowercase)
num_train_optimization_steps = None
if args.trained_model:
train_examples = processor.get_train_examples(args.data_dir,
args.train_name)
num_train_optimization_steps = int(
math.ceil(len(train_examples) / args.batch_size) /
args.gradient_accumulation_steps) * args.epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
model_bert = BertForSequenceClassification.from_pretrained(args.model,
num_labels=4)
model_bert.to(device)
if args.trained_model:
model_ = torch.load(args.trained_model,
map_location=lambda storage, loc: storage
) # load personality model
state = {}
for key in model_.state_dict().keys():
new_key = key.replace("module.", "")
state[new_key] = model_.state_dict()[key]
del state['classifier.weight'] # removing personality classifier!
del state['classifier.bias']
model_bert.load_state_dict(state, strict=False)
model_bert = model_bert.to(device)
args.freez_bert = False
evaluate(model_bert,
processor,
args,
last_bert_layers=-1,
ngram_range=(1, 1))
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--config_file",
type=str,
required=True,
help="model configuration file")
parser.add_argument("--vocab_file",
type=str,
required=True,
help="tokenizer vocab file")
parser.add_argument("--checkpoint",
type=str,
required=True,
help="pretrained model checkpoint")
# Other parameters
parser.add_argument(
"--output_dir",
default="output",
type=str,
help=
"The output directory where the model predictions and checkpoints will be written."
)
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_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--train_batch_size",
default=16,
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=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=10.0,
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("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
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(
'--fp16_opt_level',
type=str,
default='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
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")
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 synchronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = n_gpu
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
set_seed(args)
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]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
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
config = Config(args.config_file)
#
if output_mode == "classification":
model = SequenceClassification(config, num_labels=num_labels)
elif output_mode == "classification_multitask":
model = SequenceClassificationMultitask(config)
model.bert.load_state_dict(torch.load(args.checkpoint))
# Do not fine-tune BERT parameters
# for param in model.bert.parameters():
# param.requires_grad = False
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
global_step, tr_loss = 0, 0.
train_features = convert_examples_to_features(train_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
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)
if output_mode == "classification_multitask":
all_label_ids = all_label_ids = torch.tensor(
[f.label_id for f in train_features], dtype=torch.long)
num_labels = [2, 3, 3]
elif output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.float)
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)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['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
}]
t_total = len(train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
scheduler = WarmupLinearSchedule(optimizer,
warmup_steps=t_total *
args.warmup_proportion,
t_total=t_total)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
if args.fp16_opt_level == "O2":
keep_batchnorm_fp32 = False
else:
keep_batchnorm_fp32 = True
model, optimizer = amp.initialize(
model,
optimizer,
opt_level=args.fp16_opt_level,
keep_batchnorm_fp32=keep_batchnorm_fp32)
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)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
writer = open(output_eval_file, "w", encoding="utf-8")
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
epoch_iterator = tqdm(train_dataloader,
desc="Train(XX Epoch) Step(X/X) (loss=X.X)",
disable=args.local_rank not in [-1, 0])
model.train()
for step, batch in enumerate(epoch_iterator):
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
logits = model(input_ids, segment_ids, input_mask, labels=None)
if output_mode == "classification_multitask":
loss_fct0 = CrossEntropyLoss()
loss_fct1 = CrossEntropyLoss()
loss_fct2 = CrossEntropyLoss()
loss0 = loss_fct0(logits[0].view(-1, num_labels[0]),
label_ids[:, 0].view(-1))
loss1 = loss_fct1(logits[1].view(-1, num_labels[1]),
label_ids[:, 1].view(-1))
loss2 = loss_fct2(logits[2].view(-1, num_labels[2]),
