def get_model_and_tokenizer(model_name, device):
save_ckpt_path = CHECK_POINT[model_name]
if model_name == "koelectra":
model_name_or_path = "monologg/koelectra-base-discriminator"
tokenizer = ElectraTokenizer.from_pretrained(model_name_or_path)
electra_config = ElectraConfig.from_pretrained(model_name_or_path)
model = koElectraForSequenceClassification.from_pretrained(
pretrained_model_name_or_path=model_name_or_path,
config=electra_config,
num_labels=359)
elif model_name == 'kobert':
tokenizer = get_tokenizer()
model = KoBERTforSequenceClassfication()
if os.path.isfile(save_ckpt_path):
checkpoint = torch.load(save_ckpt_path, map_location=device)
pre_epoch = checkpoint['epoch']
# pre_loss = checkpoint['loss']
model.load_state_dict(checkpoint['model_state_dict'])
print(f"load pretrain from: {save_ckpt_path}, epoch={pre_epoch}")
return model, tokenizer
def token_num(data_path='./data/train.jsonl'):
data = []
with open(data_path, 'r') as json_file:
json_list = list(json_file)
bert_tok = get_tokenizer()
gpt_tok = get_kogpt2_tokenizer()
bert_tok_num = 0
gpt_tok_num = 0
count = 0
for json_str in json_list:
json_data = json.loads(json_str)
tmp_str = json_data['abstractive']
# for arti_str in json_data['article_original']:
# tmp_str += arti_str
bert_tok_num = max(
bert_tok_num,
len(bert_tok.encode(tmp_str, max_length=512, truncation=True)))
gpt_tok_num = max(
gpt_tok_num,
len(gpt_tok.encode(tmp_str, max_length=512, truncation=True)))
# print(len(json_data['article_original']))
# sum_len += len(json_data['article_original'])
# count += 1
# print('average article_original len - ', sum_len/count)
print('max bert token len:', bert_tok_num)
print('max gpt token len:', gpt_tok_num)
def _load_custom_model(self, bert_name):
custom_config = AutoConfig.from_pretrained(bert_name)
custom_config.output_hidden_states = True
custom_tokenizer = get_tokenizer()
custom_model = AutoModel.from_pretrained(bert_name,
config=custom_config)
self.model = Summarizer(custom_model=custom_model,
custom_tokenizer=custom_tokenizer)
def __init__(
self,
file_path="wellness_classification.txt",
num_label=359,
device='cpu',
max_seq_len=512, # KoBERT max_length
tokenizer=None):
self.file_path = file_path
self.device = device
self.data = []
tokenizer = tokenizer or get_tokenizer()
labels = {}
with open(self.file_path, 'r', encoding='utf-8') as f:
for line in f:
line = line.rstrip()
if not line:
continue
datas = line.split('\t')
index_of_words = tokenizer.encode(datas[0])
token_type_ids = [0] * len(index_of_words)
attention_mask = [1] * len(index_of_words)
# Padding Length
padding_length = max_seq_len - len(index_of_words)
# Zero Padding
index_of_words += [0] * padding_length
token_type_ids += [0] * padding_length
attention_mask += [0] * padding_length
labels.setdefault(datas[1], len(labels))
# Label
label = int(labels[datas[1]])
data = {
'input_ids': torch.tensor(index_of_words).to(self.device),
'token_type_ids':
torch.tensor(token_type_ids).to(self.device),
'attention_mask':
torch.tensor(attention_mask).to(self.device),
'labels': torch.tensor(label).to(self.device)
}
self.data.append(data)
print(label)
print(len(labels))
def create_loader(data_dir, model, mode, batch_size, ratio=0.8):
dataset = pd.read_csv(data_dir)
if 'bert' in model:
tokenizer = get_tokenizer()
else:
tokenizer = ElectraTokenizer.from_pretrained(
"monologg/koelectra-base-v3-discriminator")
if mode == 'train':
data, data_label = dataset['content'], dataset['info']
encodings = tokenizer(list(data.values), truncation=True, padding=True)
dataset = NewsDataset(encodings, data_label)
data_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
return data_loader, None
elif mode == 'val':
data, val, data_label, val_label = train_test_split(
dataset['content'],
dataset['info'],
train_size=ratio,
stratify=dataset['info'])
encodings = tokenizer(list(data.values), truncation=True, padding=True)
val_encodings = tokenizer(list(val.values),
truncation=True,
padding=True)
dataset = NewsDataset(encodings, data_label)
val_dataset = NewsDataset(val_encodings, val_label)
data_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
val_loader = DataLoader(val_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
return data_loader, val_loader
else:
data = dataset['content']
encodings = tokenizer(list(data.values), truncation=True, padding=True)
dataset = NewsDataset(encodings)
data_loader = DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4)
return data_loader, None
def __init__(self,
file_path = "../data/wellness_dialog_for_text_classification.txt",
num_label = 359,
device = 'cpu',
max_seq_len = 512, # KoBERT max_length
tokenizer = None
):
self.file_path = file_path
self.device = device
self.data =[]
self.tokenizer = tokenizer if tokenizer is not None else get_tokenizer()
file = open(self.file_path, 'r', encoding='utf-8')
while True:
line = file.readline()
if not line:
break
datas = line.split(" ")
index_of_words = tokenizer.encode(datas[0])
token_type_ids = [0] * len(index_of_words)
attention_mask = [1] * len(index_of_words)
# Padding Length
padding_length = max_seq_len - len(index_of_words)
# Zero Padding
index_of_words += [0] * padding_length
token_type_ids += [0] * padding_length
attention_mask += [0] * padding_length
# Label
label = int(datas[1][:-1])
data = {
'input_ids': torch.tensor(index_of_words).to(self.device),
'token_type_ids': torch.tensor(token_type_ids).to(self.device),
'attention_mask': torch.tensor(attention_mask).to(self.device),
'labels': torch.tensor(label).to(self.device)
}
self.data.append(data)
file.close()
def __init__(self):
self.root_path = '..'
self.checkpoint_path = f"{self.root_path}/checkpoint"
self.save_ckpt_path = f"{self.checkpoint_path}/kobert-wellnesee-text-classification.pth"
#답변과 카테고리 불러오기
self.category, self.answer = load_wellness_answer()
ctx = "cuda" if torch.cuda.is_available() else "cpu"
self.device = torch.device(ctx)
# 저장한 Checkpoint 불러오기
checkpoint = torch.load(self.save_ckpt_path, map_location=self.device)
self.model = KoBERTforSequenceClassfication()
self.model.load_state_dict(checkpoint['model_state_dict'])
self.model.eval()
self.tokenizer = get_tokenizer()
def __init__(self, config):
"""
Constructor for EmbeddingBERTWordPhr_kor.
@param self The object pointer.
@param config Dictionary. Configuration for EmbeddingBERTWordPhr_kor
"""
super(EmbeddingBERTWordPhr_kor, self).__init__()
self.tokenizer = get_tokenizer()
self.model = BertModel.from_pretrained('monologg/kobert')
self.embed_size = 1536
self.special_tokens = config['special_tokens']
self.fine_tune = bool(config['train'])
if self.fine_tune:
self.model.train()
self.model.requires_grad = True
else:
self.model.eval()
self.model.requires_grad = False
def setting_similarity(standard, targets):
# combine separated sentences to the only one sentence & setting encoding form
def setting_encoding_form(separated_sentences_list):
for idx, content in enumerate(separated_sentences_list):
res = ""
for sentence in content:
res += sentence + " "
separated_sentences_list[idx] = "[CLS] " + res + "[SEP]"
return separated_sentences_list
# calculate similarity
def cos_sim(A, B):
return dot(A, B) / (norm(A) * norm(B))
# merge data & separate from ids to contents
merge_data = standard + targets
ids = list(i[0] for i in merge_data)
contents = list(i[1] for i in merge_data)
# similarity function
contents = setting_encoding_form(contents)
tokenizer = get_tokenizer()
tokenized_texts = [tokenizer.tokenize(content) for content in contents]
input_ids = [
tokenizer.convert_tokens_to_ids(x) for x in tokenized_texts
]
input_ids = pad_sequences(input_ids,
maxlen=1000,
dtype=int,
truncating="post",
padding="post")
res = {}
for i in range(1, len(input_ids)):
similar_val = round(cos_sim(input_ids[0], input_ids[i]) * 100, 1)
res.update({ids[i]: {'similarity': similar_val}})
return res
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model_type",
default=None,
type=str,
required=True,
help="Model type selected in the list: " +
", ".join(MODEL_CLASSES.keys()))
parser.add_argument("--model_name_or_path",
default=None,
type=str,
required=True,
help="Path to pre-trained model or shortcut name")
parser.add_argument("--prompt", type=str, default="")
parser.add_argument("--padding_text", type=str, default="")
parser.add_argument("--length", type=int, default=20)
parser.add_argument("--temperature", type=float, default=1.0)
parser.add_argument("--top_k", type=int, default=0)
parser.add_argument("--top_p", type=float, default=0.9)
parser.add_argument("--no_cuda",
action='store_true',
help="Avoid using CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
args = parser.parse_args()
args.device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
set_seed(args)
args.model_type = args.model_type.lower()
model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
# This, too, should be args-based
#tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
tokenizer = get_tokenizer()
model = model_class.from_pretrained(args.model_name_or_path)
model.to(args.device)
model.eval()
if args.length < 0 and model.config.max_position_embeddings > 0:
args.length = model.config.max_position_embeddings
elif 0 < model.config.max_position_embeddings < args.length:
args.length = model.config.max_position_embeddings # No generation bigger than model size
elif args.length < 0:
args.length = MAX_LENGTH # avoid infinite loop
print(args)
while True:
raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
if args.model_type in ["transfo-xl", "xlnet"]:
# Models with memory likes to have a long prompt for short inputs.
raw_text = (args.padding_text
if args.padding_text else PADDING_TEXT) + raw_text
context_tokens = tokenizer.encode(raw_text)
out = sample_sequence(
model=model,
context=context_tokens,
length=args.length,
temperature=args.temperature,
top_k=args.top_k,
top_p=args.top_p,
device=args.device,
is_xlnet=bool(args.model_type == "xlnet"),
)
out = out[0, len(context_tokens):].tolist()
text = tokenizer.decode(out, clean_up_tokenization_spaces=True)
print(text)
if args.prompt:
break
return text
def main(conf):
# Prepare data
train_dev = koco.load_dataset("korean-hate-speech", mode="train_dev")
train, valid = train_dev["train"], train_dev["dev"]
# Prepare tokenizer
tokenizer = (
get_tokenizer()
if "kobert" in conf.pretrained_model
else AutoTokenizer.from_pretrained(conf.pretrained_model)
)
if conf.tokenizer.register_names:
names = pd.read_csv("entertainement_biographical_db.tsv", sep="\t")[
"name_wo_parenthesis"
].tolist()
tokenizer.add_tokens(names)
# Mapping string y_label to integer label
if conf.label.hate:
train, label2idx = map_label2idx(train, "hate")
valid, _ = map_label2idx(valid, "hate")
elif conf.label.bias:
train, label2idx = map_label2idx(train, "bias")
valid, _ = map_label2idx(valid, "bias")
# Use bias as an additional context for predicting hate
if conf.label.hate and conf.label.bias:
biases = ["gender", "others", "none"]
tokenizer.add_tokens([f"<{label}>" for label in biases])
# Prepare DataLoader
train_dataset = KoreanHateSpeechDataset(train)
valid_dataset = KoreanHateSpeechDataset(valid)
collator = KoreanHateSpeechCollator(
tokenizer, predict_hate_with_bias=(conf.label.hate and conf.label.bias)
)
train_loader = DataLoader(
train_dataset,
batch_size=conf.train_hparams.batch_size,
shuffle=True,
collate_fn=collator.collate,
)
valid_loader = DataLoader(
valid_dataset,
batch_size=conf.train_hparams.batch_size,
shuffle=False,
collate_fn=collator.collate,
)
# Prepare model
set_seeds(conf.train_hparams.seed)
model = BertForSequenceClassification.from_pretrained(
conf.pretrained_model, num_labels=len(label2idx)
)
if conf.tokenizer.register_names:
model.resize_token_embeddings(len(tokenizer))
elif conf.label.hate and conf.label.bias:
model.resize_token_embeddings(len(tokenizer))
model = model.to(device)
# Prepare optimizer and scheduler
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p
for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay": 0.01,
},
{
"params": [
p
for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay": 0.0,
},
]
optimizer = optim.AdamW(
optimizer_grouped_parameters,
lr=conf.train_hparams.lr,
eps=conf.train_hparams.adam_epsilon,
)
n_total_iterations = len(train_loader) * conf.train_hparams.n_epochs
n_warmup_steps = int(n_total_iterations * conf.train_hparams.warmup_ratio)
scheduler = get_linear_schedule_with_warmup(
optimizer, n_warmup_steps, n_total_iterations
)
criterion = nn.CrossEntropyLoss()
criterion = criterion.to(device)
# Train!
trainer = BertTrainer(conf.train_hparams)
model = trainer.train(
model, criterion, optimizer, scheduler, train_loader, valid_loader
)
makedirs(conf.checkpoint_dir)
makedirs(conf.log_dir)
checkpoint_path = f"{conf.checkpoint_dir}/{conf.model_name}.pt"
log_path = f"{conf.log_dir}/{conf.model_name}.log"
torch.save({"model": model.state_dict()}, checkpoint_path)
torch.save({"config": conf, "classes": label2idx, "tokenizer": tokenizer}, log_path)
def __init__(
self,
data_path='./data/train.jsonl',
num_label=2, # 추출할것과 추출하지 않을 것들로
device='cpu',
max_seq_len=512, # KoBERT max_length
):
self.device = device
self.data = []
self.tokenizer = get_tokenizer()
cls_token_id = self.tokenizer.cls_token_id # [CLS]
sep_token_id = self.tokenizer.sep_token_id # [SEP]
pad_token_id = self.tokenizer.pad_token_id # [PAD]
jsonl_datas = jsonl_load(data_path=data_path)
# for dict_data in jsonl_datas:
for dict_data in tqdm(jsonl_datas):
articles = dict_data['article_original']
extractive_indices = dict_data['extractive']
index_of_words = None
token_type_ids = None
label = None
token_num = None
token_type_state = False
for idx in range(len(articles)):
label_state = True if idx in extractive_indices else False
if idx == 0: # 맨 처음 문장인 경우
index_of_words = [cls_token_id]
token_type_ids = [int(token_type_state)]
label = [int(label_state)]
token_num = 1
article = articles[idx]
tmp_index = self.tokenizer.encode(article,
add_special_tokens=False)
num_tmp_index = len(tmp_index) + 1
if token_num + num_tmp_index <= max_seq_len:
index_of_words += tmp_index + [sep_token_id]
token_type_ids += [int(token_type_state)] * num_tmp_index
label += [int(label_state)] * num_tmp_index
token_num += num_tmp_index
token_type_state = not token_type_state
if token_num + num_tmp_index > max_seq_len or idx == len(
articles) - 1:
# attention mask
attention_mask = [1] * token_num
# Padding Length
padding_length = max_seq_len - token_num
# Padding
index_of_words += [pad_token_id
] * padding_length # [PAD] padding
token_type_ids += [
token_type_state
] * padding_length # last token_type_state padding
attention_mask += [0] * padding_length # zero padding
# Label Zero Padding
label += [0] * padding_length
# Data Append
data = {
'input_ids':
torch.tensor(index_of_words).to(self.device),
'token_type_ids':
torch.tensor(token_type_ids).to(self.device),
'attention_mask':
torch.tensor(attention_mask).to(self.device),
'labels':
torch.tensor(label).to(self.device)
}
self.data.append(data)
# Data Initialization
index_of_words = [cls_token_id]
token_type_ids = [int(token_type_state)]
label = [int(label_state)]
token_num = 1
token_type_state = False
target = []
valid_length = []
for s in sentence:
s = "[CLS] " + s
input_ids = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(s))
valid_length.append(len(input_ids))
target_ids = [1] * 32
target_ids[:len(input_ids)] = input_ids
target.append(target_ids)
return target, valid_length
from kobert_transformers import get_kobert_model
from kobert_transformers import get_tokenizer
tokenizer = get_tokenizer()
input_ids, valid_length = gen_input_ids(
tokenizer=tokenizer, sentence=["한국어 모델을 공유합니다.", "두번째 문장입니다."])
model = get_kobert_model()
model.eval()
input_ids = torch.LongTensor(input_ids)
attention_mask = gen_attention_mask(input_ids, valid_length)
attention_mask = torch.LongTensor(attention_mask)
token_type_ids = torch.zeros_like(input_ids)
sequence_output, pooled_output = model(input_ids, attention_mask,
token_type_ids)
pooled_output
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_data_file",
default=None,
type=str,
required=True,
help="The input training data file (a text file).")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--eval_data_file",
default=None,
type=str,
help=
"An optional input evaluation data file to evaluate the perplexity on (a text file)."
)
parser.add_argument("--model_type",
default="bert",
type=str,
help="The model architecture to be fine-tuned.")
parser.add_argument(
"--model_name_or_path",
default="bert-base-cased",
type=str,
help="The model checkpoint for weights initialization.")
parser.add_argument(
"--mlm",
action='store_true',
help=
"Train with masked-language modeling loss instead of language modeling."
)
parser.add_argument(
"--mlm_probability",
type=float,
default=0.15,
help="Ratio of tokens to mask for masked language modeling loss")
parser.add_argument(
"--config_name",
default="",
type=str,
help=
"Optional pretrained config name or path if not the same as model_name_or_path"
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help=
"Optional pretrained tokenizer name or path if not the same as model_name_or_path"
)
parser.add_argument("--tokenizer_class",
default="",
type=str,
help="Optional pretrained tokenizer clas")
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)"
)
parser.add_argument(
"--block_size",
default=-1,
type=int,
help="Optional input sequence length after tokenization."
"The training dataset will be truncated in block of this size for training."
"Default to the model max input length for single sentence inputs (take into account special tokens)."
)
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training",
action='store_true',
help="Run evaluation during training at each logging step.")
parser.add_argument('--eval_steps',
type=int,
default=100,
help="Evaluate every X updates steps.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--per_gpu_train_batch_size",
default=4,
type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size",
default=4,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.0,
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_train_epochs",
default=1.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help=
"If > 0: set total number of training steps to perform. Override num_train_epochs."
)
parser.add_argument("--warmup_samples",
default=0,
type=int,
help="Linear warmup over warmup_samples.")
parser.add_argument("--lr_decay",
action='store_true',
help="Decay LR using get_linear_schedule_with_warmup.")
parser.add_argument(
"--unfreeze_level",
default=-1,
type=int,
help="If > 0: freeze all layers except few first and last.")
parser.add_argument('--logging_steps',
type=int,
default=50,
help="Log every X updates steps.")
parser.add_argument('--save_steps',
type=int,
default=50,
help="Save checkpoint every X updates steps.")
parser.add_argument(
'--save_total_limit',
type=int,
default=None,
help=
'Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default'
)
parser.add_argument(
"--eval_all_checkpoints",
action='store_true',
help=
"Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number"
)
parser.add_argument("--no_cuda",
action='store_true',
help="Avoid using CUDA when available")
parser.add_argument('--overwrite_output_dir',
action='store_true',
help="Overwrite the content of the output directory")
parser.add_argument(
'--overwrite_cache',
action='store_true',
help="Overwrite the cached training and evaluation sets")
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='O1',
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",
type=int,
default=-1,
help="For distributed training: local_rank")
parser.add_argument('--server_ip',
type=str,
default='',
help="For distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="For distant debugging.")
args = parser.parse_args()
if args.model_type in ["bert", "roberta", "distilbert"] and not args.mlm:
raise ValueError(
"BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm "
"flag (masked language modeling).")
if args.eval_data_file is None and args.do_eval:
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir
) and args.do_train and not args.overwrite_output_dir:
raise ValueError(
f"Output directory ({args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup distant debugging if needed
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
# 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, args.device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Barrier to make sure only the first process in distributed training download model & vocab
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path)
#if args.tokenizer_class: tokenizer_class = globals()[args.tokenizer_class]
#tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
# Okay, okay, I know, will be selectable from commandline in some future
tokenizer = get_tokenizer()
if args.block_size <= 0:
args.block_size = tokenizer.max_len_single_sentence # Our input block size will be the max possible for the model
args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
config=config)
model.to(args.device)
print(200 * '/')
print(
len([
param for item in flatten_model(model)
for param in item.parameters() if param.requires_grad
])) # freeze all layers but few first and last
if args.unfreeze_level >= 0:
flat = flatten_model(model)
flat = [item for item in flat if list(item.parameters())]
i_start = 3
i_end = 1
need_grads = set(flat[:i_start + args.unfreeze_level * 3]) | set(
flat[-(i_end + args.unfreeze_level * 3):])
for item in flat:
requires_grad(item, item in need_grads)
print(200 * '/')
print(
len([
param for item in flatten_model(model)
for param in item.parameters() if param.requires_grad
]))
if args.local_rank == 0:
torch.distributed.barrier(
) # End of barrier to make sure only the first process in distributed training download model & vocab
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
if args.local_rank not in [-1, 0]:
torch.distributed.barrier(
) # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache
train_dataset = load_and_cache_examples(args,
tokenizer,
evaluate=False)
if args.local_rank == 0:
torch.distributed.barrier()
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
# Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
if args.do_train and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
save_state(args, model, tokenizer, global_step)
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(
os.path.dirname(c) for c in sorted(
glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME,
recursive=True)))
logging.getLogger("transformers.modeling_utils").setLevel(
logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split(
'-')[-1] if len(checkpoints) > 1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=global_step)
result = dict(
(k + '_{}'.format(global_step), v) for k, v in result.items())
results.update(result)
return results
def __init__(self):
self.tokenizer = get_tokenizer()
self.token2idx = self.tokenizer.token2idx
self.idx2token = {v: k for k, v in self.token2idx.items()}
