Skip to content

GitLab

Unverified Commit 3c1b6f59 authored by Julien Chaumond's avatar Julien Chaumond Committed by GitHub
Browse files

Merge branch 'master' into fix_top_k_top_p_filtering

parents a9f24a16 fa735208
Show whitespace changes
Inline Side-by-side
Showing with 2128 additions and 308 deletions
examples/distillation/scripts/extract_for_distil.py → examples/distillation/scripts/extract_distilbert.py
View file @ 3c1b6f59
......@@ -14,14 +14,15 @@
# limitations under the License.
"""
Preprocessing script before training DistilBERT.
Specific to BERT -> DistilBERT.
"""
from pytorch_transformers import BertForMaskedLM, RobertaForMaskedLM
from transformers import BertForMaskedLM, RobertaForMaskedLM
import torch
import argparse
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="Extraction some layers of the full BertForMaskedLM or RObertaForMaskedLM for Transfer Learned Distillation")
parser.add_argument("--model_type", default="bert", choices=["bert", "roberta"])
parser.add_argument("--model_type", default="bert", choices=["bert"])
parser.add_argument("--model_name", default='bert-base-uncased', type=str)
parser.add_argument("--dump_checkpoint", default='serialization_dir/tf_bert-base-uncased_0247911.pth', type=str)
parser.add_argument("--vocab_transform", action='store_true')
......@@ -31,9 +32,8 @@ if __name__ == '__main__':
if args.model_type == 'bert':
model = BertForMaskedLM.from_pretrained(args.model_name)
prefix = 'bert'
elif args.model_type == 'roberta':
model = RobertaForMaskedLM.from_pretrained(args.model_name)
prefix = 'roberta'
else:
raise ValueError(f'args.model_type should be "bert".')
state_dict = model.state_dict()
compressed_sd = {}
......@@ -68,20 +68,12 @@ if __name__ == '__main__':
state_dict[f'{prefix}.encoder.layer.{teacher_idx}.output.LayerNorm.{w}']
std_idx += 1
if args.model_type == 'bert':
compressed_sd[f'vocab_projector.weight'] = state_dict[f'cls.predictions.decoder.weight']
compressed_sd[f'vocab_projector.bias'] = state_dict[f'cls.predictions.bias']
if args.vocab_transform:
for w in ['weight', 'bias']:
compressed_sd[f'vocab_transform.{w}'] = state_dict[f'cls.predictions.transform.dense.{w}']
compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'cls.predictions.transform.LayerNorm.{w}']
elif args.model_type == 'roberta':
compressed_sd[f'vocab_projector.weight'] = state_dict[f'lm_head.decoder.weight']
compressed_sd[f'vocab_projector.bias'] = state_dict[f'lm_head.bias']
if args.vocab_transform:
for w in ['weight', 'bias']:
compressed_sd[f'vocab_transform.{w}'] = state_dict[f'lm_head.dense.{w}']
compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'lm_head.layer_norm.{w}']
print(f'N layers selected for distillation: {std_idx}')
print(f'Number of params transfered for distillation: {len(compressed_sd.keys())}')
......
examples/distillation/scripts/token_counts.py
View file @ 3c1b6f59
......@@ -13,7 +13,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Preprocessing script before training DistilBERT.
Preprocessing script before training the distilled model.
"""
from collections import Counter
import argparse
......
examples/distillation/train.py
View file @ 3c1b6f59
......@@ -13,7 +13,8 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Training DistilBERT.
Training the distilled model.
Supported architectures include: BERT -> DistilBERT, RoBERTa -> DistilRoBERTa, GPT2 -> DistilGPT2.
"""
import os
import argparse
......@@ -23,68 +24,96 @@ import shutil
import numpy as np
import torch
from pytorch_transformers import BertTokenizer, BertForMaskedLM, RobertaTokenizer, RobertaForMaskedLM
from pytorch_transformers import DistilBertForMaskedLM, DistilBertConfig
from transformers import BertConfig, BertForMaskedLM, BertTokenizer
from transformers import RobertaConfig, RobertaForMaskedLM, RobertaTokenizer
from transformers import DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer
from transformers import GPT2Config, GPT2LMHeadModel, GPT2Tokenizer
from distiller import Distiller
from utils import git_log, logger, init_gpu_params, set_seed
from dataset import Dataset
from lm_seqs_dataset import LmSeqsDataset
MODEL_CLASSES = {
'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer)
}
def sanity_checks(args):
"""
A bunch of args sanity checks to perform even starting...
"""
assert (args.mlm and args.alpha_mlm > 0.) or (not args.mlm and args.alpha_mlm == 0.)
assert (args.alpha_mlm > 0. and args.alpha_clm == 0.) or (args.alpha_mlm == 0. and args.alpha_clm > 0.)
if args.mlm:
assert os.path.isfile(args.token_counts)
assert (args.student_type in ['roberta', 'distilbert']) and (args.teacher_type in ['roberta', 'bert'])
else:
assert (args.student_type in ['gpt2']) and (args.teacher_type in ['gpt2'])
assert args.teacher_type == args.student_type or (args.student_type=='distilbert' and args.teacher_type=='bert')
assert os.path.isfile(args.student_config)
if args.student_pretrained_weights is not None:
assert os.path.isfile(args.student_pretrained_weights)
if args.freeze_token_type_embds: assert args.student_type in ['roberta']
assert args.alpha_ce >= 0.
assert args.alpha_mlm >= 0.
assert args.alpha_clm >= 0.
assert args.alpha_mse >= 0.
assert args.alpha_cos >= 0.
assert args.alpha_ce + args.alpha_mlm + args.alpha_clm + args.alpha_mse + args.alpha_cos > 0.
def freeze_pos_embeddings(student, args):
if args.student_type == 'roberta':
student.roberta.embeddings.position_embeddings.weight.requires_grad = False
elif args.student_type == 'gpt2':
student.transformer.wpe.weight.requires_grad = False
def freeze_token_type_embeddings(student, args):
if args.student_type == 'roberta':
student.roberta.embeddings.token_type_embeddings.weight.requires_grad = False
def main():
parser = argparse.ArgumentParser(description="Training")
parser.add_argument("--force", action='store_true',
help="Overwrite dump_path if it already exists.")
parser.add_argument("--dump_path", type=str, required=True,
help="The output directory (log, checkpoints, parameters, etc.)")
parser.add_argument("--data_file", type=str, required=True,
help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.")
parser.add_argument("--token_counts", type=str, required=True,
help="The token counts in the data_file for MLM.")
parser.add_argument("--force", action='store_true',
help="Overwrite dump_path if it already exists.")
parser.add_argument("--vocab_size", default=30522, type=int,
help="The vocabulary size.")
parser.add_argument("--max_position_embeddings", default=512, type=int,
help="Maximum sequence length we can model (including [CLS] and [SEP]).")
parser.add_argument("--sinusoidal_pos_embds", action='store_false',
help="If true, the position embeddings are simply fixed with sinusoidal embeddings.")
parser.add_argument("--n_layers", default=6, type=int,
help="Number of Transformer blocks.")
parser.add_argument("--n_heads", default=12, type=int,
help="Number of heads in the self-attention module.")
parser.add_argument("--dim", default=768, type=int,
help="Dimension through the network. Must be divisible by n_heads")
parser.add_argument("--hidden_dim", default=3072, type=int,
help="Intermediate dimension in the FFN.")
parser.add_argument("--dropout", default=0.1, type=float,
help="Dropout.")
parser.add_argument("--attention_dropout", default=0.1, type=float,
help="Dropout in self-attention.")
parser.add_argument("--activation", default='gelu', type=str,
help="Activation to use in self-attention")
parser.add_argument("--tie_weights_", action='store_false',
help="If true, we tie the embeddings matrix with the projection over the vocabulary matrix. Default is true.")
parser.add_argument("--from_pretrained_weights", default=None, type=str,
parser.add_argument("--student_type", type=str, choices=["distilbert", "roberta", "gpt2"], required=True,
help="The student type (DistilBERT, RoBERTa).")
parser.add_argument("--student_config", type=str, required=True,
help="Path to the student configuration.")
parser.add_argument("--student_pretrained_weights", default=None, type=str,
help="Load student initialization checkpoint.")
parser.add_argument("--from_pretrained_config", default=None, type=str,
help="Load student initialization architecture config.")
parser.add_argument("--teacher_type", default="bert", choices=["bert", "roberta"],
parser.add_argument("--teacher_type", choices=["bert", "roberta", "gpt2"], required=True,
help="Teacher type (BERT, RoBERTa).")
parser.add_argument("--teacher_name", default="bert-base-uncased", type=str,
parser.add_argument("--teacher_name", type=str, required=True,
help="The teacher model.")
parser.add_argument("--temperature", default=2., type=float,
help="Temperature for the softmax temperature.")
parser.add_argument("--alpha_ce", default=0.5, type=float,
help="Linear weight for the distillation loss. Must be >=0.")
parser.add_argument("--alpha_mlm", default=0.5, type=float,
help="Linear weight for the MLM loss. Must be >=0.")
parser.add_argument("--alpha_mlm", default=0.0, type=float,
help="Linear weight for the MLM loss. Must be >=0. Should be used in coonjunction with `mlm` flag.")
parser.add_argument("--alpha_clm", default=0.5, type=float,
help="Linear weight for the CLM loss. Must be >=0.")
parser.add_argument("--alpha_mse", default=0.0, type=float,
help="Linear weight of the MSE loss. Must be >=0.")
parser.add_argument("--alpha_cos", default=0.0, type=float,
help="Linear weight of the cosine embedding loss. Must be >=0.")
parser.add_argument("--mlm", action="store_true",
help="The LM step: MLM or CLM. If `mlm` is True, the MLM is used over CLM.")
