def __init__(self, args, dictionary, left_pad=False):
super().__init__(dictionary)
self.dropout = args.dropout
from pytorch_transformers import RobertaModel, BertModel
from pytorch_transformers.file_utils import PYTORCH_TRANSFORMERS_CACHE
from pytorch_transformers import RobertaConfig, RobertaTokenizer, BertConfig, BertTokenizer
if args.pretrained_bert_model.startswith('roberta'):
self.embed = RobertaModel.from_pretrained(
args.pretrained_bert_model,
cache_dir=PYTORCH_TRANSFORMERS_CACHE /
'distributed_{}'.format(args.distributed_rank))
# self.context = RobertaModel.from_pretrained(args.pretrained_bert_model,
# cache_dir=PYTORCH_TRANSFORMERS_CACHE / 'distributed_{}'.format(args.distributed_rank))
self.config = RobertaConfig.from_pretrained(
args.pretrained_bert_model)
self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
else:
self.embed = BertModel.from_pretrained(
args.pretrained_bert_model,
cache_dir=PYTORCH_TRANSFORMERS_CACHE /
'distributed_{}'.format(args.distributed_rank))
# self.context = BertModel.from_pretrained(args.pretrained_bert_model,
# cache_dir=PYTORCH_TRANSFORMERS_CACHE / 'distributed_{}'.format(args.distributed_rank))
self.config = BertConfig.from_pretrained(
args.pretrained_bert_model)
self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
self.padding_idx = self.tokenizer.convert_tokens_to_ids(
self.tokenizer.pad_token)
def main():
bert_base_config = BertConfig.from_pretrained('bert-base-uncased', num_labels=2)
bert_base_model = BertForSequenceClassification.from_pretrained('bert-base-uncased', config=bert_base_config)
count = 0
for name, param in bert_base_model.named_parameters():
if param.requires_grad:
size = 1
for s in param.data.size():
size = s * size
count += size
print('The total number of parameters in bert_base_uncased: ', count)
roberta_config = RobertaConfig.from_pretrained('roberta-base', num_labels=2)
roberta_model = RobertaForSequenceClassification.from_pretrained('roberta-base',config=roberta_config)
count = 0
for name, param in roberta_model.named_parameters():
if param.requires_grad:
size = 1
for s in param.data.size():
size = s * size
count += size
print('The total number of parameters in roberta: ', count)
albert_config = AlbertConfig.from_pretrained('albert-base-v2', num_labels=2)
albert_model = AlbertForSequenceClassification.from_pretrained('albert-base-v2', config=albert_config)
count = 0
for name, param in albert_model.named_parameters():
if param.requires_grad:
size = 1
for s in param.data.size():
size = s * size
count += size
print('The total number of parameters in albert: ', count)
def __init__(self, args, dictionary, embed_tokens, left_pad=False):
super().__init__(dictionary)
self.dropout = args.dropout
# from pytorch_transformers import RobertaModel
from fairseq.modules.roberta_causal_mask import RobertaCasulMaskModel, BertCasulMaskModel
from pytorch_transformers.file_utils import PYTORCH_TRANSFORMERS_CACHE
from pytorch_transformers import RobertaConfig, RobertaTokenizer, BertConfig, BertTokenizer
if args.roberta_model.startswith('roberta'):
self.roberta = RobertaCasulMaskModel.from_pretrained(
args.roberta_model,
cache_dir=PYTORCH_TRANSFORMERS_CACHE /
'distributed_{}'.format(args.distributed_rank))
self.config = RobertaConfig.from_pretrained(args.roberta_model)
self.tokenizer = RobertaTokenizer.from_pretrained(
args.roberta_model)
else:
self.roberta = BertCasulMaskModel.from_pretrained(
args.roberta_model,
cache_dir=PYTORCH_TRANSFORMERS_CACHE /
'distributed_{}'.format(args.distributed_rank))
self.config = BertConfig.from_pretrained(args.roberta_model)
self.tokenizer = BertTokenizer.from_pretrained(args.roberta_model)
self.config.output_attentions = True
self.roberta.pooler.dense.weight.requires_grad = False
self.roberta.pooler.dense.bias.requires_grad = False
embed_dim = embed_tokens.embedding_dim
self.padding_idx = embed_tokens.padding_idx
# self.embed_tokens = embed_tokens
# self.embed_scale = math.sqrt(embed_dim)
self.args = args
# if args.sentence_transformer_arch == 'fairseq':
# self.padding_idx = embed_tokens.padding_idx
# self.sent_embed_positions = PositionalEmbedding(
# 1024, embed_dim, self.padding_idx,
# left_pad=False,
# learned=args.encoder_learned_pos,
# )
# self.doc_layers = nn.ModuleList([])
# self.doc_layers.extend([
# TransformerEncoderLayer(args)
# for i in range(args.encoder_layers)
# ])
if args.sentence_transformer_arch == 'bert':
# from pytorch_transformers import RobertaConfig, RobertaTokenizer
# self.config = RobertaConfig.from_pretrained(args.roberta_model)
# self.config.output_attentions = True
# self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
embed_dim = self.config.hidden_size
print('*** padding idx before ***', embed_tokens.padding_idx)
self.padding_idx = self.tokenizer.convert_tokens_to_ids(
self.tokenizer.pad_token)
print('*** padding idx after ***', self.padding_idx)
# let's assume each document has at most 128-self.padding_idx-1 sentences
# in case of roberta, it is 126
self.sent_position_embeddings = nn.Embedding(128, embed_dim)
if args.encoder_layers:
self.config.num_hidden_layers = args.encoder_layers
if args.dropout:
self.config.hidden_dropout_prob = args.dropout
if args.attention_dropout:
self.config.attention_probs_dropout_prob = args.attention_dropout
if args.attn_type == 'attn_score':
self.sent_encoder = AttnScoreBertEncoder(self.config)
elif args.attn_type == 'attn_prob':
self.sent_encoder = BertEncoder(self.config)
else:
raise Exception('--attn-type doesn\'t support {} yet !'.format(
args.attn_type))
self.sent_encoder.apply(self._init_weights)
print('*** sentence encoder config ***')
print(self.config)
else:
raise Exception(
'--sentence-transformer-arch doesn\'t support {} yet!'.format(
args.sentence_transformer_arch))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--data_dir",
default='/hdd/lujunyu/dataset/multi_turn_corpus/ubuntu/',
type=str,
required=False,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
parser.add_argument("--task_name",
default='ubuntu',
type=str,
required=False,
help="The name of the task to train.")
parser.add_argument("--output_dir",
default='/hdd/lujunyu/model/ubuntu_roberta_new/',
type=str,
required=False,
help="The output directory where the model checkpoints will be written.")
parser.add_argument("--init_checkpoint",
default='/hdd/lujunyu/model/ubuntu_roberta_new/model.pt',
type=str,
help="Initial checkpoint (usually from a pre-trained BERT model).")
## Other parameters
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_lower_case",
default=False,
action='store_true',
help="Whether to lower case the input text. True for uncased models, False for cased models.")
parser.add_argument("--max_seq_length",
default=256,
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=750,
type=int,
help="Total batch size for eval.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
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:
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')
bert_config = RobertaConfig.from_pretrained('roberta-base', num_labels=2, type_vocab_size=2)
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
test_dataset = UbuntuDatasetForRoberta(
file_path=os.path.join(args.data_dir, "test.txt"),
max_seq_length=args.max_seq_length,
tokenizer=tokenizer
)
test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=args.eval_batch_size,
sampler=SequentialSampler(test_dataset), num_workers=8)
state_dict = torch.load(args.init_checkpoint, map_location='cpu')
model = RobertaForSequenceClassification.from_pretrained(args.init_checkpoint, config=bert_config)
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
logger.info("***** Running testing *****")
logger.info(" Num examples = %d", len(test_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
f = open(os.path.join(args.output_dir, 'logits_test.txt'), 'w')
model.eval()
test_loss = 0
nb_test_steps, nb_test_examples = 0, 0
for input_ids, segment_ids, label_ids in tqdm(test_dataloader, desc="Step"):
input_ids = input_ids.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
tmp_test_loss, logits = model(input_ids, token_type_ids=segment_ids, labels=label_ids)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
for logit, label in zip(logits, label_ids):
logit = '{},{}'.format(logit[0], logit[1])
f.write('_\t{}\t{}\n'.format(logit, label))
test_loss += tmp_test_loss.mean().item()
nb_test_examples += input_ids.size(0)
nb_test_steps += 1
f.close()
test_loss = test_loss / nb_test_steps
result = evaluate(os.path.join(args.output_dir, 'logits_test.txt'))
result.update({'test_loss':test_loss})
output_eval_file = os.path.join(args.output_dir, "results_test.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Test results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
def main():
parser = ArgumentParser()
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("--model_type", default='roberta', type=str, required=True,
help="Model type selected in the list")
parser.add_argument("--model_name_or_path", default='roberta-large', type=str, required=True,
help="Path to pre-trained model or shortcut name selected in the list: ")
parser.add_argument("--task_name", default=None, type=str, required=True,
help="The name of the task to train.")
parser.add_argument("--comment", default='', type=str,
help="The comment")
parser.add_argument('--output_dir', type=Path, default="output")
parser.add_argument("--restore", type=bool, default=True, help="Whether restore from the last checkpoint, is nochenckpoints, start from scartch")
parser.add_argument("--max_seq_length", type=int, default=256, help="max lenght of token sequence")
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", type=bool, default=False,
help="Rul evaluation during training at each logging step.")
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--adapter_transformer_layers", default=2, type=int,
help="The transformer layers of adapter.")
parser.add_argument("--adapter_size", default=768, type=int,
help="The hidden size of adapter.")
parser.add_argument("--adapter_list", default="0,11,23", type=str,
help="The layer where add an adapter")
parser.add_argument("--adapter_skip_layers", default=0, type=int,
help="The skip_layers of adapter according to bert layers")
parser.add_argument('--meta_adapter_model', type=str, help='the pretrained adapter model')
parser.add_argument("--per_gpu_train_batch_size", default=32, type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size", default=64, 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=3e-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-8, 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=3, type=int,
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_steps", default=0, type=int,
help="Linear warmup over warmup_steps.")
parser.add_argument('--logging_steps', type=int, default=10,
help="How often do we snapshot losses, for inclusion in the progress dump? (0 = disable)")
parser.add_argument('--save_steps', type=int, default=1000,
help="Save checkpoint every X updates steps.")
parser.add_argument('--eval_steps', type=int, default=None,
help="eval every X updates steps.")
parser.add_argument('--max_save_checkpoints', type=int, default=500,
help="The max amounts of checkpoint saving. Bigger than it will delete the former checkpoints")
parser.add_argument("--eval_all_checkpoints", action='store_true',
help="Evaluate all checkpoints starting with the same prefix as model_name 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.")
parser.add_argument('--negative_sample', type=int, default=0, help='how many negative samples to select')
# args
args = parser.parse_args()
args.adapter_list = args.adapter_list.split(',')
args.adapter_list = [int(i) for i in args.adapter_list]
name_prefix = 'maxlen-'+str(args.max_seq_length)+'_'+'epoch-'+str(args.num_train_epochs)+'_'+'batch-'+str(args.per_gpu_train_batch_size)+'_'+'lr-'+str(args.learning_rate)+'_'+'warmup-'+str(args.warmup_steps)+'_'+str(args.comment)
args.my_model_name = args.task_name+'_'+name_prefix
args.output_dir = os.path.join(args.output_dir, args.my_model_name)
if args.eval_steps is None:
args.eval_steps = args.save_steps*2
# 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, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
# Set seed
set_seed(args)
args.output_mode = output_modes[args.task_name]
processor = processors[args.task_name]()
label_list = processor.get_labels()
num_labels = len(label_list)
# Load pretrained model and tokenizer
if args.local_rank not in [-1, 0]:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
tokenizer = RobertaTokenizer.from_pretrained('roberta-large')
config = RobertaConfig.from_pretrained('roberta-large', output_attentions=True)
pretrained_model = PretrainedModel()
adapter_model = AdapterModel(args, pretrained_model.config)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
pretrained_model.to(args.device)
adapter_model.to(args.device)
model = (pretrained_model, adapter_model)
logger.info("Training/evaluation parameters %s", args)
val_dataset = load_and_cache_examples(args, args.task_name, tokenizer, 'dev', evaluate=True)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, 'train', evaluate=False)
global_step, tr_loss = train(args, train_dataset, val_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
logger.info("Saving model checkpoint to %s", args.output_dir)
model_to_save = model.module if hasattr(adapter_model, 'module') else adapter_model # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--data_dir",
default='/hdd/lujunyu/dataset/multi_turn_corpus/ubuntu/',
type=str,
required=False,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
parser.add_argument("--task_name",
default='ubuntu',
type=str,
required=False,
help="The name of the task to train.")
parser.add_argument("--output_dir",
default='/hdd/lujunyu/model/chatbert/check/',
type=str,
required=False,
help="The output directory where the model checkpoints will be written.")
## Other parameters
parser.add_argument("--data_augmentation",
default=False,
action='store_true',
help="Whether to use augmentation")
parser.add_argument("--max_seq_length",
default=256,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=True,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_test",
default=True,
action='store_true',
help="Whether to run eval on the test set.")
parser.add_argument("--train_batch_size",
default=400,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=100,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=20.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--warmup_steps",
default=0.0,
type=float,
help="Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--weight_decay",
default=1e-3,
type=float,
help="weight_decay")
parser.add_argument("--save_checkpoints_steps",
default=3125,
type=int,
help="How often to save the model checkpoint.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--gradient_accumulation_steps',
type=int,
default=5,
help="Number of updates steps to accumualte before performing a backward/update pass.")
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:
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 %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError("At least one of `do_train` or `do_eval` must be True.")
bert_config = RobertaConfig.from_pretrained('roberta-base', num_labels=2)
if args.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"Cannot use sequence length {} because the BERT model was only trained up to sequence length {}".format(
args.max_seq_length, bert_config.max_position_embeddings))
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
if args.do_train:
raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
else:
os.makedirs(args.output_dir, exist_ok=True)
tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
if args.data_augmentation:
train_dataset = UbuntuDatasetForRoberta(
file_path=os.path.join(args.data_dir, "train_augment_ubuntu.txt"),
max_seq_length=args.max_seq_length,
tokenizer=tokenizer
)
else:
train_dataset = UbuntuDatasetForRoberta(
file_path=os.path.join(args.data_dir, "train.txt"),
max_seq_length=args.max_seq_length,
tokenizer=tokenizer
)
eval_dataset = UbuntuDatasetForRoberta(
file_path=os.path.join(args.data_dir, "valid.txt"), ### TODO:change
max_seq_length=args.max_seq_length,
tokenizer=tokenizer
)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=args.train_batch_size,
sampler=RandomSampler(train_dataset), num_workers=8)
eval_dataloader = torch.utils.data.DataLoader(eval_dataset, batch_size=args.eval_batch_size,
sampler=SequentialSampler(eval_dataset), num_workers=8)
model = RobertaForSequenceClassification.from_pretrained('roberta-base',config=bert_config)
model.to(device)
num_train_steps = None
if args.do_train:
num_train_steps = int(
len(train_dataset) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# remove pooler, which is not used thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer]
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': args.weight_decay}, {
'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)
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_steps)
else:
optimizer = None
scheduler = None
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
for name, param in model.named_parameters():
if param.requires_grad:
print(name, param.data)
global_step = 0
best_metric = 0.0
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_steps)
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, label_ids = batch
loss, _ = model(input_ids, labels=label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
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:
optimizer.step() # We have accumulated enought gradients
scheduler.step()
model.zero_grad()
global_step += 1
if (step + 1) % args.save_checkpoints_steps == 0:
model.eval()
f = open(os.path.join(args.output_dir, 'logits_dev.txt'), 'w')
eval_loss = 0
nb_eval_steps, nb_eval_examples = 0, 0
logits_all = []
for input_ids, label_ids in eval_dataloader:
input_ids = input_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss, logits = model(input_ids, labels=label_ids)
logits = logits.detach().cpu().numpy()
logits_all.append(logits)
label_ids = label_ids.cpu().numpy()
for logit, label in zip(logits, label_ids):
logit = '{},{}'.format(logit[0], logit[1])
f.write('_\t{}\t{}\n'.format(logit, label))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
f.close()
logits_all = np.concatenate(logits_all,axis=0)
eval_loss = eval_loss / nb_eval_steps
result = evaluate(os.path.join(args.output_dir, 'logits_dev.txt'))
result.update({'eval_loss': eval_loss})
output_eval_file = os.path.join(args.output_dir, "eval_results_dev.txt")
with open(output_eval_file, "a") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
### Save the best checkpoint
if best_metric < result['R10@1'] + result['R10@2']:
try: ### Remove 'module' prefix when using DataParallel
state_dict = model.module.state_dict()
except AttributeError:
state_dict = model.state_dict()
torch.save(state_dict, os.path.join(args.output_dir, "model.pt"))
best_metric = result['R10@1'] + result['R10@2']
logger.info('Saving the best model in {}'.format(os.path.join(args.output_dir, "model.pt")))
### visualize bad cases of the best model
logger.info('Saving Bad cases...')
visualize_bad_cases(
logits=logits_all,
input_file_path=os.path.join(args.data_dir, 'valid.txt'),
output_file_path=os.path.join(args.output_dir, 'valid_bad_cases.txt')
)
model.train()