label_ids[:, 2].view(-1))
loss = loss0 + loss1 + loss2
elif output_mode == "classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), label_ids.view(-1))
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:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
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:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scheduler.step() # Update learning rate schedule
optimizer.step()
optimizer.zero_grad()
global_step += 1
epoch_iterator.set_description(
"Train(%d Epoch) Step(%d / %d) (loss=%5.5f)" %
(_, global_step, t_total, loss.item()))
model_checkpoint = "%s_%d.bin" % (args.task_name, _)
logger.info(model_checkpoint)
output_model_file = os.path.join(args.output_dir, model_checkpoint)
if n_gpu > 1 or args.local_rank != -1:
logger.info("***** Saving file *****(module)")
torch.save(model.module.state_dict(), output_model_file)
else:
logger.info("***** Saving file *****")
torch.save(model.state_dict(), output_model_file)
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,
output_mode)
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)
if output_mode == "classification_multitask":
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.long)
elif output_mode == "classification":
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.float)
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 = []
preds0 = []
preds1 = []
preds2 = []
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,
labels=None)
# create eval loss and other metric required by the task
if output_mode == "classification_multitask":
loss_fct0 = CrossEntropyLoss()
loss_fct1 = CrossEntropyLoss()
loss_fct2 = CrossEntropyLoss()
loss0 = loss_fct0(logits[0].view(-1, num_labels[0]),
label_ids[:, 0].view(-1))
loss1 = loss_fct1(logits[1].view(-1, num_labels[1]),
label_ids[:, 1].view(-1))
loss2 = loss_fct2(logits[2].view(-1, num_labels[2]),
label_ids[:, 2].view(-1))
tmp_eval_loss = loss0 + loss1 + loss2
elif output_mode == "classification":
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
tmp_eval_loss = loss_fct(logits.view(-1),
label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if output_mode == "classification_multitask":
if len(preds0) == 0:
preds0.append(logits[0].detach().cpu().numpy())
preds1.append(logits[1].detach().cpu().numpy())
preds2.append(logits[2].detach().cpu().numpy())
else:
preds0[0] = np.append(preds0[0],
logits[0].detach().cpu().numpy(),
axis=0)
preds1[0] = np.append(preds1[0],
logits[1].detach().cpu().numpy(),
axis=0)
preds2[0] = np.append(preds2[0],
logits[2].detach().cpu().numpy(),
axis=0)
else:
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
if output_mode == "classification_multitask":
preds0 = preds0[0]
preds1 = preds1[0]
preds2 = preds2[0]
preds0 = np.argmax(preds0, axis=1)
preds1 = np.argmax(preds1, axis=1)
preds2 = np.argmax(preds2, axis=1)
preds = [preds0, preds1, preds2]
elif output_mode == "classification":
preds = preds[0]
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = preds[0]
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / global_step
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
logger.info("***** Eval results *****")
writer.write("%d_epoch\n" % (_ + 1))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("\n")
writer.close()
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--corpus_path",
default=None,
type=str,
required=True,
help="The input training data file (a text file).")
parser.add_argument("--model_name",
type=str,
required=True,
help="model name")
parser.add_argument("--gen_config_file",
type=str,
required=True,
help="generator model configuration file")
parser.add_argument("--disc_config_file",
type=str,
required=True,
help="discriminator model configuration file")
## Other parameters
parser.add_argument(
"--output_dir",
default='checkpoint',
type=str,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument(
"--max_seq_length",
default=512,
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("--train_batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for training.")
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("--learning_rate",
default=5e-4,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="Weight deay if we apply some.")
parser.add_argument("--adam_epsilon",
default=1e-6,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--num_steps",
default=1000000,
type=int,
help="Total number of training epochs to perform.")
parser.add_argument("--warmup_steps",
default=10000,
type=int,
help="Linear warmup over warmup_steps.")
parser.add_argument(
"--mask_prob",
default=0.15,
type=float,
help="Masking probability (Small, Base=0.15 Large=0.25)")
parser.add_argument('--logging_steps',
type=int,
default=8,
help="Log every X updates steps.")
parser.add_argument('--save_steps',
type=int,
default=50000,
help="Save checkpoint every X updates steps.")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument(
'--loss_scale',
type=float,
default=0.0,
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("--local_rank",
type=int,
default=-1,
help="For distributed training: local_rank")
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
if args.local_rank == -1:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
# Load pretrained model and tokenizer
tokenizer = BertTokenizer(vocab_file='dataset/wiki_vocab_32k_0213.txt',
do_lower_case=args.do_lower_case,
max_len=args.max_seq_length,
do_basic_tokenize=True)
generator_config = Config(args.gen_config_file)
discriminator_config = Config(args.disc_config_file)
model = Electra(generator_config, discriminator_config)
gen_params = count_parameters(model.generator)
disc_params = count_parameters(model.discriminator)
logger.info("Generator Model Parameter: %d" % gen_params)
logger.info("Discriminator Model Parameter: %d" % disc_params)
logger.info("Generator Configuration: %s" % generator_config)
logger.info("Discriminator Configuration: %s" % discriminator_config)
model.to(args.device)
args.vocab_size = generator_config.vocab_size
logger.info("Training parameters %s", args)
# Training
train(args, model, tokenizer)
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--config_file",
type=str,
required=True,
help="model configuration file")
parser.add_argument("--vocab_file",
type=str,
required=True,
help="tokenizer vocab file")
parser.add_argument("--checkpoint",
default=None,
type=str,
required=True,
help="fine-tuned model checkpoint")
# Other parameters
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("--eval_batch_size",
default=32,
type=int,
help="Total batch size for eval.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
args = parser.parse_args()
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
args.n_gpu = n_gpu
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: {} n_gpu: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
set_seed(args)
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]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
test_examples = processor.get_test_examples(args.data_dir)
# Prepare model
config = Config(args.config_file)
model = SequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
if args.fp16:
model = model.half()
test_features = convert_examples_to_features(test_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running Evaluation *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in test_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in test_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in test_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in test_features],
dtype=torch.float)
test_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_label_ids)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.eval_batch_size)
model.eval()
preds = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
test_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, labels=None)
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
preds = preds[0]
if output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
def train(hp):
tokenizer = BertTokenizer.from_pretrained(hp.bert_model_dir, do_lower_case=False)
if hp.dataset == "lm_raw_data_finance":
dict_file = "../data/dataset_finance/raw_data/annual_report_entity_list"
elif hp.dataset == "lm_raw_data_novel":
dict_file = "../data/dataset_book9/raw_data/entity_book9"
elif hp.dataset == "lm_raw_data_thuner":
dict_file = "../data/dataset_thuner/raw_data/thu_entity.txt"
entity_dict = EntityDict(hp.dataset, dict_file)
entity_dict.load(os.path.join(hp.bert_model_dir, 'entity.dict'))
device = 'cuda' if torch.cuda.is_available() else 'cpu'
ner_label = NerLabel([hp.trainset, hp.validset])
fname = os.path.join(hp.logdir, 'dict.pt')
ner_label.save(fname)
train_dataset = NerDataset(hp.trainset, ner_label, tokenizer, entity_dict)
eval_dataset = NerDataset(hp.validset, ner_label, tokenizer, entity_dict)
test_dataset = NerDataset(hp.testset, ner_label, tokenizer, entity_dict)
model = Net(hp.bert_model_dir, hp.top_rnns, len(ner_label.VOCAB), entity_dict.entity_num, device, hp.finetuning).to(device)
device_ids = [0, 1]
model = nn.DataParallel(model, device_ids=device_ids)
train_iter = data.DataLoader(dataset=train_dataset,
batch_size=hp.batch_size,
shuffle=True,
num_workers=4,
collate_fn=pad)
eval_iter = data.DataLoader(dataset=eval_dataset,
batch_size=hp.batch_size,
shuffle=False,
num_workers=4,
collate_fn=pad)
test_iter = data.DataLoader(dataset=test_dataset,
batch_size=hp.batch_size,
shuffle=False,
num_workers=4,
collate_fn=pad)
optimizer = optim.Adam(model.parameters(), lr=hp.lr)
criterion = nn.CrossEntropyLoss(ignore_index=0)
## train the model
best_eval = -10
for epoch in range(1, hp.n_epochs + 1):
train_epoch(model, train_iter, optimizer, criterion, tokenizer)
print(f"=========eval at epoch={epoch}=========")
if not os.path.exists(hp.logdir): os.makedirs(hp.logdir)
fname = os.path.join(hp.logdir, 'model')
precision, recall, f1 = evaluate(model, eval_iter, fname, ner_label, verbose=False)
if f1 > best_eval:
best_eval = f1
print("epoch{} get the best eval f-score:{}".format(epoch, best_eval))
torch.save(model.state_dict(), f"{fname}.pt")
print(f"weights were saved to {fname}.pt")
print(f"=========test at epoch={epoch}=========")
if not os.path.exists(hp.logdir): os.makedirs(hp.logdir)
fname = os.path.join(hp.logdir, str(epoch))
precision, recall, f1 = evaluate(model, test_iter, fname, ner_label, verbose=False)
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(args.vocab_file,
max_len=args.max_seq_length,
do_basic_tokenize=True)
# Prepare model
config = Config.from_json_file(args.config_file)
model = QuestionAnswering(config)
model.bert.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
logger.info("Training hyper-parameters %s", args)
# Training
train_dataset = load_and_cache_examples(args, tokenizer)
train(args, train_dataset, model)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint",
default='output/korquad_3.bin',
type=str,
help="checkpoint")
parser.add_argument(
"--output_dir",
default='debug',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--model_config", type=str)
parser.add_argument("--vocab", type=str)
## Other parameters
parser.add_argument(
"--predict_file",
default='data/KorQuAD_v1.0_dev.json',
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument(
"--max_seq_length",
default=512,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument(
"--doc_stride",
default=64,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
parser.add_argument(
"--max_query_length",
default=96,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument(
"--n_best_size",
default=20,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions.json output file."
)
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help=
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument(
"--verbose_logging",
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
parser.add_argument('--ensemble', default='false', type=str)
args = parser.parse_args()
args.ensemble = args.ensemble.lower() == 'true'
# Setup CUDA, GPU & distributed training
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: %s, n_gpu: %s, 16-bits training: %s", device,
args.n_gpu, args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(vocab_file=args.vocab,
do_basic_tokenize=True,
max_len=args.max_seq_length)
config = Config.from_json_file(args.model_config)
model = QuestionAnswering(config)
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
if not args.ensemble:
logger.info(" Load Model: %s ", args.checkpoint)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
results = evaluate(args, model, examples, features)
else:
list_ckpts = []
with open(os.path.join(args.output_dir, "ckpt_list.txt"), 'r') as f:
for line in f:
list_ckpts.append(line.strip())
list_results = []
for i, ckpt in enumerate(list_ckpts):
fn = os.path.join(args.output_dir, ckpt)
logger.info(" Load Model: %s ", fn)
model.load_state_dict(torch.load(fn))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
results = evaluate(args, model, examples, features)
list_results.append(results)
list_qid = []
for example in examples:
list_qid.append(example.qas_id)
all_predictions = collections.OrderedDict()
for qid in list_qid:
max_prob, answer = 0.0, ""
for results in list_results:
prob, text = 0.0, None
for output in results[qid]:
if output["text"]:
prob = output["probability"]
text = output["text"]
break
if prob > max_prob:
max_prob = prob
answer = text
all_predictions[qid] = answer
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
with open(output_prediction_file, "w") as writer:
writer.write(json.dumps(all_predictions, indent=4) + "\n")
expected_version = 'KorQuAD_v1.0'
with open(args.predict_file) as dataset_file:
dataset_json = json.load(dataset_file)
read_version = "_".join(dataset_json['version'].split("_")[:-1])
if (read_version != expected_version):
logger.info('Evaluation expects ' + expected_version +
', but got dataset with ' + read_version,
file=sys.stderr)
dataset = dataset_json['data']
with open(os.path.join(args.output_dir,
"predictions.json")) as prediction_file:
predictions = json.load(prediction_file)
logger.info(json.dumps(korquad_eval(dataset, predictions)))
try:
from apex.fp16_utils import FP16_Optimizer
except:
logger.info('WARNING: apex not installed, ignoring --fp16 option')
args.fp16 = False
# object of <class 'torch.device'>
device = torch.device(
'cuda' if torch.cuda.is_available() and args.cuda else 'cpu')
n_gpu = torch.cuda.device_count()
########################################################################################################################
# Load Data
########################################################################################################################
## using the Bert Word Vocabulary
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
vocab_size = len(tokenizer.vocab)
corpus = load_lm_data(args.entity_dict, args.data, args.output_dir,
args.dataset, tokenizer)
## Training Dataset
train_iter = corpus.get_iterator('train',
args.batch_size,
args.max_seq_length,
args.max_doc_length,
device=device)
## total batch numbers and optim updating steps
total_train_steps = int(train_iter.batch_steps * args.num_train_epochs)
########################################################################################################################
# Building the model
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint",
default='output/korquad_3.bin',
type=str,
help="checkpoint")
parser.add_argument(
"--output_dir",
default='debug',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
## Other parameters
parser.add_argument(
"--predict_file",
default='data/KorQuAD_v1.0_dev.json',
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument(
"--config_name",
default="data/bert_small.json",
type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument(
"--max_seq_length",
default=512,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument(
"--doc_stride",
default=64,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
parser.add_argument(
"--max_query_length",
default=96,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument(
"--n_best_size",
default=20,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions.json output file."
)
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help=
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument(
"--verbose_logging",
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: %s, n_gpu: %s, 16-bits training: %s", device,
args.n_gpu, args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(vocab_file='data/wiki_vocab_32k.txt',
do_basic_tokenize=True,
max_len=args.max_seq_length)
config = Config.from_json_file(args.config_name)
model = QuestionAnswering(config)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
evaluate(args, model, examples, features)
def main():
parser = argparse.ArgumentParser()
# Parameters
parser.add_argument(
"--output_dir",
default='debug',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--checkpoint",
default='output/korquad_3.bin',
type=str,
help="fine-tuned model checkpoint")
parser.add_argument("--config_file",
default='data/large_config.json',
type=str,
help="model configuration file")
parser.add_argument("--vocab_file",
default='data/large_v1_32k_vocab.txt',
type=str,
help="tokenizer vocab file")
parser.add_argument("--check_list", default=False, type=bool)
parser.add_argument("--eda_type",
default="no_eda",
type=str,
help="sr, ri , rd, rs")
parser.add_argument(
"--predict_file",
default='data/korquad/KorQuAD_v1.0_dev.json',
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument(
"--max_seq_length",
default=512,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument(
"--doc_stride",
default=64,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help=
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument("--batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument(
"--n_best_size",
default=20,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions."
"json output file.")
parser.add_argument(
"--verbose_logging",
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--just_eval", default=False, type=bool)
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: %s, n_gpu: %s, 16-bits training: %s", device,
args.n_gpu, args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(vocab_file=args.vocab_file,
do_basic_tokenize=True,
max_len=args.max_seq_length)
config = Config.from_json_file(args.config_file)
model = QuestionAnswering(config)
if args.check_list:
model_list = os.listdir(args.checkpoint)
model_list.sort()
result_dict = defaultdict(list)
eda_list = ["st_rs", "st_rd", "rs", "rd", "sr", "ri", "no_eda"]
for m in model_list:
model.load_state_dict(torch.load(args.checkpoint + '/' + m))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
for eda in eda_list:
if eda in m:
args.eda_type = eda
break
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
r = evaluate(args, model, examples, features)
em = r["exact_match"]
f1 = r["f1"]
result_dict[args.eda_type].append([em, f1, m])
print(result_dict)
else:
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
evaluate(args, model, examples, features)