parser.add_argument("--mlm_mask_prop", default=0.15, type=float,
help="Proportion of tokens for which we need to make a prediction.")
parser.add_argument("--word_mask", default=0.8, type=float,
......@@ -95,17 +124,20 @@ def main():
help="Proportion of tokens to randomly replace.")
parser.add_argument("--mlm_smoothing", default=0.7, type=float,
help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).")
parser.add_argument("--token_counts", type=str,
help="The token counts in the data_file for MLM.")
parser.add_argument("--restrict_ce_to_mask", action='store_true',
help="If true, compute the distilation loss only the [MLM] prediction distribution.")
parser.add_argument("--freeze_pos_embs", action="store_true",
help="Freeze positional embeddings during distillation. For student_type in ['roberta', 'gpt2'] only.")
parser.add_argument("--freeze_token_type_embds", action="store_true",
help="Freeze token type embeddings during distillation if existent. For student_type in ['roberta'] only.")
parser.add_argument("--n_epoch", type=int, default=3,
help="Number of pass on the whole dataset.")
parser.add_argument("--batch_size", type=int, default=5,
help="Batch size (for each process).")
parser.add_argument("--tokens_per_batch", type=int, default=-1,
help="If specified, modify the batches so that they have approximately this number of tokens.")
parser.add_argument("--shuffle", action='store_false',
help="If true, shuffle the sequence order. Default is true.")
parser.add_argument("--group_by_size", action='store_false',
help="If true, group sequences that have similar length into the same batch. Default is true.")
......@@ -141,6 +173,7 @@ def main():
parser.add_argument("--checkpoint_interval", type=int, default=4000,
help="Checkpoint interval.")
args = parser.parse_args()
sanity_checks(args)
## ARGS ##
......@@ -164,21 +197,19 @@ def main():
with open(os.path.join(args.dump_path, 'parameters.json'), 'w') as f:
json.dump(vars(args), f, indent=4)
git_log(args.dump_path)
assert (args.from_pretrained_weights is None and args.from_pretrained_config is None) or \
(args.from_pretrained_weights is not None and args.from_pretrained_config is not None)
student_config_class, student_model_class, _ = MODEL_CLASSES[args.student_type]
teacher_config_class, teacher_model_class, teacher_tokenizer_class = MODEL_CLASSES[args.teacher_type]
### TOKENIZER ###
if args.teacher_type == 'bert':
tokenizer = BertTokenizer.from_pretrained(args.teacher_name)
elif args.teacher_type == 'roberta':
tokenizer = RobertaTokenizer.from_pretrained(args.teacher_name)
tokenizer = teacher_tokenizer_class.from_pretrained(args.teacher_name)
special_tok_ids = {}
for tok_name, tok_symbol in tokenizer.special_tokens_map.items():
idx = tokenizer.all_special_tokens.index(tok_symbol)
special_tok_ids[tok_name] = tokenizer.all_special_ids[idx]
logger.info(f'Special tokens {special_tok_ids}')
args.special_tok_ids = special_tok_ids
args.max_model_input_size = tokenizer.max_model_input_sizes[args.teacher_name]
## DATA LOADER ##
......@@ -187,35 +218,34 @@ def main():
data = pickle.load(fp)
assert os.path.isfile(args.token_counts)
if args.mlm:
logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
with open(args.token_counts, 'rb') as fp:
counts = pickle.load(fp)
assert len(counts) == args.vocab_size
token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
for idx in special_tok_ids.values():
token_probs[idx] = 0. # do not predict special tokens
token_probs = torch.from_numpy(token_probs)
else:
token_probs = None
train_dataloader = Dataset(params=args, data=data)
train_lm_seq_dataset = LmSeqsDataset(params=args, data=data)
logger.info(f'Data loader created.')
## STUDENT ##
if args.from_pretrained_weights is not None:
assert os.path.isfile(args.from_pretrained_weights)
assert os.path.isfile(args.from_pretrained_config)
logger.info(f'Loading pretrained weights from {args.from_pretrained_weights}')
logger.info(f'Loading pretrained config from {args.from_pretrained_config}')
stu_architecture_config = DistilBertConfig.from_json_file(args.from_pretrained_config)
logger.info(f'Loading student config from {args.student_config}')
stu_architecture_config = student_config_class.from_pretrained(args.student_config)
stu_architecture_config.output_hidden_states = True
student = DistilBertForMaskedLM.from_pretrained(args.from_pretrained_weights,
if args.student_pretrained_weights is not None:
logger.info(f'Loading pretrained weights from {args.student_pretrained_weights}')
student = student_model_class.from_pretrained(args.student_pretrained_weights,
config=stu_architecture_config)
else:
args.vocab_size_or_config_json_file = args.vocab_size
stu_architecture_config = DistilBertConfig(**vars(args), output_hidden_states=True)
student = DistilBertForMaskedLM(stu_architecture_config)
student = student_model_class(stu_architecture_config)
if args.n_gpu > 0:
......@@ -224,18 +254,31 @@ def main():
## TEACHER ##
if args.teacher_type == 'bert':
teacher = BertForMaskedLM.from_pretrained(args.teacher_name, output_hidden_states=True)
elif args.teacher_type == 'roberta':
teacher = RobertaForMaskedLM.from_pretrained(args.teacher_name, output_hidden_states=True)
teacher = teacher_model_class.from_pretrained(args.teacher_name, output_hidden_states=True)
if args.n_gpu > 0:
teacher.to(f'cuda:{args.local_rank}')
logger.info(f'Teacher loaded from {args.teacher_name}.')
## FREEZING ##
if args.freeze_pos_embs:
freeze_pos_embeddings(student, args)
if args.freeze_token_type_embds:
freeze_token_type_embeddings(student, args)
## SANITY CHECKS ##
assert student.config.vocab_size == teacher.config.vocab_size
assert student.config.hidden_size == teacher.config.hidden_size
assert student.config.max_position_embeddings == teacher.config.max_position_embeddings
if args.mlm:
assert token_probs.size(0) == stu_architecture_config.vocab_size
## DISTILLER ##
torch.cuda.empty_cache()
distiller = Distiller(params=args,
dataloader=train_dataloader,
dataset=train_lm_seq_dataset,
token_probs=token_probs,
student=student,
teacher=teacher)
......
examples/distillation/training_configs/distilbert-base-uncased.json 0 → 100644
View file @ 3c1b6f59
{
"activation": "gelu",
"attention_dropout": 0.1,
"dim": 768,
"dropout": 0.1,
"hidden_dim": 3072,
"initializer_range": 0.02,
"max_position_embeddings": 512,
"n_heads": 12,
"n_layers": 6,
"sinusoidal_pos_embds": true,
"tie_weights_": true,
"vocab_size": 30522
}
\ No newline at end of file
examples/distillation/training_configs/distilgpt2.json 0 → 100644
View file @ 3c1b6f59
{
"initializer_range": 0.02,
"layer_norm_epsilon": 0.00001,
"n_ctx": 1024,
"n_embd": 768,
"n_head": 12,
"n_layer": 6,
"n_positions": 1024,
"vocab_size": 50257
}
\ No newline at end of file
examples/requirements.txt
View file @ 3c1b6f59
tensorboardX
tensorboard
scikit-learn
seqeval
examples/run_bertology.py
View file @ 3c1b6f59
......@@ -32,7 +32,7 @@ from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subse
from torch.utils.data.distributed import DistributedSampler
from torch.nn import CrossEntropyLoss, MSELoss
from pytorch_transformers import (WEIGHTS_NAME,
from transformers import (WEIGHTS_NAME,
BertConfig, BertForSequenceClassification, BertTokenizer,
XLMConfig, XLMForSequenceClassification, XLMTokenizer,
XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer)
......
examples/run_generation.py
View file @ 3c1b6f59
......@@ -14,7 +14,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/Transformer-XL/XLNet)
""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/CTRL/Transformer-XL/XLNet)
"""
from __future__ import absolute_import, division, print_function, unicode_literals
......@@ -26,12 +26,14 @@ import torch
import torch.nn.functional as F
import numpy as np
from pytorch_transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig
from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
from pytorch_transformers import GPT2LMHeadModel, GPT2Tokenizer
from pytorch_transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
from pytorch_transformers import XLNetLMHeadModel, XLNetTokenizer
from pytorch_transformers import TransfoXLLMHeadModel, TransfoXLTokenizer
from transformers import GPT2LMHeadModel, GPT2Tokenizer
from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
from transformers import XLNetLMHeadModel, XLNetTokenizer
from transformers import TransfoXLLMHeadModel, TransfoXLTokenizer
from transformers import CTRLLMHeadModel, CTRLTokenizer
from transformers import XLMWithLMHeadModel, XLMTokenizer
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
......@@ -41,13 +43,15 @@ logger = logging.getLogger(__name__)
MAX_LENGTH = int(10000) # Hardcoded max length to avoid infinite loop
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig)), ())
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
MODEL_CLASSES = {
'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
'xlm': (XLMWithLMHeadModel, XLMTokenizer),
}
# Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
......@@ -103,7 +107,8 @@ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')
return logits
def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, is_xlnet=False, device='cpu'):
def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
context = torch.tensor(context, dtype=torch.long, device=device)
context = context.unsqueeze(0).repeat(num_samples, 1)
generated = context
......@@ -121,9 +126,26 @@ def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=
target_mapping[0, 0, -1] = 1.0 # predict last token
inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
outputs = model(**inputs) # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
next_token_logits = outputs[0][:, -1, :] / temperature
if is_xlm_mlm and xlm_mask_token:
# XLM MLM models are direct models (predict same token, not next token)
# => need one additional dummy token in the input (will be masked and guessed)
input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
inputs = {'input_ids': input_ids}
if xlm_lang is not None:
inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
outputs = model(**inputs) # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
# reptition penalty from CTRL (https://arxiv.org/abs/1909.05858)
for _ in set(generated.view(-1).tolist()):
next_token_logits[_] /= repetition_penalty
filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
if temperature == 0: #greedy sampling:
next_token = torch.argmax(filtered_logits).unsqueeze(0)
else:
next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
generated = torch.cat((generated, next_token), dim=1)
return generated
......@@ -137,15 +159,21 @@ def main():
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
parser.add_argument("--prompt", type=str, default="")
parser.add_argument("--padding_text", type=str, default="")
parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
parser.add_argument("--length", type=int, default=20)
parser.add_argument("--temperature", type=float, default=1.0)
parser.add_argument("--num_samples", type=int, default=1)
parser.add_argument("--temperature", type=float, default=1.0,
help="temperature of 0 implies greedy sampling")
parser.add_argument("--repetition_penalty", type=float, default=1.0,
help="primarily useful for CTRL model; in that case, use 1.2")
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")
parser.add_argument('--stop_token', type=str, default=None,
help="Token at which text generation is stopped")
args = parser.parse_args()
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
......@@ -167,13 +195,39 @@ def main():
elif args.length < 0:
args.length = MAX_LENGTH # avoid infinite loop
print(args)
logger.info(args)
if args.model_type in ["ctrl"]:
if args.temperature > 0.7:
logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
while True:
xlm_lang = None
# XLM Language usage detailed in the issues #1414
if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
and model.config.use_lang_emb:
if args.xlm_lang:
language = args.xlm_lang
else:
language = None
while language not in tokenizer.lang2id.keys():
language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
xlm_lang = tokenizer.lang2id[language]
# XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
if is_xlm_mlm:
xlm_mask_token = tokenizer.mask_token_id
else:
xlm_mask_token = None
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)
context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
if args.model_type == "ctrl":
if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
out = sample_sequence(
model=model,
context=context_tokens,
......@@ -182,13 +236,20 @@ def main():
temperature=args.temperature,
top_k=args.top_k,
top_p=args.top_p,
device=args.device,
repetition_penalty=args.repetition_penalty,
is_xlnet=bool(args.model_type == "xlnet"),
is_xlm_mlm=is_xlm_mlm,
xlm_mask_token=xlm_mask_token,
xlm_lang=xlm_lang,
device=args.device,
)
out = out[:, len(context_tokens):].tolist()
for o in out:
text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
text = text[: text.find(args.stop_token) if args.stop_token else None]
print(text)
if args.prompt:
break
return text
......
examples/run_glue.py
View file @ 3c1b6f59
......@@ -28,10 +28,15 @@ import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data.distributed import DistributedSampler
from tensorboardX import SummaryWriter
try:
from torch.utils.tensorboard import SummaryWriter
except:
from tensorboardX import SummaryWriter
from tqdm import tqdm, trange
from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
from transformers import (WEIGHTS_NAME, BertConfig,
BertForSequenceClassification, BertTokenizer,
RobertaConfig,
RobertaForSequenceClassification,
......@@ -39,22 +44,29 @@ from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
XLMConfig, XLMForSequenceClassification,
XLMTokenizer, XLNetConfig,
XLNetForSequenceClassification,
XLNetTokenizer)
XLNetTokenizer,
DistilBertConfig,
DistilBertForSequenceClassification,
DistilBertTokenizer)
from pytorch_transformers import AdamW, WarmupLinearSchedule
from transformers import AdamW, WarmupLinearSchedule
from utils_glue import (compute_metrics, convert_examples_to_features,
output_modes, processors)
from transformers import glue_compute_metrics as compute_metrics
from transformers import glue_output_modes as output_modes
from transformers import glue_processors as processors
from transformers import glue_convert_examples_to_features as convert_examples_to_features
logger = logging.getLogger(__name__)
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig)), ())
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig,
RobertaConfig, DistilBertConfig)), ())
MODEL_CLASSES = {
'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
}
......@@ -128,10 +140,11 @@ def train(args, train_dataset, model, tokenizer):
batch = tuple(t.to(args.device) for t in batch)
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None, # XLM and RoBERTa don't use segment_ids
'labels': batch[3]}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
......@@ -141,15 +154,18 @@ def train(args, train_dataset, model, tokenizer):
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
scheduler.step() # Update learning rate schedule
if (step + 1) % args.gradient_accumulation_steps == 0 and not args.tpu:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
......@@ -173,6 +189,11 @@ def train(args, train_dataset, model, tokenizer):
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
if args.tpu:
args.xla_model.optimizer_step(optimizer, barrier=True)
model.zero_grad()
global_step += 1
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
......@@ -218,8 +239,9 @@ def evaluate(args, model, tokenizer, prefix=""):
with torch.no_grad():
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None, # XLM and RoBERTa don't use segment_ids
'labels': batch[3]}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert', 'xlnet'] else None # XLM, DistilBERT and RoBERTa don't use segment_ids
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
......@@ -240,7 +262,7 @@ def evaluate(args, model, tokenizer, prefix=""):
result = compute_metrics(eval_task, preds, out_label_ids)
results.update(result)
output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
......@@ -262,7 +284,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length),
str(task)))
if os.path.exists(cached_features_file):
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
else:
......@@ -272,12 +294,11 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
# HACK(label indices are swapped in RoBERTa pretrained model)
label_list[1], label_list[2] = label_list[2], label_list[1]
examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer, output_mode,
cls_token_at_end=bool(args.model_type in ['xlnet']), # xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
sep_token=tokenizer.sep_token,
sep_token_extra=bool(args.model_type in ['roberta']), # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
features = convert_examples_to_features(examples,
tokenizer,
label_list=label_list,
max_length=args.max_seq_length,
output_mode=output_mode,
pad_on_left=bool(args.model_type in ['xlnet']), # pad on the left for xlnet
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
......@@ -291,14 +312,14 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long)
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.float)
all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
return dataset
......@@ -372,6 +393,15 @@ def main():
parser.add_argument('--seed', type=int, default=42,
help="random seed for initialization")
parser.add_argument('--tpu', action='store_true',
help="Whether to run on the TPU defined in the environment variables")
parser.add_argument('--tpu_ip_address', type=str, default='',
help="TPU IP address if none are set in the environment variables")
parser.add_argument('--tpu_name', type=str, default='',
help="TPU name if none are set in the environment variables")
parser.add_argument('--xrt_tpu_config', type=str, default='',
help="XRT TPU config if none are set in the environment variables")
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',
......@@ -405,6 +435,23 @@ def main():
args.n_gpu = 1
args.device = device
if args.tpu:
if args.tpu_ip_address:
os.environ["TPU_IP_ADDRESS"] = args.tpu_ip_address
if args.tpu_name:
os.environ["TPU_NAME"] = args.tpu_name
if args.xrt_tpu_config:
os.environ["XRT_TPU_CONFIG"] = args.xrt_tpu_config
assert "TPU_IP_ADDRESS" in os.environ
assert "TPU_NAME" in os.environ
assert "XRT_TPU_CONFIG" in os.environ
import torch_xla
import torch_xla.core.xla_model as xm
args.device = xm.xla_device()
args.xla_model = xm
# Setup logging
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
......@@ -450,7 +497,7 @@ def main():
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0) and not args.tpu:
# Create output directory if needed
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
......@@ -478,13 +525,15 @@ def main():
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("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
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 ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=global_step)
result = evaluate(args, model, tokenizer, prefix=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
results.update(result)
......
examples/run_lm_finetuning.py
View file @ 3c1b6f59
......@@ -27,19 +27,27 @@ import logging
import os
import pickle
import random
import re
import shutil
import numpy as np
import torch
from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
from torch.utils.data.distributed import DistributedSampler
from tensorboardX import SummaryWriter
try:
from torch.utils.tensorboard import SummaryWriter
except:
from tensorboardX import SummaryWriter
from tqdm import tqdm, trange
from pytorch_transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
from transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
BertConfig, BertForMaskedLM, BertTokenizer,
GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
RobertaConfig, RobertaForMaskedLM, RobertaTokenizer)
RobertaConfig, RobertaForMaskedLM, RobertaTokenizer,
DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer)
logger = logging.getLogger(__name__)
......@@ -49,7 +57,8 @@ MODEL_CLASSES = {
'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer)
'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer)
}
......@@ -57,7 +66,7 @@ class TextDataset(Dataset):
def __init__(self, tokenizer, file_path='train', block_size=512):
assert os.path.isfile(file_path)
directory, filename = os.path.split(file_path)
cached_features_file = os.path.join(directory, f'cached_lm_{block_size}_{filename}')
cached_features_file = os.path.join(directory, 'cached_lm_' + str(block_size) + '_' + filename)
if os.path.exists(cached_features_file):
logger.info("Loading features from cached file %s", cached_features_file)
......@@ -72,9 +81,8 @@ class TextDataset(Dataset):
tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))
while len(tokenized_text) >= block_size: # Truncate in block of block_size
self.examples.append(tokenizer.add_special_tokens_single_sentence(tokenized_text[:block_size]))
tokenized_text = tokenized_text[block_size:]
for i in range(0, len(tokenized_text)-block_size+1, block_size): # Truncate in block of block_size
self.examples.append(tokenizer.build_inputs_with_special_tokens(tokenized_text[i:i+block_size]))
# Note that we are loosing the last truncated example here for the sake of simplicity (no padding)
# If your dataset is small, first you should loook for a bigger one :-) and second you
# can change this behavior by adding (model specific) padding.
......@@ -103,11 +111,43 @@ def set_seed(args):
torch.cuda.manual_seed_all(args.seed)
def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False):
if not args.save_total_limit:
return
if args.save_total_limit <= 0:
return
# Check if we should delete older checkpoint(s)
glob_checkpoints = glob.glob(os.path.join(args.output_dir, '{}-*'.format(checkpoint_prefix)))
if len(glob_checkpoints) <= args.save_total_limit:
return
ordering_and_checkpoint_path = []
for path in glob_checkpoints:
if use_mtime:
ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
else:
regex_match = re.match('.*{}-([0-9]+)'.format(checkpoint_prefix), path)
if regex_match and regex_match.groups():
ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))
checkpoints_sorted = sorted(ordering_and_checkpoint_path)
checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
for checkpoint in checkpoints_to_be_deleted:
logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint))
shutil.rmtree(checkpoint)
def mask_tokens(inputs, tokenizer, args):
""" Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
labels = inputs.clone()
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
masked_indices = torch.bernoulli(torch.full(labels.shape, args.mlm_probability)).bool()
probability_matrix = torch.full(labels.shape, args.mlm_probability)
special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
masked_indices = torch.bernoulli(probability_matrix).bool()
labels[~masked_indices] = -1 # We only compute loss on masked tokens
# 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
......@@ -186,7 +226,7 @@ def train(args, train_dataset, model, tokenizer):
labels = labels.to(args.device)
model.train()
outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
......@@ -221,8 +261,9 @@ def train(args, train_dataset, model, tokenizer):
logging_loss = tr_loss
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
checkpoint_prefix = 'checkpoint'
# Save model checkpoint
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
output_dir = os.path.join(args.output_dir, '{}-{}'.format(checkpoint_prefix, global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
......@@ -230,6 +271,8 @@ def train(args, train_dataset, model, tokenizer):
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
_rotate_checkpoints(args, checkpoint_prefix)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
......@@ -266,10 +309,12 @@ def evaluate(args, model, tokenizer, prefix=""):
model.eval()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = batch.to(args.device)
inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
inputs = inputs.to(args.device)
labels = labels.to(args.device)
with torch.no_grad():
outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
lm_loss = outputs[0]
eval_loss += lm_loss.mean().item()
nb_eval_steps += 1
......@@ -281,7 +326,7 @@ def evaluate(args, model, tokenizer, prefix=""):
"perplexity": perplexity
}
output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
......@@ -358,6 +403,8 @@ def main():
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',
......@@ -380,7 +427,7 @@ def main():
parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
args = parser.parse_args()
if args.model_type in ["bert", "roberta"] and not args.mlm:
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:
......@@ -479,13 +526,15 @@ def main():
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("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
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 ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=global_step)
result = evaluate(args, model, tokenizer, prefix=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
results.update(result)
......
examples/run_multiple_choice.py
View file @ 3c1b6f59
......@@ -29,16 +29,21 @@ import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data.distributed import DistributedSampler
from tensorboardX import SummaryWriter
try:
from torch.utils.tensorboard import SummaryWriter
except:
from tensorboardX import SummaryWriter
from tqdm import tqdm, trange
from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
from transformers import (WEIGHTS_NAME, BertConfig,
BertForMultipleChoice, BertTokenizer,
XLNetConfig, XLNetForMultipleChoice,
XLNetTokenizer, RobertaConfig,
RobertaForMultipleChoice, RobertaTokenizer)
from pytorch_transformers import AdamW, WarmupLinearSchedule
from transformers import AdamW, WarmupLinearSchedule
from utils_multiple_choice import (convert_examples_to_features, processors)
......@@ -141,7 +146,7 @@ def train(args, train_dataset, model, tokenizer):
'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None, # XLM don't use segment_ids
'labels': batch[3]}
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
......@@ -293,7 +298,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length),
str(task)))
if os.path.exists(cached_features_file):
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
else:
......@@ -306,14 +311,14 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
else:
examples = processor.get_train_examples(args.data_dir)
logger.info("Training number: %s", str(len(examples)))
features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer,
cls_token_at_end=bool(args.model_type in ['xlnet']), # xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
sep_token=tokenizer.sep_token,
sep_token_extra=bool(args.model_type in ['roberta']),
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
features = convert_examples_to_features(
examples,
label_list,
args.max_seq_length,
tokenizer,
pad_on_left=bool(args.model_type in ['xlnet']), # pad on the left for xlnet
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0)
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, cached_features_file)
......@@ -362,7 +367,7 @@ def main():
help="Whether to run eval on the dev set.")
parser.add_argument("--do_test", action='store_true', help='Whether to run test on the test set')
parser.add_argument("--evaluate_during_training", action='store_true',
help="Rul evaluation during training at each logging step.")
help="Run 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.")
......@@ -508,13 +513,15 @@ def main():
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("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
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 ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=global_step)
result = evaluate(args, model, tokenizer, prefix=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
results.update(result)
......@@ -524,13 +531,15 @@ def main():
checkpoints = [args.output_dir]
# if args.eval_all_checkpoints: # can not use this to do test!!
# checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
# logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
# 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 ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
model = model_class.from_pretrained(checkpoint)
model.to(args.device)
result = evaluate(args, model, tokenizer, prefix=global_step, test=True)
result = evaluate(args, model, tokenizer, prefix=prefix, test=True)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
results.update(result)
if best_steps:
......
examples/run_ner.py 0 → 100644
View file @ 3c1b6f59
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """
from __future__ import absolute_import, division, print_function
import argparse
import glob
import logging
import os
import random
import numpy as np
import torch
from seqeval.metrics import precision_score, recall_score, f1_score
from tensorboardX import SummaryWriter
from torch.nn import CrossEntropyLoss
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange
from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
from transformers import AdamW, WarmupLinearSchedule
from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, BertTokenizer
from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
logger = logging.getLogger(__name__)
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig)),
())
MODEL_CLASSES = {
"bert": (BertConfig, BertForTokenClassification, BertTokenizer),
"roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer)
}
def set_seed(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
""" Train the model """
if args.local_rank in [-1, 0]:
tb_writer = SummaryWriter()
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
# Prepare optimizer and schedule (linear warmup and decay)
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": 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 = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps * (
torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
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],
"token_type_ids": batch[2] if args.model_type in ["bert", "xlnet"] else None,
# XLM and RoBERTa don"t use segment_ids
"labels": batch[3]}
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel training
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
scheduler.step() # Update learning rate schedule
optimizer.step()
model.zero_grad()
global_step += 1
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, labels, pad_token_label_id)
for key, value in results.items():
tb_writer.add_scalar("eval_{}".format(key), value, global_step)
tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
logging_loss = tr_loss
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)
model_to_save = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
if args.local_rank in [-1, 0]:
tb_writer.close()
return global_step, tr_loss / global_step
def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""):
eval_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode=mode)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# Eval!
logger.info("***** Running evaluation %s *****", prefix)
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
preds = None
out_label_ids = None
model.eval()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": batch[2] if args.model_type in ["bert", "xlnet"] else None,
# XLM and RoBERTa don"t use segment_ids
"labels": batch[3]}
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
if args.n_gpu > 1:
tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu parallel evaluating
eval_loss += tmp_eval_loss.item()
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs["labels"].detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = np.argmax(preds, axis=2)
label_map = {i: label for i, label in enumerate(labels)}
out_label_list = [[] for _ in range(out_label_ids.shape[0])]
preds_list = [[] for _ in range(out_label_ids.shape[0])]
for i in range(out_label_ids.shape[0]):
for j in range(out_label_ids.shape[1]):
if out_label_ids[i, j] != pad_token_label_id:
out_label_list[i].append(label_map[out_label_ids[i][j]])
preds_list[i].append(label_map[preds[i][j]])
results = {
"loss": eval_loss,
"precision": precision_score(out_label_list, preds_list),
"recall": recall_score(out_label_list, preds_list),
"f1": f1_score(out_label_list, preds_list)
}
logger.info("***** Eval results %s *****", prefix)
for key in sorted(results.keys()):
logger.info(" %s = %s", key, str(results[key]))
return results, preds_list
def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
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
cached_features_file = os.path.join(args.data_dir, "cached_{}_{}_{}".format(mode,
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length)))
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
features = torch.load(cached_features_file)
else:
logger.info("Creating features from dataset file at %s", args.data_dir)
examples = read_examples_from_file(args.data_dir, mode)
features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
cls_token_at_end=bool(args.model_type in ["xlnet"]),
# xlnet has a cls token at the end
cls_token=tokenizer.cls_token,
cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
sep_token=tokenizer.sep_token,
sep_token_extra=bool(args.model_type in ["roberta"]),
# roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
pad_on_left=bool(args.model_type in ["xlnet"]),
# pad on the left for xlnet
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
pad_token_label_id=pad_token_label_id
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
torch.save(features, 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
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
return dataset
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--data_dir", default=None, type=str, required=True,
help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
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 selected in the list: " + ", ".join(ALL_MODELS))
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("--labels", default="", type=str,
help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
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,
help="Pretrained tokenizer name or path if not the same as model_name")
parser.add_argument("--cache_dir", default="", type=str,
help="Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument("--max_seq_length", default=128, type=int,
help="The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.")
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("--do_predict", action="store_true",
help="Whether to run predictions on the test set.")
parser.add_argument("--evaluate_during_training", action="store_true",
help="Whether to run 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("--per_gpu_train_batch_size", default=8, type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size", default=8, 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 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,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs", default=3.0, 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_steps", default=0, type=int,
help="Linear warmup over warmup_steps.")
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("--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.")
args = parser.parse_args()
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(
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
args.output_dir))
# 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)
# Prepare CONLL-2003 task
labels = get_labels(args.labels)
num_labels = len(labels)
# Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later
pad_token_label_id = CrossEntropyLoss().ignore_index
# 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
args.model_type = args.model_type.lower()
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,
num_labels=num_labels)
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)
model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path),
config=config)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
model.to(args.device)
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode="train")
global_step, tr_loss = train(args, train_dataset, model, tokenizer, labels, pad_token_label_id)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
# Create output directory if needed
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)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Good practice: save your training arguments together with the trained model
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
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("pytorch_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, labels, pad_token_label_id, mode="dev", prefix=global_step)
if global_step:
result = {"{}_{}".format(global_step, k): v for k, v in result.items()}
results.update(result)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
for key in sorted(results.keys()):
writer.write("{} = {}\n".format(key, str(results[key])))
if args.do_predict and args.local_rank in [-1, 0]:
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
model = model_class.from_pretrained(args.output_dir)
model.to(args.device)
result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")
# Save results
output_test_results_file = os.path.join(args.output_dir, "test_results.txt")
with open(output_test_results_file, "w") as writer:
for key in sorted(result.keys()):
writer.write("{} = {}\n".format(key, str(result[key])))
# Save predictions
output_test_predictions_file = os.path.join(args.output_dir, "test_predictions.txt")
with open(output_test_predictions_file, "w") as writer:
with open(os.path.join(args.data_dir, "test.txt"), "r") as f:
example_id = 0
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
writer.write(line)
if not predictions[example_id]:
example_id += 1
elif predictions[example_id]:
output_line = line.split()[0] + " " + predictions[example_id].pop(0) + "\n"
writer.write(output_line)
else:
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
return results
if __name__ == "__main__":
main()
examples/run_squad.py
View file @ 3c1b6f59
......@@ -13,7 +13,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Finetuning the library models for question-answering on SQuAD (Bert, XLM, XLNet)."""
""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
from __future__ import absolute_import, division, print_function
......@@ -28,18 +28,23 @@ import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data.distributed import DistributedSampler
from tqdm import tqdm, trange
from tensorboardX import SummaryWriter
try:
from torch.utils.tensorboard import SummaryWriter
except:
from tensorboardX import SummaryWriter
from tqdm import tqdm, trange
from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
from transformers import (WEIGHTS_NAME, BertConfig,
BertForQuestionAnswering, BertTokenizer,
XLMConfig, XLMForQuestionAnswering,
XLMTokenizer, XLNetConfig,
XLNetForQuestionAnswering,
XLNetTokenizer)
XLNetTokenizer,
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
from pytorch_transformers import AdamW, WarmupLinearSchedule
from transformers import AdamW, WarmupLinearSchedule
from utils_squad import (read_squad_examples, convert_examples_to_features,
RawResult, write_predictions,
......@@ -59,6 +64,7 @@ MODEL_CLASSES = {
'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
}
def set_seed(args):
......@@ -133,14 +139,15 @@ def train(args, train_dataset, model, tokenizer):
batch = tuple(t.to(args.device) for t in batch)
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': None if args.model_type == 'xlm' else batch[2],
'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]})
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
if args.n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
......@@ -150,13 +157,16 @@ def train(args, train_dataset, model, tokenizer):
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
......@@ -216,9 +226,10 @@ def evaluate(args, model, tokenizer, prefix=""):
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
'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]
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4],
......@@ -473,6 +484,16 @@ def main():
logger.info("Training/evaluation parameters %s", args)
# Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
# Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
# remove the need for this code, but it is still valid.
if args.fp16:
try:
import apex
apex.amp.register_half_function(torch, 'einsum')
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
......@@ -508,7 +529,7 @@ def main():
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("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
logger.info("Evaluate the following checkpoints: %s", checkpoints)
......
examples/run_summarization_finetuning.py 0 → 100644
View file @ 3c1b6f59
# coding=utf-8
# Copyright 2019 The HuggingFace Inc. team.
# Copyright (c) 2019 The HuggingFace Inc. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Finetuning seq2seq models for sequence generation."""
import argparse
import functools
import logging
import os
import random
import sys
import numpy as np
from tqdm import tqdm, trange
import torch
from torch.optim import Adam
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from transformers import (
AutoTokenizer,
BertForMaskedLM,
BertConfig,
PreTrainedEncoderDecoder,
Model2Model,
)
from utils_summarization import (
CNNDailyMailDataset,
encode_for_summarization,
fit_to_block_size,
build_lm_labels,
build_mask,
compute_token_type_ids,
)
logger = logging.getLogger(__name__)
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
def set_seed(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# ------------
# Load dataset
# ------------
def load_and_cache_examples(args, tokenizer):
dataset = CNNDailyMailDataset(tokenizer, data_dir=args.data_dir)
return dataset
def collate(data, tokenizer, block_size):
""" List of tuple as an input. """
# remove the files with empty an story/summary, encode and fit to block
data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data)
data = [
encode_for_summarization(story, summary, tokenizer) for story, summary in data
]
data = [
(
fit_to_block_size(story, block_size, tokenizer.pad_token_id),
fit_to_block_size(summary, block_size, tokenizer.pad_token_id),
)
for story, summary in data
]
stories = torch.tensor([story for story, summary in data])
summaries = torch.tensor([summary for story, summary in data])
encoder_token_type_ids = compute_token_type_ids(stories, tokenizer.cls_token_id)
encoder_mask = build_mask(stories, tokenizer.pad_token_id)
decoder_mask = build_mask(summaries, tokenizer.pad_token_id)
lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id)
return (
stories,
summaries,
encoder_token_type_ids,
encoder_mask,
decoder_mask,
lm_labels,
)
# ----------
# Optimizers
# ----------
class BertSumOptimizer(object):
""" Specific optimizer for BertSum.
As described in [1], the authors fine-tune BertSum for abstractive
summarization using two Adam Optimizers with different warm-up steps and
learning rate. They also use a custom learning rate scheduler.
[1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
arXiv preprint arXiv:1908.08345 (2019).
"""
def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8):
self.encoder = model.encoder
self.decoder = model.decoder
self.lr = lr
self.warmup_steps = warmup_steps
self.optimizers = {
"encoder": Adam(
model.encoder.parameters(),
lr=lr["encoder"],
betas=(beta_1, beta_2),
eps=eps,
),
"decoder": Adam(
model.decoder.parameters(),
lr=lr["decoder"],
betas=(beta_1, beta_2),
eps=eps,
),
}
self._step = 0
def _update_rate(self, stack):
return self.lr[stack] * min(
self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-0.5)
)
def zero_grad(self):
self.optimizer_decoder.zero_grad()
self.optimizer_encoder.zero_grad()
def step(self):
self._step += 1
for stack, optimizer in self.optimizers.items():
new_rate = self._update_rate(stack)
for param_group in optimizer.param_groups:
param_group["lr"] = new_rate
optimizer.step()
# ------------
# Train
# ------------
def train(args, model, tokenizer):
""" Fine-tune the pretrained model on the corpus. """
set_seed(args)
# Load the data
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_dataset = load_and_cache_examples(args, tokenizer)
train_sampler = RandomSampler(train_dataset)
model_collate_fn = functools.partial(collate, tokenizer=tokenizer, block_size=512)
train_dataloader = DataLoader(
train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
collate_fn=model_collate_fn,
)
# Training schedule
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = t_total // (
len(train_dataloader) // args.gradient_accumulation_steps + 1
)
else:
t_total = (
len(train_dataloader)
// args.gradient_accumulation_steps
* args.num_train_epochs
)
# Prepare the optimizer
lr = {"encoder": 0.002, "decoder": 0.2}
warmup_steps = {"encoder": 20000, "decoder": 10000}
optimizer = BertSumOptimizer(model, lr, warmup_steps)
# Train
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(
" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size
)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps
# * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
model.zero_grad()
train_iterator = trange(args.num_train_epochs, desc="Epoch", disable=True)
global_step = 0
tr_loss = 0.0
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=True)
for step, batch in enumerate(epoch_iterator):
source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
source = source.to(args.device)
target = target.to(args.device)
encoder_token_type_ids = encoder_token_type_ids.to(args.device)
encoder_mask = encoder_mask.to(args.device)
decoder_mask = decoder_mask.to(args.device)
lm_labels = lm_labels.to(args.device)
model.train()
outputs = model(
source,
target,
encoder_token_type_ids=encoder_token_type_ids,
encoder_attention_mask=encoder_mask,
decoder_attention_mask=decoder_mask,
decoder_lm_labels=lm_labels,
)
loss = outputs[0]
print(loss)
if args.gradient_accumulation_steps > 1:
loss /= args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
model.zero_grad()
global_step += 1
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
return global_step, tr_loss / global_step
# ------------
# Train
# ------------
def evaluate(args, model, tokenizer, prefix=""):
set_seed(args)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(
eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size
)
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(eval_dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
eval_loss = 0.0
nb_eval_steps = 0
model.eval()
for batch in tqdm(eval_dataloader, desc="Evaluating"):
source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
source = source.to(args.device)
target = target.to(args.device)
encoder_token_type_ids = encoder_token_type_ids.to(args.device)
encoder_mask = encoder_mask.to(args.device)
decoder_mask = decoder_mask.to(args.device)
lm_labels = lm_labels.to(args.device)
with torch.no_grad():
outputs = model(
source,
target,
encoder_token_type_ids=encoder_token_type_ids,
encoder_attention_mask=encoder_mask,
decoder_attention_mask=decoder_mask,
decoder_lm_labels=lm_labels,
)
lm_loss = outputs[0]
eval_loss += lm_loss.mean().item()
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
perplexity = torch.exp(torch.tensor(eval_loss))
result = {"perplexity": perplexity}
# Save the evaluation's results
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(prefix))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
return result
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--data_dir",
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.",
)
# Optional parameters
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(
"--do_evaluate",
type=bool,
default=False,
help="Run model evaluation on out-of-sample data.",
)
parser.add_argument("--do_train", type=bool, default=False, help="Run training.")
parser.add_argument(
"--do_overwrite_output_dir",
type=bool,
default=False,
help="Whether to overwrite the output dir.",
)
parser.add_argument(
"--model_name_or_path",
default="bert-base-cased",
type=str,
help="The model checkpoint to initialize the encoder and decoder's weights with.",
)
parser.add_argument(
"--model_type",
default="bert",
type=str,
help="The decoder architecture to be fine-tuned.",
)
parser.add_argument(
"--max_grad_norm", default=1.0, type=float, help="Max gradient norm."
)
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(
"--to_cpu", default=False, type=bool, help="Whether to force training on CPU."
)
parser.add_argument(
"--num_train_epochs",
default=10,
type=int,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--per_gpu_train_batch_size",
default=4,
type=int,
help="Batch size per GPU/CPU for training.",
)
parser.add_argument("--seed", default=42, type=int)
args = parser.parse_args()
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.do_overwrite_output_dir
):
raise ValueError(
"Output directory ({}) already exists and is not empty. Use --do_overwrite_output_dir to overwrite.".format(
args.output_dir
)
)
# Set up training device
if args.to_cpu or not torch.cuda.is_available():
args.device = torch.device("cpu")
args.n_gpu = 0
else:
args.device = torch.device("cuda")
args.n_gpu = torch.cuda.device_count()
# Load pretrained model and tokenizer. The decoder's weights are randomly initialized.
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
config = BertConfig.from_pretrained(args.model_name_or_path)
decoder_model = BertForMaskedLM(config)
model = Model2Model.from_pretrained(
args.model_name_or_path, decoder_model=decoder_model
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
0,
args.device,
args.n_gpu,
False,
False,
)
logger.info("Training/evaluation parameters %s", args)
# Train the model
model.to(args.device)
if args.do_train:
global_step, tr_loss = train(args, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
logger.info("Saving model checkpoint to %s", args.output_dir)
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
model_to_save = (
model.module if hasattr(model, "module") else 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_arguments.bin"))
# Evaluate the model
results = {}
if args.do_evaluate:
checkpoints = []
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
encoder_checkpoint = os.path.join(checkpoint, "encoder")
decoder_checkpoint = os.path.join(checkpoint, "decoder")
model = PreTrainedEncoderDecoder.from_pretrained(
encoder_checkpoint, decoder_checkpoint
)
model.to(args.device)
results = "placeholder"
return results
if __name__ == "__main__":
main()
examples/run_tf_glue.py 0 → 100644
View file @ 3c1b6f59
import os
import tensorflow as tf
import tensorflow_datasets
from transformers import BertTokenizer, TFBertForSequenceClassification, BertConfig, glue_convert_examples_to_features, BertForSequenceClassification, glue_processors
# script parameters
BATCH_SIZE = 32
EVAL_BATCH_SIZE = BATCH_SIZE * 2
USE_XLA = False
USE_AMP = False
EPOCHS = 3
TASK = "mrpc"
if TASK == "sst-2":
TFDS_TASK = "sst2"
elif TASK == "sts-b":
TFDS_TASK = "stsb"
else:
TFDS_TASK = TASK
num_labels = len(glue_processors[TASK]().get_labels())
print(num_labels)
tf.config.optimizer.set_jit(USE_XLA)
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": USE_AMP})
# Load tokenizer and model from pretrained model/vocabulary. Specify the number of labels to classify (2+: classification, 1: regression)
config = BertConfig.from_pretrained("bert-base-cased", num_labels=num_labels)
tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
model = TFBertForSequenceClassification.from_pretrained('bert-base-cased', config=config)
# Load dataset via TensorFlow Datasets
data, info = tensorflow_datasets.load(f'glue/{TFDS_TASK}', with_info=True)
train_examples = info.splits['train'].num_examples
# MNLI expects either validation_matched or validation_mismatched
valid_examples = info.splits['validation'].num_examples
# Prepare dataset for GLUE as a tf.data.Dataset instance
train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, 128, TASK)
# MNLI expects either validation_matched or validation_mismatched
valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, 128, TASK)
train_dataset = train_dataset.shuffle(128).batch(BATCH_SIZE).repeat(-1)
valid_dataset = valid_dataset.batch(EVAL_BATCH_SIZE)
# Prepare training: Compile tf.keras model with optimizer, loss and learning rate schedule
opt = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08)
if USE_AMP:
# loss scaling is currently required when using mixed precision
opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, 'dynamic')
if num_labels == 1:
loss = tf.keras.losses.MeanSquaredError()
else:
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
model.compile(optimizer=opt, loss=loss, metrics=[metric])
# Train and evaluate using tf.keras.Model.fit()
train_steps = train_examples//BATCH_SIZE
valid_steps = valid_examples//EVAL_BATCH_SIZE
history = model.fit(train_dataset, epochs=EPOCHS, steps_per_epoch=train_steps,
validation_data=valid_dataset, validation_steps=valid_steps)
# Save TF2 model
os.makedirs('./save/', exist_ok=True)
model.save_pretrained('./save/')
# Load the TensorFlow model in PyTorch for inspection
pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
# Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
sentence_0 = 'This research was consistent with his findings.'
sentence_1 = 'His findings were compatible with this research.'
sentence_2 = 'His findings were not compatible with this research.'
inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
del inputs_1["special_tokens_mask"]
del inputs_2["special_tokens_mask"]
pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
print('sentence_1 is', 'a paraphrase' if pred_1 else 'not a paraphrase', 'of sentence_0')
print('sentence_2 is', 'a paraphrase' if pred_2 else 'not a paraphrase', 'of sentence_0')
examples/utils_multiple_choice.py
View file @ 3c1b6f59
......@@ -13,7 +13,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" BERT multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension """
""" Multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension """
from __future__ import absolute_import, division, print_function
......@@ -26,6 +26,8 @@ import json
import csv
import glob
import tqdm
from typing import List
from transformers import PreTrainedTokenizer
logger = logging.getLogger(__name__)
......@@ -40,7 +42,7 @@ class InputExample(object):
Args:
example_id: Unique id for the example.
contexts: list of str. The untokenized text of the first sequence (context of corresponding question).
question: string. The untokenized text of the second sequence (qustion).
question: string. The untokenized text of the second sequence (question).
endings: list of str. multiple choice's options. Its length must be equal to contexts' length.
label: (Optional) string. The label of the example. This should be
specified for train and dev examples, but not for test examples.
......@@ -66,7 +68,7 @@ class InputFeatures(object):
'input_mask': input_mask,
'segment_ids': segment_ids
}
for _, input_ids, input_mask, segment_ids in choices_features
for input_ids, input_mask, segment_ids in choices_features
]
self.label = label
......@@ -192,7 +194,7 @@ class SwagProcessor(DataProcessor):
return lines
def _create_examples(self, lines, type):
def _create_examples(self, lines: List[List[str]], type: str):
"""Creates examples for the training and dev sets."""
if type == "train" and lines[0][-1] != 'label':
raise ValueError(
......@@ -300,24 +302,18 @@ class ArcProcessor(DataProcessor):
return examples
def convert_examples_to_features(examples, label_list, max_seq_length,
tokenizer,
cls_token_at_end=False,
cls_token='[CLS]',
cls_token_segment_id=1,
sep_token='[SEP]',
sequence_a_segment_id=0,
sequence_b_segment_id=1,
sep_token_extra=False,
def convert_examples_to_features(
examples: List[InputExample],
label_list: List[str],
max_length: int,
tokenizer: PreTrainedTokenizer,
pad_token_segment_id=0,
pad_on_left=False,
pad_token=0,
mask_padding_with_zero=True):
""" Loads a data file into a list of `InputBatch`s
`cls_token_at_end` define the location of the CLS token:
- False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
- True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
`cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
mask_padding_with_zero=True,
) -> List[InputFeatures]:
"""
Loads a data file into a list of `InputFeatures`
"""
label_map = {label : i for i, label in enumerate(label_list)}
......@@ -328,125 +324,70 @@ def convert_examples_to_features(examples, label_list, max_seq_length,
logger.info("Writing example %d of %d" % (ex_index, len(examples)))
choices_features = []
for ending_idx, (context, ending) in enumerate(zip(example.contexts, example.endings)):
tokens_a = tokenizer.tokenize(context)
tokens_b = None
text_a = context
if example.question.find("_") != -1:
#this is for cloze question
tokens_b = tokenizer.tokenize(example.question.replace("_", ending))
else:
tokens_b = tokenizer.tokenize(example.question + " " + ending)
# you can add seq token between quesiotn and ending. This does not make too much difference.
# tokens_b = tokenizer.tokenize(example.question)
# tokens_b += [sep_token]
# if sep_token_extra:
# tokens_b += [sep_token]
# tokens_b += tokenizer.tokenize(ending)
special_tokens_count = 4 if sep_token_extra else 3
_truncate_seq_pair(tokens_a, tokens_b, max_seq_length - special_tokens_count)
# The convention in BERT is:
# (a) For sequence pairs:
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
# (b) For single sequences:
# tokens: [CLS] the dog is hairy . [SEP]
# type_ids: 0 0 0 0 0 0 0
#
# Where "type_ids" are used to indicate whether this is the first
# sequence or the second sequence. The embedding vectors for `type=0` and
# `type=1` were learned during pre-training and are added to the wordpiece
# embedding vector (and position vector). This is not *strictly* necessary
# since the [SEP] token unambiguously separates the sequences, but it makes
# it easier for the model to learn the concept of sequences.
#
# For classification tasks, the first vector (corresponding to [CLS]) is
# used as as the "sentence vector". Note that this only makes sense because
# the entire model is fine-tuned.
tokens = tokens_a + [sep_token]
if sep_token_extra:
# roberta uses an extra separator b/w pairs of sentences
tokens += [sep_token]
segment_ids = [sequence_a_segment_id] * len(tokens)
if tokens_b:
tokens += tokens_b + [sep_token]
segment_ids += [sequence_b_segment_id] * (len(tokens_b) + 1)
if cls_token_at_end:
tokens = tokens + [cls_token]
segment_ids = segment_ids + [cls_token_segment_id]
# this is for cloze question
text_b = example.question.replace("_", ending)
else:
tokens = [cls_token] + tokens
segment_ids = [cls_token_segment_id] + segment_ids
text_b = example.question + " " + ending
inputs = tokenizer.encode_plus(
text_a,
text_b,
add_special_tokens=True,
max_length=max_length,
)
if 'num_truncated_tokens' in inputs and inputs['num_truncated_tokens'] > 0:
logger.info('Attention! you are cropping tokens (swag task is ok). '
'If you are training ARC and RACE and you are poping question + options,'
'you need to try to use a bigger max seq length!')
input_ids = tokenizer.convert_tokens_to_ids(tokens)
input_ids, token_type_ids = inputs["input_ids"], inputs["token_type_ids"]
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
attention_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
# Zero-pad up to the sequence length.
padding_length = max_seq_length - len(input_ids)
padding_length = max_length - len(input_ids)
if pad_on_left:
input_ids = ([pad_token] * padding_length) + input_ids
input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
attention_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + attention_mask
token_type_ids = ([pad_token_segment_id] * padding_length) + token_type_ids
else:
input_ids = input_ids + ([pad_token] * padding_length)
input_mask = input_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
segment_ids = segment_ids + ([pad_token_segment_id] * padding_length)
attention_mask = attention_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
token_type_ids = token_type_ids + ([pad_token_segment_id] * padding_length)
assert len(input_ids) == max_length
assert len(attention_mask) == max_length
assert len(token_type_ids) == max_length
choices_features.append((input_ids, attention_mask, token_type_ids))
assert len(input_ids) == max_seq_length
assert len(input_mask) == max_seq_length
assert len(segment_ids) == max_seq_length
choices_features.append((tokens, input_ids, input_mask, segment_ids))
label = label_map[example.label]
if ex_index < 2:
logger.info("*** Example ***")
logger.info("race_id: {}".format(example.example_id))
for choice_idx, (tokens, input_ids, input_mask, segment_ids) in enumerate(choices_features):
for choice_idx, (input_ids, attention_mask, token_type_ids) in enumerate(choices_features):
logger.info("choice: {}".format(choice_idx))
logger.info("tokens: {}".format(' '.join(tokens)))
logger.info("input_ids: {}".format(' '.join(map(str, input_ids))))
logger.info("input_mask: {}".format(' '.join(map(str, input_mask))))
logger.info("segment_ids: {}".format(' '.join(map(str, segment_ids))))
logger.info("attention_mask: {}".format(' '.join(map(str, attention_mask))))
logger.info("token_type_ids: {}".format(' '.join(map(str, token_type_ids))))
logger.info("label: {}".format(label))
features.append(
InputFeatures(
example_id = example.example_id,
choices_features = choices_features,
label = label
example_id=example.example_id,
choices_features=choices_features,
label=label,
)
)
return features
def _truncate_seq_pair(tokens_a, tokens_b, max_length):
"""Truncates a sequence pair in place to the maximum length."""
# This is a simple heuristic which will always truncate the longer sequence
# one token at a time. This makes more sense than truncating an equal percent
# of tokens from each, since if one sequence is very short then each token
# that's truncated likely contains more information than a longer sequence.
# However, since we'd better not to remove tokens of options and questions, you can choose to use a bigger
# length or only pop from context
while True:
total_length = len(tokens_a) + len(tokens_b)
if total_length <= max_length:
break
if len(tokens_a) > len(tokens_b):
tokens_a.pop()
else:
logger.info('Attention! you are removing from token_b (swag task is ok). '
'If you are training ARC and RACE (you are poping question + options), '
'you need to try to use a bigger max seq length!')
tokens_b.pop()
processors = {
......@@ -456,7 +397,7 @@ processors = {
}
GLUE_TASKS_NUM_LABELS = {
MULTIPLE_CHOICE_TASKS_NUM_LABELS = {
"race", 4,
"swag", 4,
"arc", 4
......
examples/utils_ner.py 0 → 100644
View file @ 3c1b6f59
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Named entity recognition fine-tuning: utilities to work with CoNLL-2003 task. """
from __future__ import absolute_import, division, print_function
import logging
import os
from io import open
logger = logging.getLogger(__name__)
class InputExample(object):
"""A single training/test example for token classification."""
def __init__(self, guid, words, labels):
"""Constructs a InputExample.
Args:
guid: Unique id for the example.
words: list. The words of the sequence.
labels: (Optional) list. The labels for each word of the sequence. This should be
specified for train and dev examples, but not for test examples.
"""
self.guid = guid
self.words = words
self.labels = labels
class InputFeatures(object):
"""A single set of features of data."""
def __init__(self, input_ids, input_mask, segment_ids, label_ids):
self.input_ids = input_ids
self.input_mask = input_mask
self.segment_ids = segment_ids
self.label_ids = label_ids
def read_examples_from_file(data_dir, mode):
file_path = os.path.join(data_dir, "{}.txt".format(mode))
guid_index = 1
examples = []
with open(file_path, encoding="utf-8") as f:
words = []
labels = []
for line in f:
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
if words:
examples.append(InputExample(guid="{}-{}".format(mode, guid_index),
words=words,
labels=labels))
guid_index += 1
words = []
labels = []
else:
splits = line.split(" ")
words.append(splits[0])
if len(splits) > 1:
labels.append(splits[-1].replace("\n", ""))
else:
# Examples could have no label for mode = "test"
labels.append("O")
if words:
examples.append(InputExample(guid="%s-%d".format(mode, guid_index),
words=words,
labels=labels))
return examples
def convert_examples_to_features(examples,
label_list,
max_seq_length,
tokenizer,
cls_token_at_end=False,
cls_token="[CLS]",
cls_token_segment_id=1,
sep_token="[SEP]",
sep_token_extra=False,
pad_on_left=False,
pad_token=0,
pad_token_segment_id=0,
pad_token_label_id=-1,
sequence_a_segment_id=0,
mask_padding_with_zero=True):
""" Loads a data file into a list of `InputBatch`s
`cls_token_at_end` define the location of the CLS token:
- False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
- True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
`cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
"""
label_map = {label: i for i, label in enumerate(label_list)}
features = []
for (ex_index, example) in enumerate(examples):
if ex_index % 10000 == 0:
logger.info("Writing example %d of %d", ex_index, len(examples))
tokens = []
label_ids = []
for word, label in zip(example.words, example.labels):
word_tokens = tokenizer.tokenize(word)
tokens.extend(word_tokens)
# Use the real label id for the first token of the word, and padding ids for the remaining tokens
label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))
# Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
special_tokens_count = 3 if sep_token_extra else 2
if len(tokens) > max_seq_length - special_tokens_count:
tokens = tokens[:(max_seq_length - special_tokens_count)]
label_ids = label_ids[:(max_seq_length - special_tokens_count)]
# The convention in BERT is:
# (a) For sequence pairs:
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
# (b) For single sequences:
# tokens: [CLS] the dog is hairy . [SEP]
# type_ids: 0 0 0 0 0 0 0
#
# Where "type_ids" are used to indicate whether this is the first
# sequence or the second sequence. The embedding vectors for `type=0` and
# `type=1` were learned during pre-training and are added to the wordpiece
# embedding vector (and position vector). This is not *strictly* necessary
# since the [SEP] token unambiguously separates the sequences, but it makes
# it easier for the model to learn the concept of sequences.
#
# For classification tasks, the first vector (corresponding to [CLS]) is
# used as as the "sentence vector". Note that this only makes sense because
# the entire model is fine-tuned.
tokens += [sep_token]
label_ids += [pad_token_label_id]
if sep_token_extra:
# roberta uses an extra separator b/w pairs of sentences
tokens += [sep_token]
label_ids += [pad_token_label_id]
segment_ids = [sequence_a_segment_id] * len(tokens)
if cls_token_at_end:
tokens += [cls_token]
label_ids += [pad_token_label_id]
segment_ids += [cls_token_segment_id]
else:
tokens = [cls_token] + tokens
label_ids = [pad_token_label_id] + label_ids
segment_ids = [cls_token_segment_id] + segment_ids
input_ids = tokenizer.convert_tokens_to_ids(tokens)
# The mask has 1 for real tokens and 0 for padding tokens. Only real
# tokens are attended to.
input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
# Zero-pad up to the sequence length.
padding_length = max_seq_length - len(input_ids)
if pad_on_left:
input_ids = ([pad_token] * padding_length) + input_ids
input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
label_ids = ([pad_token_label_id] * padding_length) + label_ids
else:
input_ids += ([pad_token] * padding_length)
input_mask += ([0 if mask_padding_with_zero else 1] * padding_length)
segment_ids += ([pad_token_segment_id] * padding_length)
label_ids += ([pad_token_label_id] * padding_length)
assert len(input_ids) == max_seq_length
assert len(input_mask) == max_seq_length
assert len(segment_ids) == max_seq_length
assert len(label_ids) == max_seq_length
if ex_index < 5:
logger.info("*** Example ***")
logger.info("guid: %s", example.guid)
logger.info("tokens: %s", " ".join([str(x) for x in tokens]))
logger.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
logger.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
logger.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
logger.info("label_ids: %s", " ".join([str(x) for x in label_ids]))
features.append(
InputFeatures(input_ids=input_ids,
input_mask=input_mask,
segment_ids=segment_ids,
label_ids=label_ids))
return features
def get_labels(path):
if path:
with open(path, "r") as f:
labels = f.read().splitlines()
if "O" not in labels:
labels = ["O"] + labels
return labels
else:
return ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
examples/utils_squad.py
View file @ 3c1b6f59
......@@ -24,7 +24,7 @@ import math
import collections
from io import open
from pytorch_transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
from transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
# Required by XLNet evaluation method to compute optimal threshold (see write_predictions_extended() method)
from utils_squad_evaluate import find_all_best_thresh_v2, make_qid_to_has_ans, get_raw_scores
......
examples/utils_summarization.py 0 → 100644
View file @ 3c1b6f59
from collections import deque
import os
import torch
from torch.utils.data import Dataset
# ------------
# Data loading
# ------------
class CNNDailyMailDataset(Dataset):
""" Abstracts the dataset used to train seq2seq models.
CNN/Daily News:
The CNN/Daily News raw datasets are downloaded from [1]. The stories are
stored in different files; the summary appears at the end of the story as
sentences that are prefixed by the special `@highlight` line. To process
the data, untar both datasets in the same folder, and pass the path to this
folder as the "data_dir argument. The formatting code was inspired by [2].
[1] https://cs.nyu.edu/~kcho/
[2] https://github.com/abisee/cnn-dailymail/
"""
def __init__(self, tokenizer, prefix="train", data_dir=""):
assert os.path.isdir(data_dir)
self.tokenizer = tokenizer
# We initialize the class by listing all the files that contain
# stories and summaries. Files are not read in memory given
# the size of the corpus.
self.stories_path = []
datasets = ("cnn", "dailymail")
for dataset in datasets:
path_to_stories = os.path.join(data_dir, dataset, "stories")
story_filenames_list = os.listdir(path_to_stories)
for story_filename in story_filenames_list:
path_to_story = os.path.join(path_to_stories, story_filename)
if not os.path.isfile(path_to_story):
continue
self.stories_path.append(path_to_story)
def __len__(self):
return len(self.stories_path)
def __getitem__(self, idx):
story_path = self.stories_path[idx]
with open(story_path, encoding="utf-8") as source:
raw_story = source.read()
story_lines, summary_lines = process_story(raw_story)
return story_lines, summary_lines
def process_story(raw_story):
""" Extract the story and summary from a story file.
Attributes:
raw_story (str): content of the story file as an utf-8 encoded string.
Raises:
IndexError: If the stoy is empty or contains no highlights.
"""
nonempty_lines = list(
filter(lambda x: len(x) != 0, [line.strip() for line in raw_story.split("\n")])
)
# for some unknown reason some lines miss a period, add it
nonempty_lines = [_add_missing_period(line) for line in nonempty_lines]
# gather article lines
story_lines = []
lines = deque(nonempty_lines)
while True:
try:
element = lines.popleft()
if element.startswith("@highlight"):
break
story_lines.append(element)
except IndexError:
# if "@highlight" is absent from the file we pop
# all elements until there is None.
return story_lines, []
# gather summary lines
summary_lines = list(filter(lambda t: not t.startswith("@highlight"), lines))
return story_lines, summary_lines
def _add_missing_period(line):
END_TOKENS = [".", "!", "?", "...", "'", "`", '"', u"\u2019", u"\u2019", ")"]
if line.startswith("@highlight"):
return line
if line[-1] in END_TOKENS:
return line
return line + "."
# --------------------------
# Encoding and preprocessing
# --------------------------
def fit_to_block_size(sequence, block_size, pad_token):
""" Adapt the source and target sequences' lengths to the block size.
If the sequence is shorter than the block size we pad it with -1 ids
which correspond to padding tokens.
"""
if len(sequence) > block_size:
return sequence[:block_size]
else:
sequence.extend([pad_token] * (block_size - len(sequence)))
return sequence
def build_lm_labels(sequence, pad_token):
""" Padding token, encoded as 0, are represented by the value -1 so they
are not taken into account in the loss computation. """
padded = sequence.clone()
padded[padded == pad_token] = -1
return padded
def build_mask(sequence, pad_token):
""" Builds the mask. The attention mechanism will only attend to positions
with value 1. """
mask = torch.ones_like(sequence)
idx_pad_tokens = sequence == pad_token
mask[idx_pad_tokens] = 0
return mask
def encode_for_summarization(story_lines, summary_lines, tokenizer):
""" Encode the story and summary lines, and join them
as specified in [1] by using `[SEP] [CLS]` tokens to separate
sentences.
"""
story_lines_token_ids = [
tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
for line in story_lines
]
summary_lines_token_ids = [
tokenizer.add_special_tokens_single_sequence(tokenizer.encode(line))
for line in summary_lines
]
story_token_ids = [
token for sentence in story_lines_token_ids for token in sentence
]
summary_token_ids = [
token for sentence in summary_lines_token_ids for token in sentence
]
return story_token_ids, summary_token_ids
def compute_token_type_ids(batch, separator_token_id):
""" Segment embeddings as described in [1]
The values {0,1} were found in the repository [2].
Attributes:
batch: torch.Tensor, size [batch_size, block_size]
Batch of input.
separator_token_id: int
The value of the token that separates the segments.
[1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
arXiv preprint arXiv:1908.08345 (2019).
[2] https://github.com/nlpyang/PreSumm (/src/prepro/data_builder.py, commit fac1217)
"""
batch_embeddings = []
for sequence in batch:
sentence_num = 0
embeddings = []
for s in sequence:
if s == separator_token_id:
sentence_num += 1
embeddings.append(sentence_num % 2)
batch_embeddings.append(embeddings)
return torch.tensor(batch_embeddings)
examples/utils_summarization_test.py 0 → 100644
View file @ 3c1b6f59
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
import torch
from utils_summarization import (
compute_token_type_ids,
fit_to_block_size,
build_mask,
build_lm_labels,
process_story,
)
class SummarizationDataProcessingTest(unittest.TestCase):
def setUp(self):
self.block_size = 10
def test_fit_to_block_sequence_too_small(self):
""" Pad the sequence with 0 if the sequence is smaller than the block size."""
sequence = [1, 2, 3, 4]
expected_output = [1, 2, 3, 4, 0, 0, 0, 0, 0, 0]
self.assertEqual(
fit_to_block_size(sequence, self.block_size, 0), expected_output
)
def test_fit_to_block_sequence_fit_exactly(self):
""" Do nothing if the sequence is the right size. """
sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
self.assertEqual(
fit_to_block_size(sequence, self.block_size, 0), expected_output
)
def test_fit_to_block_sequence_too_big(self):
""" Truncate the sequence if it is too long. """
sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
self.assertEqual(
fit_to_block_size(sequence, self.block_size, 0), expected_output
)
def test_process_story_no_highlights(self):
""" Processing a story with no highlights returns an empty list for the summary.
"""
raw_story = """It was the year of Our Lord one thousand seven hundred and
seventy-five.\n\nSpiritual revelations were conceded to England at that
favoured period, as at this."""
_, summary_lines = process_story(raw_story)
self.assertEqual(summary_lines, [])
def test_process_empty_story(self):
""" An empty story returns an empty collection of lines.
"""
raw_story = ""
story_lines, summary_lines = process_story(raw_story)
self.assertEqual(story_lines, [])
self.assertEqual(summary_lines, [])
def test_process_story_with_missing_period(self):
raw_story = (
"It was the year of Our Lord one thousand seven hundred and "
"seventy-five\n\nSpiritual revelations were conceded to England "
"at that favoured period, as at this.\n@highlight\n\nIt was the best of times"
)
story_lines, summary_lines = process_story(raw_story)
expected_story_lines = [
"It was the year of Our Lord one thousand seven hundred and seventy-five.",
"Spiritual revelations were conceded to England at that favoured period, as at this.",
]
self.assertEqual(expected_story_lines, story_lines)
expected_summary_lines = ["It was the best of times."]
self.assertEqual(expected_summary_lines, summary_lines)
def test_build_lm_labels_no_padding(self):
sequence = torch.tensor([1, 2, 3, 4])
expected = sequence
np.testing.assert_array_equal(
build_lm_labels(sequence, 0).numpy(), expected.numpy()
)
def test_build_lm_labels(self):
sequence = torch.tensor([1, 2, 3, 4, 0, 0, 0])
expected = torch.tensor([1, 2, 3, 4, -1, -1, -1])
np.testing.assert_array_equal(
build_lm_labels(sequence, 0).numpy(), expected.numpy()
)
def test_build_mask_no_padding(self):
sequence = torch.tensor([1, 2, 3, 4])
expected = torch.tensor([1, 1, 1, 1])
np.testing.assert_array_equal(build_mask(sequence, 0).numpy(), expected.numpy())
def test_build_mask(self):
sequence = torch.tensor([1, 2, 3, 4, 23, 23, 23])
expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
np.testing.assert_array_equal(
build_mask(sequence, 23).numpy(), expected.numpy()
)
def test_build_mask_with_padding_equal_to_one(self):
sequence = torch.tensor([8, 2, 3, 4, 1, 1, 1])
expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
np.testing.assert_array_equal(build_mask(sequence, 1).numpy(), expected.numpy())
def test_compute_token_type_ids(self):
separator = 101
batch = torch.tensor(
[[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]]
)
expected = torch.tensor(
[[0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1], [0, 1, 1, 1, 0, 0]]
)
result = compute_token_type_ids(batch, separator)
np.testing.assert_array_equal(result, expected)
if __name__ == "__main__":
unittest.main()
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment