SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Merge branch 'master' into fix-xlnet-squad2.0

Browse Source
This commit is contained in:
Thomas Wolf
2019-12-21 12:48:10 +01:00
committed by GitHub
parent ca99a2d500 8618bf15d6
commit 562f864038
200 changed files with 20341 additions and 3889 deletions
Download Patch File Download Diff File Expand all files Collapse all files

221
examples/run_squad.py
View File

@@ -16,6 +16,8 @@
""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
from __future__ import absolute_import, division, print_function
from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor, SquadResult
from transformers.data.metrics.squad_metrics import compute_predictions_logits, compute_predictions_log_probs, squad_evaluate
import argparse
import logging
@@ -23,11 +25,9 @@ import os
import random
import glob
import timeit
import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)
from torch.utils.data.distributed import DistributedSampler
try:
@@ -43,18 +43,12 @@ from transformers import (WEIGHTS_NAME, BertConfig,
XLMTokenizer, XLNetConfig,
XLNetForQuestionAnswering,
XLNetTokenizer,
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer,
XLMConfig, XLMForQuestionAnswering, XLMTokenizer,
)
from transformers import AdamW, get_linear_schedule_with_warmup
from utils_squad import (read_squad_examples, convert_examples_to_features,
RawResult, write_predictions,
RawResultExtended, write_predictions_extended)
# The follwing import is the official SQuAD evaluation script (2.0).
# You can remove it from the dependencies if you are using this script outside of the library
# We've added it here for automated tests (see examples/test_examples.py file)
from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
logger = logging.getLogger(__name__)
@@ -65,7 +59,8 @@ MODEL_CLASSES = {
'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
}
def set_seed(args):
@@ -98,14 +93,16 @@ def train(args, train_dataset, model, tokenizer):
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': args.weight_decay},
{'params': [p for n, p in model.named_parameters() 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 = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=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.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
@@ -128,22 +125,28 @@ def train(args, train_dataset, model, tokenizer):
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
global_step = 1
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
set_seed(args) # Added here for reproductibility (even between python 2 and 3)
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'start_positions': batch[3],
'end_positions': batch[4]}
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'start_positions': batch[3],
'end_positions': batch[4]
}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[5],
'p_mask': batch[6]})
@@ -175,8 +178,8 @@ def train(args, train_dataset, model, tokenizer):
model.zero_grad()
global_step += 1
# Log metrics
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
@@ -185,8 +188,8 @@ def train(args, train_dataset, model, tokenizer):
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
logging_loss = tr_loss
# Save model checkpoint
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
@@ -215,50 +218,72 @@ def evaluate(args, model, tokenizer, prefix=""):
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# multi-gpu evaluate
if args.n_gpu > 1:
if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
model = torch.nn.DataParallel(model)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
all_results = []
start_time = timeit.default_timer()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {'input_ids': batch[0],
'attention_mask': batch[1]
}
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1]
}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
example_indices = batch[3]
# XLNet and XLM use more arguments for their predictions
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4],
'p_mask': batch[5]})
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
outputs = model(**inputs)
for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id)
if args.model_type in ['xlnet', 'xlm']:
# XLNet uses a more complex post-processing procedure
result = RawResultExtended(unique_id = unique_id,
start_top_log_probs = to_list(outputs[0][i]),
start_top_index = to_list(outputs[1][i]),
end_top_log_probs = to_list(outputs[2][i]),
end_top_index = to_list(outputs[3][i]),
cls_logits = to_list(outputs[4][i]))
output = [to_list(output[i]) for output in outputs]
# Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
# models only use two.
if len(output) >= 5:
start_logits = output[0]
start_top_index = output[1]
end_logits = output[2]
end_top_index = output[3]
cls_logits = output[4]
result = SquadResult(
unique_id, start_logits, end_logits,
start_top_index=start_top_index,
end_top_index=end_top_index,
cls_logits=cls_logits
)
else:
result = RawResult(unique_id = unique_id,
start_logits = to_list(outputs[0][i]),
end_logits = to_list(outputs[1][i]))
start_logits, end_logits = output
result = SquadResult(
unique_id, start_logits, end_logits
)
all_results.append(result)
evalTime = timeit.default_timer() - start_time
@@ -267,63 +292,84 @@ def evaluate(args, model, tokenizer, prefix=""):
# Compute predictions
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
if args.version_2_with_negative:
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
else:
output_null_log_odds_file = None
# XLNet and XLM use a more complex post-processing procedure
if args.model_type in ['xlnet', 'xlm']:
# XLNet uses a more complex post-processing procedure
write_predictions_extended(examples, features, all_results, args.n_best_size,
start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
args.max_answer_length, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.predict_file,
model.config.start_n_top, model.config.end_n_top,
output_nbest_file, output_null_log_odds_file,
start_n_top, end_n_top,
args.version_2_with_negative, tokenizer, args.verbose_logging)
else:
write_predictions(examples, features, all_results, args.n_best_size,
predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
args.max_answer_length, args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative, args.null_score_diff_threshold)
# Evaluate with the official SQuAD script
evaluate_options = EVAL_OPTS(data_file=args.predict_file,
pred_file=output_prediction_file,
na_prob_file=output_null_log_odds_file)
results = evaluate_on_squad(evaluate_options)
# Compute the F1 and exact scores.
results = squad_evaluate(examples, predictions)
return results
def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# Load data features from cache or dataset file
input_file = args.predict_file if evaluate else args.train_file
cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
input_dir = args.data_dir if args.data_dir else "."
cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
'dev' if evaluate else 'train',
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length)))
str(args.max_seq_length))
)
# Init features and dataset from cache if it exists
if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
features_and_dataset = torch.load(cached_features_file)
features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
else:
logger.info("Creating features from dataset file at %s", input_file)
examples = read_squad_examples(input_file=input_file,
is_training=not evaluate,
version_2_with_negative=args.version_2_with_negative)
features = convert_examples_to_features(examples=examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=not evaluate,
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
pad_token_segment_id=3 if args.model_type in ['xlnet'] else 0,
cls_token_at_end=True if args.model_type in ['xlnet'] else False,
sequence_a_is_doc=True if args.model_type in ['xlnet'] else False)
logger.info("Creating features from dataset file at %s", input_dir)
if not args.data_dir and ((evaluate and not args.predict_file) or (not evaluate and not args.train_file)):
try:
import tensorflow_datasets as tfds
except ImportError:
raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
if args.version_2_with_negative:
logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")
tfds_examples = tfds.load("squad")
examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
else:
processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
if evaluate:
examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
else:
examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
features, dataset = squad_convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=not evaluate,
return_dataset='pt'
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, cached_features_file)
torch.save({"features": features, "dataset": dataset}, cached_features_file)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
@@ -355,10 +401,6 @@ def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_file", default=None, type=str, required=True,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument("--predict_file", default=None, type=str, required=True,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
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,
@@ -367,6 +409,15 @@ def main():
help="The output directory where the model checkpoints and predictions will be written.")
## Other parameters
parser.add_argument("--data_dir", default=None, type=str,
help="The input data dir. Should contain the .json files for the task." +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--train_file", default=None, type=str,
help="The input training file. If a data dir is specified, will look for the file there" +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--predict_file", default=None, type=str,
help="The input evaluation file. If a data dir is specified, will look for the file there" +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--config_name", default="", type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument("--tokenizer_name", default="", type=str,
@@ -405,7 +456,7 @@ def main():
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("--weight_decay", default=0.0, type=float,
help="Weight deay if we apply some.")
help="Weight decay 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,
@@ -540,7 +591,7 @@ def main():
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
model = model_class.from_pretrained(args.output_dir, force_download=True)
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
model.to(args.device)
@@ -548,17 +599,23 @@ def main():
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
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 model loading logs
if args.do_train:
logger.info("Loading checkpoints saved during training for evaluation")
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 model loading logs
else:
logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
checkpoints = [args.model_name_or_path]
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
# Reload the model
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
model = model_class.from_pretrained(checkpoint)
model = model_class.from_pretrained(checkpoint, force_download=True)
model.to(args.device)
# Evaluate