def bertForNextSentencePrediction(*args, **kwargs):
"""
BERT model with next sentence prediction head.
This module comprises the BERT model followed by the next sentence
classification head.
Example:
# Load the tokenizer
>>> import torch
>>> tokenizer = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertTokenizer', 'bert-base-cased', do_basic_tokenize=False)
# Prepare tokenized input
>>> text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
>>> tokenized_text = tokenizer.tokenize(text)
>>> indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
>>> segments_ids = [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1]
>>> tokens_tensor = torch.tensor([indexed_tokens])
>>> segments_tensors = torch.tensor([segments_ids])
# Load bertForNextSentencePrediction
>>> model = torch.hub.load('huggingface/pytorch-pretrained-BERT', 'bertForNextSentencePrediction', 'bert-base-cased')
>>> model.eval()
# Predict the next sentence classification logits
>>> with torch.no_grad():
next_sent_classif_logits = model(tokens_tensor, segments_tensors)
"""
model = BertForNextSentencePrediction.from_pretrained(*args, **kwargs)
return model
def bertForNextSentencePrediction(*args, **kwargs):
"""
BERT model with next sentence prediction head.
This module comprises the BERT model followed by the next sentence
classification head.
"""
model = BertForNextSentencePrediction.from_pretrained(*args, **kwargs)
return model
def test_BertForNextSentencePrediction():
input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])
input_mask = torch.LongTensor([[1, 1, 1], [1, 1, 0]])
token_type_ids = torch.LongTensor([[0, 0, 1], [0, 1, 0]])
config = BertConfig(vocab_size_or_config_json_file=32000,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072)
model = BertForNextSentencePrediction(config)
print(model(input_ids, token_type_ids, input_mask))
def __init__(
self
, question_file
, answer_file
, sample_n
, random_state
, bert_cache
, logger
, device
, max_seq_length
, batch_size
):
# initializes some vars
self.question_file = question_file
self.answer_file = answer_file
self.sample_n = sample_n
self.random_state = random_state
self.bert_cache = bert_cache
self.logger = logger
self.device = device
self.max_seq_length = max_seq_length
self.batch_size = batch_size
# gets the pre-trained tokenizer
self.tokenizer = BertTokenizer.from_pretrained(
"bert-base-uncased"
, do_lower_case = True
, cache_dir = self.bert_cache
)
# gets the pre-trained model
self.model = BertForNextSentencePrediction.from_pretrained(
"bert-base-uncased"
, cache_dir = self.bert_cache
).to(self.device)
# instantiates the helper class
self.ceshiner = Ceshiner()
if label == pred_label:
correct += 1
print(
f'Accuracy: {correct}/{len(labels)} = {correct/len(labels)*100:.2f}%')
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='CSQA using BERT NSP model')
parser.add_argument('--input', help='input dataset')
parser.add_argument('--bert-vocab', help='bert vocab file')
parser.add_argument('--bert-model', help='pretrained bert model')
parser.add_argument('--batch-size',
type=int,
default=8,
help='batch size for BERT')
parser.add_argument('--gpu-id', '-g', type=int, default=0, help='GPU ID')
args = parser.parse_args()
print('Initialize BERT model...')
TOKENIZER = WordTokenizer(word_splitter=BertBasicWordSplitter())
WORD_INDEXER = PretrainedBertIndexer(pretrained_model=args.bert_vocab)
VOCAB = Vocabulary()
GPU_ID = args.gpu_id
BERT_NEXT_SENTENCE = BertForNextSentencePrediction.from_pretrained(
args.bert_model).to(torch.device(f"cuda:{GPU_ID}"))
BERT_NEXT_SENTENCE.eval()
main()
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default="../data/bert",
type=str,
required=False,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default='bert-base-uncased',
type=str,
required=False,
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("--task_name",
default='MRPC',
type=str,
required=False,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default="../data/bert/output",
type=str,
required=False,
help="The output directory where the model checkpoints will be written."
)
# Other parameters
parser.add_argument(
"--max_seq_length",
default=192,
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_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=True,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=128,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=128,
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=2.0,
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",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
parser.add_argument(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
args = parser.parse_args()
processors = {"mrpc": MrpcProcessor}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
# if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
# raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForNextSentencePrediction.from_pretrained(args.bert_model)
# model = BertForNextSentencePrediction.from_pretrained(args.bert_model,
# cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [
(n, param.clone().detach().to('cpu').float().requires_grad_())
for n, param in model.named_parameters()
]
elif args.optimize_on_cpu:
param_optimizer = [(n,
param.clone().detach().to('cpu').requires_grad_())
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
if param.grad is not None:
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info(
"FP16 TRAINING: Nan in gradients, reducing loss scaling"
)
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
model.eval()
torch.save(model.state_dict(),
"../data/models/base-uncased-192-2of9-ep%s.pt" % epoch)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': tr_loss / nb_tr_steps
}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w", encoding='utf-8') 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])))
from pytorch_pretrained_bert.modeling import BertConfig, BertForNextSentencePrediction
# from next_sentence.processing import create_features, read_file, MAX_LENGTH
from utils.utils import features_translation, tokenizer
SOURCE_PATH = '/home/gump/Software/pycharm-2018.1.6/projects/bert-for-classificaion/' \
'next_sentence/data/'
CONFIG_FILE = SOURCE_PATH+'model/model_2.2.2/config_4.json'
MODEL_FILE = SOURCE_PATH+'model/model_2.2.2/model_4.bin'
TEST_FILE = '/home/gump/Software/pycharm-2018.1.6/projects/bert-for-classificaion/' \
'next_sentence/data/chat_bot_2.2.0.csv'
MAX_LENGTH = 30
# load model
config = BertConfig(CONFIG_FILE)
model = BertForNextSentencePrediction(config)
model.load_state_dict(torch.load(MODEL_FILE))
if torch.cuda.is_available():
model.cuda()
model.eval()
# load data
tokenizer = tokenizer()
def predict(text_a, text_b):
token_a = tokenizer.tokenize(text_a)
token_b = tokenizer.tokenize(text_b)
token_text = ['[CLS]'] + token_a + ['[SEP]'] + token_b + ['[SEP]']
tokens_ids = tokenizer.convert_tokens_to_ids(token_text)
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
default="../data/bert",
type=str,
required=False,
help=
"The input data dir. Should contain the _p.tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default='bert-base-uncased',
type=str,
required=False,
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("--task_name",
default='MRPC',
type=str,
required=False,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default="../data/bert/output",
type=str,
required=False,
help="The output directory where the model checkpoints will be written."
)
# Other parameters
parser.add_argument(
"--max_seq_length",
default=192,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_predict",
default=True,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=True,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=128,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=128,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
parser.add_argument(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
parser.add_argument('--load_model_path',
default='../data/models/base-uncased-192-2of9-ep1.pt',
help='Load model for prediction')
args = parser.parse_args()
processors = {"mrpc": MrpcProcessor}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_eval` or `do_predict` must be True."
)
# if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
# raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForNextSentencePrediction.from_pretrained(args.bert_model)
# model = BertForNextSentencePrediction.from_pretrained(args.bert_model,
# cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
print('model loading')
model.load_state_dict(torch.load(args.load_model_path))
print('model loaded')
# Prepare optimizer
if args.fp16:
param_optimizer = [
(n, param.clone().detach().to('cpu').float().requires_grad_())
for n, param in model.named_parameters()
]
elif args.optimize_on_cpu:
param_optimizer = [(n,
param.clone().detach().to('cpu').requires_grad_())
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_predict and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_test_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
count = 0
predictions = [0]
for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy() #predicted label
logits = logits[:, 1]
predictions.extend(logits)
count += args.eval_batch_size
print(count / len(eval_examples) * 100)
data = pd.read_csv('../data/bert/eval2_unlabelled_p.tsv',
sep='\t',
header=None)
data[len(data.columns)] = predictions
data.to_csv('../data/pre-answer.tsv',
sep='\t',
header=False,
index=False)
# Converting to submission format
data = pd.read_csv('../data/pre-answer.tsv',
sep='\t',
names=[
'query_id', 'query_text', 'passage_text',
'passage_id', 'cs'
])
uniq, index = np.unique(data['query_id'], return_index=True)
query_id = uniq[index.argsort()]
scores = data['cs'].values.reshape(-1, 10)
print(scores.shape)
answer = np.column_stack((query_id, scores))
answer = pd.DataFrame(answer)
answer.iloc[:, 0] = answer.iloc[:, 0].astype('int')
answer.to_csv('../data/answer.tsv', sep='\t', header=None, index=False)
def main():
parser = argparse.ArgumentParser()
#files arguments
parser.add_argument("-model_name",
default="classifacation_2",
help="Name of the model. To be used in filename")
parser.add_argument("-model_group",
default="baseline",
help="dir to put model in")
parser.add_argument("-old_model_name",
default=None,
help="filename of model to be loaded")
parser.add_argument("-model_version",
default=0,
help="version to be added to the models filename")
parser.add_argument("-bert_model",
default="bert-base-uncased",
help="Bert pre-trained model")
parser.add_argument("-train",
default=True,
help="true if you want to train the model")
parser.add_argument("-val",
default=True,
help="true if you want to evaluate the model")
parser.add_argument("-train_batch_size",
default=65,
help="batch size for training dataset")
parser.add_argument("-val_batch_size",
default=10,
help="batch size for validation dataset")
parser.add_argument("-max_len",
default=60,
help="max length for input sequence")
parser.add_argument("-learning_rate",
default=0.00005,
help="learning weight for optimization")
parser.add_argument("-gradient_accumulation_steps", default=1, help="")
parser.add_argument("-num_epochs",
default=2,
help="number of epochs for training, and validation")
parser.add_argument("-num_outpout_checkpoints_train",
default=-2,
help="determines how many times the loss is outputed "
"during training")
parser.add_argument("-num_outpout_checkpoints_val",
default=1,
help="determines how many times metrics are outputed "
"during training")
parser.add_argument(
"-warmup_proportion",
default=0.1,
help="Proportion of training to perform linear learning"
" rate warmup for.")
args = parser.parse_args()
args.project_file = os.getcwd()
args.dataset_path = "{}/data/".format(args.project_file)
args.output_dir = '{}/classification/{}/{}_{}/'.format(
args.project_file, args.model_group, args.model_name,
args.model_version)
os.makedirs(args.output_dir, exist_ok=True)
# check_files(file)
use_old_model = args.old_model_name is not None
if use_old_model:
args.old_model_filename = '{}/classification/{}/{}' \
'/'.format(args.project_file,
args.model_group, args.old_model_name)
args = load_args("{}args.json".format(args.old_model_filename), args)
model = cuda(BertForNextSentencePrediction.from_pretrained(
args.bert_model))
phases = []
data_loaders = []
#loads features from file that is created in make_features.py
#it takes a long time to create features which is why there is a seperate
#file
args.features_path = "{}/classification/features/max_{}/".format(
args.project_file, args.max_len)
if not os.path.exists(args.features_path):
raise ValueError(
"you must create features with classification/make_features.py "
"prior to running this model.\n need file: {}".format(
args.features_path))
#get train features, and make optimizer
if args.train:
train_features = load_features("{}train.pkl".format(
args.features_path))
dataloader_train = make_dataloader(train_features,
args.train_batch_size)
data_loaders.append(dataloader_train)
phases.append('train')
num_train_steps = int(
len(train_features) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_epochs)
if args.num_outpout_checkpoints_train < 0:
args.num_outpout_checkpoints_train = len(train_features) / \
args.train_batch_size / (args.num_outpout_checkpoints_train * -1)
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
}]
t_total = num_train_steps
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
args.train_len = len(train_features)
else:
optimizer = None
# get val examples
if args.val:
val_features = load_features("{}val.pkl".format(args.features_path))
dataloader_val = make_dataloader(val_features, args.val_batch_size)
data_loaders.append(dataloader_val)
phases.append('val')
args.val_len = len(val_features)
#outoput args
string = ""
for k, v in vars(args).items():
string += "{}: {}\n".format(k, v)
print(string)
output = string + '\n'
#load checkpoint
if use_old_model:
model, optimizer = load_checkpoint(
"{}model".format(args.old_model_filename), model, optimizer)
#output results from last model
results = load_results("{}metrics.json".format(
args.old_model_filename))
string = "\n--loaded model--\ntrain loss: {}\nval loss: {}\n" \
"num_epoch: {}\n".format(results[
"average_train_epoch_losses"][-1], results["val_loss"][-1],
results["epoch"])
output += string
print(string)
outfile = open("{}output".format(args.output_dir), 'w')
outfile.write(output)
outfile.close()
with open("{}args.json".format(args.output_dir), 'w') as fp:
json.dump(vars(args), fp, indent=4, sort_keys=True)
best_model = model
best_optimizer = optimizer
metrics = {
"accuracy": [],
"precision": [],
"recall": [],
"f1": [],
"lowest_loss": 100,
"average_train_epoch_losses": [],
"train_epoch_losses": [],
"val_loss": [],
"best_epoch": 0
}
highest_acc = 0
for epoch in range(0, args.num_epochs):
start = time.clock()
string = 'Epoch: {}\n'.format(epoch)
print(string, end='')
output = output + '\n' + string
metrics["epoch"] = epoch
#if epoch == 6:
# model.unfreeze_embeddings()
# parameters = list(model.parameters())
# optimizer = torch.optim.Adam(
# parameters, amsgrad=True, weight_decay=weight_decay)
#use when you validate before training, and what to validate on last epoch
#if epoch == params["nb_epochs"] -1 and params["val"] and params["train"]:
# phases.append('val')
# data_loaders.append(dataloader_val)
for phase, data_loader in zip(phases, data_loaders):
if phase == 'train':
model.train()
intervals = args.num_outpout_checkpoints_train
string = '--Train-- \n'
else:
model.eval()
intervals = args.num_outpout_checkpoints_val
string = '--Validation-- \n'
print(string, end='')
output = output + '\n' + string
epoch_loss = []
epoch_accuracy = []
epoch_precision = []
epoch_recall = []
epoch_f1 = []
j = 1
for i, batch in enumerate(tqdm(data_loader, desc="batch")):
batch = tuple(variable(t) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if phase == 'val':
outputs = model(input_ids, segment_ids, input_mask)
epoch_loss.append(float(loss))
average_epoch_loss = np.mean(epoch_loss)
if phase == 'train':
loss.backward()
#torch.nn.utils.clip_grad_norm_(parameters,
# params["max_grad_norm"])
optimizer.step()
if (len(data_loader) / intervals) * j <= i + 1:
#if len(data_loader) == i + 1:
string = ('Example {:03d} | {} loss: {:.3f}'.format(
i, phase, average_epoch_loss))
#print(string, end='\n')
output = output + string + '\n'
outfile = open("{}output".format(args.output_dir), 'w')
outfile.write(output)
outfile.close()
j += 1
optimizer.zero_grad()
else:
# get result metrics
targets = label_ids.cpu().numpy()
predicted = torch.argmax(outputs.view(-1, 2),
-1).cpu().numpy()
accuracy, precision, recall, f1 = classifier_accuracy(
targets, predicted)
#print('{},{},{},{}'.format(accuracy, precision, recall,
# f1))
epoch_accuracy.append(accuracy)
epoch_precision.append(precision)
epoch_recall.append(recall)
epoch_f1.append(f1)
if (len(data_loader) / intervals) * j <= i + 1:
# if len(data_loader) == i + 1:
string = ('Example {:03d} | {} loss: {:.3f}'.format(
i, phase, average_epoch_loss))
# print(string, end='\n')
output = output + string + '\n'
average_epoch_accuracy = np.mean(epoch_accuracy)
average_epoch_precision = np.mean(epoch_precision)
average_epoch_recall = np.mean(epoch_recall)
average_epoch_f1 = np.mean(epoch_f1)
string = "Accuracy: {:.3f}\nPrecision: {:.3f}\n" \
"Recall: {:.3f}\nF1: {:.3f}\n".format(
average_epoch_accuracy, average_epoch_precision,
average_epoch_recall, average_epoch_f1)
output = output + string + '\n'
outfile = open("{}output".format(args.output_dir), 'w')
outfile.write(output)
outfile.close()
j += 1
# print random sentence
if phase == 'val':
time_taken = time.clock() - start
metrics["val_loss"].append(average_epoch_loss)
string = ' {} loss: {:.3f} | time: {:.3f}'.format(
phase, average_epoch_loss, time_taken)
string += ' | lowest loss: {:.3f} highest accuracy:' \
' {:.3f}'.format(metrics["lowest_loss"], highest_acc)
#print(string, end='\n')
output = output + '\n' + string + '\n'
average_epoch_accuracy = np.mean(epoch_accuracy)
average_epoch_precision = np.mean(epoch_precision)
average_epoch_recall = np.mean(epoch_recall)
average_epoch_f1 = np.mean(epoch_f1)
metrics["accuracy"].append(average_epoch_accuracy),
metrics["precision"].append(average_epoch_precision)
metrics["recall"].append(average_epoch_recall)
metrics["f1"].append(average_epoch_f1)
if average_epoch_loss < metrics["lowest_loss"]:
best_model = model
best_optimizer = optimizer
metrics["best_epoch"] = epoch
metrics["lowest_loss"] = average_epoch_loss
save_checkpoint("{}model".format(args.output_dir), best_model,
best_optimizer, epoch, model, optimizer)
with open("{}metrics.json".format(args.output_dir), 'w') as fp:
json.dump(metrics, fp, indent=4, sort_keys=True)
string = "Accuracy: {:.3f}\nPrecision: {:.3f}\nRecall:" \
" {:.3f}\nF1: {:.3f}\n".format(
average_epoch_accuracy, average_epoch_precision,
average_epoch_recall, average_epoch_f1)
# print(string, end='\n')
output = output + string + '\n'
"""
random_idx = np.random.randint(len(dataset_val))
sentence_1, sentence_2, labels = dataset_val[random_idx]
batch = tuple(variable(t) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
outputs_var = model(sentence_1_var.unsqueeze(0),
sentence_2_var.unsqueeze(0)) # unsqueeze
# to get the batch dimension
outputs = outputs_var.squeeze(0).data.cpu().numpy()
string = '> {}\n'.format(get_sentence_from_indices(
sentence_1, dataset_val.vocab, PairsDataset.EOS_TOKEN))
string = string + u'> {}\n'.format(get_sentence_from_indices(
sentence_2, dataset_val.vocab, PairsDataset.EOS_TOKEN))
string = string + u'target:{}| P false:{:.3f}, P true:' \
u' {:.3f}'.format(targets, float(outputs[0]), float(outputs[1]))
print(string, end='\n\n')
output = output + string + '\n' + '\n'
"""
else:
metrics["average_train_epoch_losses"].append(
average_epoch_loss)
metrics["train_epoch_losses"].append(epoch_loss)
outfile = open("{}output".format(args.output_dir), 'w')
outfile.write(output)
outfile.close()
val_features = create_features(getattr(args, 'val_file'))
val_features_data = features_translation(val_features)
logging.info('create batch data')
train_data = DataLoader(train_features_data,
batch_size=getattr(args, 'batch_size'),
shuffle=True,
drop_last=True)
val_data = DataLoader(val_features_data,
batch_size=getattr(args, 'batch_size'),
shuffle=True,
drop_last=True)
# load model
logging.info('create model')
model = BertForNextSentencePrediction.from_pretrained(BERT_PRETRAINED_PATH,
cache_dir='data/cache')
if args.fp16:
model.half()
if torch.cuda.is_available():
model.cuda()
# optimizer
parameters = list(model.named_parameters())
# parameters = [n for n in parameters if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in parameters if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {