Skip to content

GitLab

Commit fa84ae26 authored by Aymeric Augustin's avatar Aymeric Augustin
Browse files

Reformat source code with black. 

This is the result of:

    $ black --line-length 119 examples templates transformers utils hubconf.py setup.py

There's a lot of fairly long lines in the project. As a consequence, I'm
picking the longest widely accepted line length, 119 characters.

This is also Thomas' preference, because it allows for explicit variable
names, to make the code easier to understand.
parent 63e3827c
Show whitespace changes
Inline Side-by-side
Showing with 2550 additions and 1952 deletions
examples/run_bertology.py
View file @ fa84ae26
......@@ -32,10 +32,18 @@ from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subse
from torch.utils.data.distributed import DistributedSampler
from torch.nn import CrossEntropyLoss, MSELoss
from transformers import (WEIGHTS_NAME,
BertConfig, BertForSequenceClassification, BertTokenizer,
XLMConfig, XLMForSequenceClassification, XLMTokenizer,
XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer)
from transformers import (
WEIGHTS_NAME,
BertConfig,
BertForSequenceClassification,
BertTokenizer,
XLMConfig,
XLMForSequenceClassification,
XLMTokenizer,
XLNetConfig,
XLNetForSequenceClassification,
XLNetTokenizer,
)
from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES
......@@ -63,7 +71,9 @@ def print_2d_tensor(tensor):
logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:d}" for x in tensor[row].cpu().data))
def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None):
def compute_heads_importance(
args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None
):
""" This method shows how to compute:
- head attention entropy
- head importance scores according to http://arxiv.org/abs/1905.10650
......@@ -85,8 +95,14 @@ def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True,
input_ids, input_mask, segment_ids, label_ids = batch
# Do a forward pass (not with torch.no_grad() since we need gradients for importance score - see below)
outputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids, head_mask=head_mask)
loss, logits, all_attentions = outputs[0], outputs[1], outputs[-1] # Loss and logits are the first, attention the last
outputs = model(
input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids, head_mask=head_mask
)
loss, logits, all_attentions = (
outputs[0],
outputs[1],
outputs[-1],
) # Loss and logits are the first, attention the last
loss.backward() # Backpropagate to populate the gradients in the head mask
if compute_entropy:
......@@ -113,15 +129,15 @@ def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True,
# Layerwise importance normalization
if not args.dont_normalize_importance_by_layer:
exponent = 2
norm_by_layer = torch.pow(torch.pow(head_importance, exponent).sum(-1), 1/exponent)
norm_by_layer = torch.pow(torch.pow(head_importance, exponent).sum(-1), 1 / exponent)
head_importance /= norm_by_layer.unsqueeze(-1) + 1e-20
if not args.dont_normalize_global_importance:
head_importance = (head_importance - head_importance.min()) / (head_importance.max() - head_importance.min())
# Print/save matrices
np.save(os.path.join(args.output_dir, 'attn_entropy.npy'), attn_entropy.detach().cpu().numpy())
np.save(os.path.join(args.output_dir, 'head_importance.npy'), head_importance.detach().cpu().numpy())
np.save(os.path.join(args.output_dir, "attn_entropy.npy"), attn_entropy.detach().cpu().numpy())
np.save(os.path.join(args.output_dir, "head_importance.npy"), head_importance.detach().cpu().numpy())
logger.info("Attention entropies")
print_2d_tensor(attn_entropy)
......@@ -129,7 +145,9 @@ def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True,
print_2d_tensor(head_importance)
logger.info("Head ranked by importance scores")
head_ranks = torch.zeros(head_importance.numel(), dtype=torch.long, device=args.device)
head_ranks[head_importance.view(-1).sort(descending=True)[1]] = torch.arange(head_importance.numel(), device=args.device)
head_ranks[head_importance.view(-1).sort(descending=True)[1]] = torch.arange(
head_importance.numel(), device=args.device
)
head_ranks = head_ranks.view_as(head_importance)
print_2d_tensor(head_ranks)
......@@ -152,7 +170,7 @@ def mask_heads(args, model, eval_dataloader):
while current_score >= original_score * args.masking_threshold:
head_mask = new_head_mask.clone() # save current head mask
# heads from least important to most - keep only not-masked heads
head_importance[head_mask == 0.0] = float('Inf')
head_importance[head_mask == 0.0] = float("Inf")
current_heads_to_mask = head_importance.view(-1).sort()[1]
if len(current_heads_to_mask) <= num_to_mask:
......@@ -167,14 +185,21 @@ def mask_heads(args, model, eval_dataloader):
print_2d_tensor(new_head_mask)
# Compute metric and head importance again
_, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, head_mask=new_head_mask)
_, head_importance, preds, labels = compute_heads_importance(
args, model, eval_dataloader, compute_entropy=False, head_mask=new_head_mask
)
preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
current_score = compute_metrics(args.task_name, preds, labels)[args.metric_name]
logger.info("Masking: current score: %f, remaning heads %d (%.1f percents)", current_score, new_head_mask.sum(), new_head_mask.sum()/new_head_mask.numel() * 100)
logger.info(
"Masking: current score: %f, remaning heads %d (%.1f percents)",
current_score,
new_head_mask.sum(),
new_head_mask.sum() / new_head_mask.numel() * 100,
)
logger.info("Final head mask")
print_2d_tensor(head_mask)
np.save(os.path.join(args.output_dir, 'head_mask.npy'), head_mask.detach().cpu().numpy())
np.save(os.path.join(args.output_dir, "head_mask.npy"), head_mask.detach().cpu().numpy())
return head_mask
......@@ -186,8 +211,9 @@ def prune_heads(args, model, eval_dataloader, head_mask):
# Try pruning and test time speedup
# Pruning is like masking but we actually remove the masked weights
before_time = datetime.now()
_, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
compute_entropy=False, compute_importance=False, head_mask=head_mask)
_, _, preds, labels = compute_heads_importance(
args, model, eval_dataloader, compute_entropy=False, compute_importance=False, head_mask=head_mask
)
preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
score_masking = compute_metrics(args.task_name, preds, labels)[args.metric_name]
original_time = datetime.now() - before_time
......@@ -199,73 +225,127 @@ def prune_heads(args, model, eval_dataloader, head_mask):
pruned_num_params = sum(p.numel() for p in model.parameters())
before_time = datetime.now()
_, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
compute_entropy=False, compute_importance=False, head_mask=None)
_, _, preds, labels = compute_heads_importance(
args, model, eval_dataloader, compute_entropy=False, compute_importance=False, head_mask=None
)
preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
score_pruning = compute_metrics(args.task_name, preds, labels)[args.metric_name]
new_time = datetime.now() - before_time
logger.info("Pruning: original num of params: %.2e, after pruning %.2e (%.1f percents)", original_num_params, pruned_num_params, pruned_num_params/original_num_params * 100)
logger.info(
"Pruning: original num of params: %.2e, after pruning %.2e (%.1f percents)",
original_num_params,
pruned_num_params,
pruned_num_params / original_num_params * 100,
)
logger.info("Pruning: score with masking: %f score with pruning: %f", score_masking, score_pruning)
logger.info("Pruning: speed ratio (new timing / original timing): %f percents", original_time/new_time * 100)
logger.info("Pruning: speed ratio (new timing / original timing): %f percents", original_time / new_time * 100)
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 .tsv files (or other data files) for the task.")
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("--task_name", default=None, type=str, required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model predictions and checkpoints will be written.")
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_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(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
)
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("--config_name", default="", type=str,
help="Pretrained config name or path if not the same as model_name_or_path")
parser.add_argument("--tokenizer_name", default="", type=str,
help="Pretrained tokenizer name or path if not the same as model_name_or_path")
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("--data_subset", type=int, default=-1,
help="If > 0: limit the data to a subset of data_subset instances.")
parser.add_argument("--overwrite_output_dir", action='store_true',
help="Whether to overwrite data in output directory")
parser.add_argument('--overwrite_cache', action='store_true',
help="Overwrite the cached training and evaluation sets")
parser.add_argument("--dont_normalize_importance_by_layer", action='store_true',
help="Don't normalize importance score by layers")
parser.add_argument("--dont_normalize_global_importance", action='store_true',
help="Don't normalize all importance scores between 0 and 1")
parser.add_argument("--try_masking", action='store_true',
help="Whether to try to mask head until a threshold of accuracy.")
parser.add_argument("--masking_threshold", default=0.9, type=float,
help="masking threshold in term of metrics (stop masking when metric < threshold * original metric value).")
parser.add_argument("--masking_amount", default=0.1, type=float,
help="Amount to heads to masking at each masking step.")
parser.add_argument("--metric_name", default="acc", type=str,
help="Metric to use for head masking.")
parser.add_argument("--max_seq_length", default=128, type=int,
parser.add_argument(
"--config_name",
default="",
type=str,
help="Pretrained config name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Pretrained tokenizer name or path if not the same as model_name_or_path",
)
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(
"--data_subset", type=int, default=-1, help="If > 0: limit the data to a subset of data_subset instances."
)
parser.add_argument(
"--overwrite_output_dir", action="store_true", help="Whether to overwrite data in output directory"
)
parser.add_argument(
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
)
parser.add_argument(
"--dont_normalize_importance_by_layer", action="store_true", help="Don't normalize importance score by layers"
)
parser.add_argument(
"--dont_normalize_global_importance",
action="store_true",
help="Don't normalize all importance scores between 0 and 1",
)
parser.add_argument(
"--try_masking", action="store_true", help="Whether to try to mask head until a threshold of accuracy."
)
parser.add_argument(
"--masking_threshold",
default=0.9,
type=float,
help="masking threshold in term of metrics (stop masking when metric < threshold * original metric value).",
)
parser.add_argument(
"--masking_amount", default=0.1, type=float, help="Amount to heads to masking at each masking step."
)
parser.add_argument("--metric_name", default="acc", type=str, help="Metric to use for head masking.")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, sequences shorter padded.")
"Sequences longer than this will be truncated, sequences shorter padded.",
)
parser.add_argument("--batch_size", default=1, type=int, help="Batch size.")
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available")
parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
parser.add_argument("--no_cuda", action="store_true", help="Whether not to use CUDA when available")
parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
args = parser.parse_args()
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()
......@@ -278,10 +358,10 @@ def main():
torch.cuda.set_device(args.local_rank)
args.device = torch.device("cuda", args.local_rank)
args.n_gpu = 1
torch.distributed.init_process_group(backend='nccl') # Initializes the distributed backend
torch.distributed.init_process_group(backend="nccl") # Initializes the distributed backend
# Setup logging
logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logging.basicConfig(level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info("device: {} n_gpu: {}, distributed: {}".format(args.device, args.n_gpu, bool(args.local_rank != -1)))
# Set seeds
......@@ -306,17 +386,23 @@ def main():
args.model_type = key # take the first match in model types
break
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,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
output_attentions=True,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
cache_dir=args.cache_dir if args.cache_dir else None,
)
tokenizer = tokenizer_class.from_pretrained(
args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
......@@ -324,14 +410,14 @@ def main():
# Distributed and parallel training
model.to(args.device)
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)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
)
elif args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Print/save training arguments
torch.save(args, os.path.join(args.output_dir, 'run_args.bin'))
torch.save(args, os.path.join(args.output_dir, "run_args.bin"))
logger.info("Training/evaluation parameters %s", args)
# Prepare dataset for the GLUE task
......@@ -341,11 +427,9 @@ def main():
eval_sampler = SequentialSampler(eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)
# Compute head entropy and importance score
compute_heads_importance(args, model, eval_dataloader)
# Try head masking (set heads to zero until the score goes under a threshole)
# and head pruning (remove masked heads and see the effect on the network)
if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
......@@ -353,5 +437,5 @@ def main():
prune_heads(args, model, eval_dataloader, head_mask)
if __name__ == '__main__':
if __name__ == "__main__":
main()
examples/run_generation.py
View file @ fa84ae26
......@@ -33,9 +33,7 @@ from transformers import XLMWithLMHeadModel, XLMTokenizer
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO,
)
logger = logging.getLogger(__name__)
......@@ -71,6 +69,7 @@ def set_seed(args):
if args.n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
#
# Functions to prepare models' input
#
......@@ -78,15 +77,11 @@ def set_seed(args):
def prepare_ctrl_input(args, _, tokenizer, prompt_text):
if args.temperature > 0.7:
logger.info(
"CTRL typically works better with lower temperatures (and lower top_k)."
)
logger.info("CTRL typically works better with lower temperatures (and lower top_k).")
encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)
if not any(encoded_prompt[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"
)
logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
return prompt_text
......@@ -102,11 +97,7 @@ def prepare_xlm_input(args, model, tokenizer, prompt_text):
else:
language = None
while language not in available_languages:
language = input(
"Using XLM. Select language in "
+ str(list(available_languages))
+ " >>> "
)
language = input("Using XLM. Select language in " + str(list(available_languages)) + " >>> ")
# kwargs["language"] = tokenizer.lang2id[language]
# TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
......@@ -148,17 +139,34 @@ def adjust_length_to_model(length, max_sequence_length):
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model_type", default=None, type=str, required=True,
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
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(MODEL_CLASSES.keys()),
)
parser.add_argument("--prompt", type=str, default="")
parser.add_argument("--length", type=int, default=20)
parser.add_argument("--stop_token", type=str, default=None, help="Token at which text generation is stopped")
parser.add_argument("--temperature", type=float, default=1.0, help="temperature of 1.0 has no effect, lower tend toward 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(
"--temperature",
type=float,
default=1.0,
help="temperature of 1.0 has no effect, lower tend toward 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("--k", type=int, default=0)
parser.add_argument("--p", type=float, default=0.9)
......@@ -169,9 +177,7 @@ def main():
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
args = parser.parse_args()
args.device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu"
)
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
set_seed(args)
......@@ -181,17 +187,13 @@ def main():
args.model_type = args.model_type.lower()
model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
except KeyError:
raise KeyError(
"the model {} you specified is not supported. You are welcome to add it and open a PR :)"
)
raise KeyError("the model {} you specified is not supported. You are welcome to add it and open a PR :)")
tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
model = model_class.from_pretrained(args.model_name_or_path)
model.to(args.device)
args.length = adjust_length_to_model(
args.length, max_sequence_length=model.config.max_position_embeddings
)
args.length = adjust_length_to_model(args.length, max_sequence_length=model.config.max_position_embeddings)
logger.info(args)
prompt_text = args.prompt if args.prompt else input("Model prompt >>> ")
......@@ -201,7 +203,7 @@ def main():
if requires_preprocessing:
prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
prompt_text = prepare_input(args, model, tokenizer, prompt_text)
encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors='pt')
encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors="pt")
output_sequences = model.generate(
input_ids=encoded_prompt,
......
examples/run_glue.py
View file @ fa84ae26
......@@ -26,8 +26,7 @@ import json
import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
try:
......@@ -37,13 +36,18 @@ except:
from tqdm import tqdm, trange
from transformers import (WEIGHTS_NAME, BertConfig,
BertForSequenceClassification, BertTokenizer,
from transformers import (
WEIGHTS_NAME,
BertConfig,
BertForSequenceClassification,
BertTokenizer,
RobertaConfig,
RobertaForSequenceClassification,
RobertaTokenizer,
XLMConfig, XLMForSequenceClassification,
XLMTokenizer, XLNetConfig,
XLMConfig,
XLMForSequenceClassification,
XLMTokenizer,
XLNetConfig,
XLNetForSequenceClassification,
XLNetTokenizer,
DistilBertConfig,
......@@ -55,7 +59,7 @@ from transformers import (WEIGHTS_NAME, BertConfig,
XLMRobertaConfig,
XLMRobertaForSequenceClassification,
XLMRobertaTokenizer,
)
)
from transformers import AdamW, get_linear_schedule_with_warmup
......@@ -66,17 +70,22 @@ from transformers import glue_convert_examples_to_features as convert_examples_t
logger = logging.getLogger(__name__)
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig,
RobertaConfig, DistilBertConfig)), ())
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),
'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer),
'xlmroberta': (XLMRobertaConfig, XLMRobertaForSequenceClassification, XLMRobertaTokenizer),
"bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
"xlnet": (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
"xlm": (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
"roberta": (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
"distilbert": (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
"albert": (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer),
"xlmroberta": (XLMRobertaConfig, XLMRobertaForSequenceClassification, XLMRobertaTokenizer),
}
......@@ -104,20 +113,27 @@ def train(args, train_dataset, model, tokenizer):
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']
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}
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
......@@ -132,17 +148,21 @@ def train(args, train_dataset, model, tokenizer):
# 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)
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(
" 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)
......@@ -152,7 +172,7 @@ def train(args, train_dataset, model, tokenizer):
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
......@@ -163,7 +183,9 @@ def train(args, train_dataset, model, tokenizer):
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
train_iterator = trange(
epochs_trained, 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])
......@@ -176,11 +198,11 @@ def train(args, train_dataset, model, tokenizer):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'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
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "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 transformers (see doc)
......@@ -209,36 +231,40 @@ def train(args, train_dataset, model, tokenizer):
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
logs = {}
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
eval_key = 'eval_{}'.format(key)
eval_key = "eval_{}".format(key)
logs[eval_key] = value
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
learning_rate_scalar = scheduler.get_lr()[0]
logs['learning_rate'] = learning_rate_scalar
logs['loss'] = loss_scalar
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
logging_loss = tr_loss
for key, value in logs.items():
tb_writer.add_scalar(key, value, global_step)
print(json.dumps({**logs, **{'step': global_step}}))
print(json.dumps({**logs, **{"step": global_step}}))
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))
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 = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
......@@ -257,7 +283,7 @@ def train(args, train_dataset, model, tokenizer):
def evaluate(args, model, tokenizer, prefix=""):
# Loop to handle MNLI double evaluation (matched, mis-matched)
eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,)
eval_outputs_dirs = (args.output_dir, args.output_dir + "-MM") if args.task_name == "mnli" else (args.output_dir,)
results = {}
for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
......@@ -288,11 +314,11 @@ 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],
'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
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "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]
......@@ -300,10 +326,10 @@ def evaluate(args, model, tokenizer, prefix=""):
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs['labels'].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)
out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
......@@ -330,29 +356,36 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
processor = processors[task]()
output_mode = output_modes[task]
# Load data features from cache or dataset file
cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format(
'dev' if evaluate else 'train',
list(filter(None, args.model_name_or_path.split('/'))).pop(),
cached_features_file = os.path.join(
args.data_dir,
"cached_{}_{}_{}_{}".format(
"dev" if evaluate else "train",
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
str(task)))
str(task),
),
)
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)
label_list = processor.get_labels()
if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta', 'xlmroberta']:
if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta", "xlmroberta"]:
# 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,
examples = (
processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
)
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_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_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)
......@@ -378,90 +411,149 @@ 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 .tsv files (or other data files) for the 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("--task_name", default=None, type=str, required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model predictions and checkpoints will be written.")
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=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(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
)
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("--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,
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("--evaluate_during_training", action='store_true',
help="Rul evaluation during training at each logging step.")
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--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',
"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(
"--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
)
parser.add_argument(
"--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
)
parser.add_argument("--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.")
"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))
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()
......@@ -473,16 +565,24 @@ def main():
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')
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)
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)
......@@ -502,17 +602,23 @@ def main():
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,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
......@@ -521,14 +627,12 @@ def main():
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use 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
......@@ -538,36 +642,39 @@ def main():
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 = (
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'))
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
model.to(args.device)
# 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)))
checkpoints = list(
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
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=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
return results
......
examples/run_lm_finetuning.py
View file @ fa84ae26
......@@ -42,37 +42,55 @@ except:
from tqdm import tqdm, trange
from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup,
BertConfig, BertForMaskedLM, BertTokenizer,
GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
RobertaConfig, RobertaForMaskedLM, RobertaTokenizer,
DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer,
CamembertConfig, CamembertForMaskedLM, CamembertTokenizer)
from transformers import (
WEIGHTS_NAME,
AdamW,
get_linear_schedule_with_warmup,
BertConfig,
BertForMaskedLM,
BertTokenizer,
GPT2Config,
GPT2LMHeadModel,
GPT2Tokenizer,
OpenAIGPTConfig,
OpenAIGPTLMHeadModel,
OpenAIGPTTokenizer,
RobertaConfig,
RobertaForMaskedLM,
RobertaTokenizer,
DistilBertConfig,
DistilBertForMaskedLM,
DistilBertTokenizer,
CamembertConfig,
CamembertForMaskedLM,
CamembertTokenizer,
)
logger = logging.getLogger(__name__)
MODEL_CLASSES = {
'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
'camembert': (CamembertConfig, CamembertForMaskedLM, CamembertTokenizer)
"gpt2": (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
"openai-gpt": (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
"bert": (BertConfig, BertForMaskedLM, BertTokenizer),
"roberta": (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
"distilbert": (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
"camembert": (CamembertConfig, CamembertForMaskedLM, CamembertTokenizer),
}
class TextDataset(Dataset):
def __init__(self, tokenizer, args, file_path='train', block_size=512):
def __init__(self, tokenizer, args, 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, args.model_name_or_path + '_cached_lm_' + str(block_size) + '_' + filename)
cached_features_file = os.path.join(
directory, args.model_name_or_path + "_cached_lm_" + str(block_size) + "_" + filename
)
if os.path.exists(cached_features_file) and not args.overwrite_cache:
logger.info("Loading features from cached file %s", cached_features_file)
with open(cached_features_file, 'rb') as handle:
with open(cached_features_file, "rb") as handle:
self.examples = pickle.load(handle)
else:
logger.info("Creating features from dataset file at %s", directory)
......@@ -83,14 +101,14 @@ class TextDataset(Dataset):
tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))
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]))
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.
logger.info("Saving features into cached file %s", cached_features_file)
with open(cached_features_file, 'wb') as handle:
with open(cached_features_file, "wb") as handle:
pickle.dump(self.examples, handle, protocol=pickle.HIGHEST_PROTOCOL)
def __len__(self):
......@@ -101,7 +119,12 @@ class TextDataset(Dataset):
def load_and_cache_examples(args, tokenizer, evaluate=False):
dataset = TextDataset(tokenizer, args, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
dataset = TextDataset(
tokenizer,
args,
file_path=args.eval_data_file if evaluate else args.train_data_file,
block_size=args.block_size,
)
return dataset
......@@ -120,7 +143,7 @@ def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False):
return
# Check if we should delete older checkpoint(s)
glob_checkpoints = glob.glob(os.path.join(args.output_dir, '{}-*'.format(checkpoint_prefix)))
glob_checkpoints = glob.glob(os.path.join(args.output_dir, "{}-*".format(checkpoint_prefix)))
if len(glob_checkpoints) <= args.save_total_limit:
return
......@@ -129,7 +152,7 @@ def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False):
if use_mtime:
ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
else:
regex_match = re.match('.*{}-([0-9]+)'.format(checkpoint_prefix), path)
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))
......@@ -147,7 +170,9 @@ def mask_tokens(inputs, tokenizer, args):
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)
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()]
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] = -100 # We only compute loss on masked tokens
......@@ -181,19 +206,26 @@ def train(args, train_dataset, model, tokenizer):
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']
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}
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
......@@ -208,17 +240,21 @@ def train(args, train_dataset, model, tokenizer):
# 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)
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(
" 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)
......@@ -228,7 +264,7 @@ def train(args, train_dataset, model, tokenizer):
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
......@@ -239,11 +275,13 @@ def train(args, train_dataset, model, tokenizer):
tr_loss, logging_loss = 0.0, 0.0
model_to_resize = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
model_to_resize = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training
model_to_resize.resize_token_embeddings(len(tokenizer))
model.zero_grad()
train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
train_iterator = trange(
epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
)
set_seed(args) # Added here for reproducibility (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])
......@@ -285,31 +323,35 @@ def train(args, train_dataset, model, tokenizer):
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
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
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)
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:
checkpoint_prefix = 'checkpoint'
checkpoint_prefix = "checkpoint"
# Save model checkpoint
output_dir = os.path.join(args.output_dir, '{}-{}'.format(checkpoint_prefix, 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
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
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)
torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
......@@ -365,9 +407,7 @@ def evaluate(args, model, tokenizer, prefix=""):
eval_loss = eval_loss / nb_eval_steps
perplexity = torch.exp(torch.tensor(eval_loss))
result = {
"perplexity": perplexity
}
result = {"perplexity": perplexity}
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
with open(output_eval_file, "w") as writer:
......@@ -383,107 +423,167 @@ def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_data_file", default=None, type=str, required=True,
help="The input training data file (a text file).")
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model predictions and checkpoints will be written.")
parser.add_argument(
"--train_data_file", default=None, type=str, required=True, help="The input training data file (a text file)."
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.",
)
## Other parameters
parser.add_argument("--eval_data_file", default=None, type=str,
help="An optional input evaluation data file to evaluate the perplexity on (a text file).")
parser.add_argument("--model_type", default="bert", type=str,
help="The model architecture to be fine-tuned.")
parser.add_argument("--model_name_or_path", default="bert-base-cased", type=str,
help="The model checkpoint for weights initialization.")
parser.add_argument("--mlm", action='store_true',
help="Train with masked-language modeling loss instead of language modeling.")
parser.add_argument("--mlm_probability", type=float, default=0.15,
help="Ratio of tokens to mask for masked language modeling loss")
parser.add_argument("--config_name", default="", type=str,
help="Optional pretrained config name or path if not the same as model_name_or_path")
parser.add_argument("--tokenizer_name", default="", type=str,
help="Optional pretrained tokenizer name or path if not the same as model_name_or_path")
parser.add_argument("--cache_dir", default="", type=str,
help="Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)")
parser.add_argument("--block_size", default=-1, type=int,
parser.add_argument(
"--eval_data_file",
default=None,
type=str,
help="An optional input evaluation data file to evaluate the perplexity on (a text file).",
)
parser.add_argument("--model_type", default="bert", type=str, help="The model architecture to be fine-tuned.")
parser.add_argument(
"--model_name_or_path",
default="bert-base-cased",
type=str,
help="The model checkpoint for weights initialization.",
)
parser.add_argument(
"--mlm", action="store_true", help="Train with masked-language modeling loss instead of language modeling."
)
parser.add_argument(
"--mlm_probability", type=float, default=0.15, help="Ratio of tokens to mask for masked language modeling loss"
)
parser.add_argument(
"--config_name",
default="",
type=str,
help="Optional pretrained config name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
help="Optional pretrained tokenizer name or path if not the same as model_name_or_path",
)
parser.add_argument(
"--cache_dir",
default="",
type=str,
help="Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)",
)
parser.add_argument(
"--block_size",
default=-1,
type=int,
help="Optional input sequence length after tokenization."
"The training dataset will be truncated in block of this size for training."
"Default to the model max input length for single sentence inputs (take into account special tokens).")
parser.add_argument("--do_train", action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval", action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--evaluate_during_training", action='store_true',
help="Run evaluation during training at each logging step.")
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--per_gpu_train_batch_size", default=4, type=int,
help="Batch size per GPU/CPU for training.")
parser.add_argument("--per_gpu_eval_batch_size", default=4, type=int,
help="Batch size per GPU/CPU for evaluation.")
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--learning_rate", default=5e-5, type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay", default=0.0, type=float,
help="Weight 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=1.0, type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--max_steps", default=-1, type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
parser.add_argument("--warmup_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('--save_total_limit', type=int, default=None,
help='Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default')
parser.add_argument("--eval_all_checkpoints", action='store_true',
help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number")
parser.add_argument("--no_cuda", action='store_true',
help="Avoid using CUDA when available")
parser.add_argument('--overwrite_output_dir', action='store_true',
help="Overwrite the content of the output directory")
parser.add_argument('--overwrite_cache', action='store_true',
help="Overwrite the cached training and evaluation sets")
parser.add_argument('--seed', type=int, default=42,
help="random seed for initialization")
parser.add_argument('--fp16', action='store_true',
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
parser.add_argument('--fp16_opt_level', type=str, default='O1',
"Default to the model max input length for single sentence inputs (take into account special tokens).",
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
)
parser.add_argument(
"--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
)
parser.add_argument("--per_gpu_train_batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.")
parser.add_argument(
"--per_gpu_eval_batch_size", default=4, type=int, help="Batch size per GPU/CPU for evaluation."
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight 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=1.0, type=float, help="Total number of training epochs to perform."
)
parser.add_argument(
"--max_steps",
default=-1,
type=int,
help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
)
parser.add_argument("--warmup_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(
"--save_total_limit",
type=int,
default=None,
help="Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default",
)
parser.add_argument(
"--eval_all_checkpoints",
action="store_true",
help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number",
)
parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
parser.add_argument(
"--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
)
parser.add_argument(
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
)
parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
)
parser.add_argument(
"--fp16_opt_level",
type=str,
default="O1",
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument("--local_rank", type=int, default=-1,
help="For distributed training: local_rank")
parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
args = parser.parse_args()
if args.model_type in ["bert", "roberta", "distilbert", "camembert"] 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).")
raise ValueError(
"BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm "
"flag (masked language modeling)."
)
if args.eval_data_file is None and args.do_eval:
raise ValueError("Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
raise ValueError(
"Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
"or remove the --do_eval argument."
)
if (
os.path.exists(args.output_dir)
and os.listdir(args.output_dir)
and args.do_train
and not args.overwrite_output_dir
):
raise ValueError(
"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()
......@@ -495,16 +595,24 @@ def main():
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')
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)
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)
......@@ -514,18 +622,26 @@ def main():
torch.distributed.barrier() # Barrier to make sure only the first process in distributed training download model & vocab
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.block_size <= 0:
args.block_size = tokenizer.max_len_single_sentence # Our input block size will be the max possible for the model
args.block_size = (
tokenizer.max_len_single_sentence
) # Our input block size will be the max possible for the model
args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
model.to(args.device)
if args.local_rank == 0:
......@@ -546,7 +662,6 @@ def main():
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
# Create output directory if needed
......@@ -556,35 +671,38 @@ def main():
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 = (
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'))
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
checkpoints = list(
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
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=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
return results
......
examples/run_multiple_choice.py
View file @ fa84ae26
......@@ -26,8 +26,7 @@ import random
import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
try:
......@@ -37,34 +36,38 @@ except:
from tqdm import tqdm, trange
from transformers import (WEIGHTS_NAME, BertConfig,
BertForMultipleChoice, BertTokenizer,
XLNetConfig, XLNetForMultipleChoice,
XLNetTokenizer, RobertaConfig,
RobertaForMultipleChoice, RobertaTokenizer)
from transformers import (
WEIGHTS_NAME,
BertConfig,
BertForMultipleChoice,
BertTokenizer,
XLNetConfig,
XLNetForMultipleChoice,
XLNetTokenizer,
RobertaConfig,
RobertaForMultipleChoice,
RobertaTokenizer,
)
from transformers import AdamW, get_linear_schedule_with_warmup
from utils_multiple_choice import (convert_examples_to_features, processors)
from utils_multiple_choice import convert_examples_to_features, processors
logger = logging.getLogger(__name__)
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, RobertaConfig)), ())
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, RobertaConfig)), ()
)
MODEL_CLASSES = {
'bert': (BertConfig, BertForMultipleChoice, BertTokenizer),
'xlnet': (XLNetConfig, XLNetForMultipleChoice, XLNetTokenizer),
'roberta': (RobertaConfig, RobertaForMultipleChoice, RobertaTokenizer)
"bert": (BertConfig, BertForMultipleChoice, BertTokenizer),
"xlnet": (XLNetConfig, XLNetForMultipleChoice, XLNetTokenizer),
"roberta": (RobertaConfig, RobertaForMultipleChoice, RobertaTokenizer),
}
def select_field(features, field):
return [
[
choice[field]
for choice in feature.choices_features
]
for feature in features
]
return [[choice[field] for choice in feature.choices_features] for feature in features]
def simple_accuracy(preds, labels):
......@@ -95,13 +98,18 @@ def train(args, train_dataset, model, tokenizer):
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']
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}
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
if args.fp16:
try:
from apex import amp
......@@ -115,17 +123,21 @@ def train(args, train_dataset, model, tokenizer):
# 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)
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(
" 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)
......@@ -141,10 +153,14 @@ def train(args, train_dataset, model, tokenizer):
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 don't use segment_ids
'labels': batch[3]}
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"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 transformers (see doc)
......@@ -171,10 +187,12 @@ def train(args, train_dataset, model, tokenizer):
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
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
tb_writer.add_scalar("eval_{}".format(key), value, global_step)
if results["eval_acc"] > best_dev_acc:
best_dev_acc = results["eval_acc"]
best_dev_loss = results["eval_loss"]
......@@ -182,22 +200,33 @@ def train(args, train_dataset, model, tokenizer):
if args.do_test:
results_test = evaluate(args, model, tokenizer, test=True)
for key, value in results_test.items():
tb_writer.add_scalar('test_{}'.format(key), value, global_step)
logger.info("test acc: %s, loss: %s, global steps: %s", str(results_test['eval_acc']), str(results_test['eval_loss']), str(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)
logger.info("Average loss: %s at global step: %s", str((tr_loss - logging_loss)/args.logging_steps), str(global_step))
tb_writer.add_scalar("test_{}".format(key), value, global_step)
logger.info(
"test acc: %s, loss: %s, global steps: %s",
str(results_test["eval_acc"]),
str(results_test["eval_loss"]),
str(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)
logger.info(
"Average loss: %s at global step: %s",
str((tr_loss - logging_loss) / args.logging_steps),
str(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))
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 = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_vocabulary(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
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:
......@@ -246,10 +275,14 @@ def evaluate(args, model, tokenizer, prefix="", test=False):
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 don't use segment_ids
'labels': batch[3]}
inputs = {
"input_ids": batch[0],
"attention_mask": batch[1],
"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)
tmp_eval_loss, logits = outputs[:2]
......@@ -257,10 +290,10 @@ def evaluate(args, model, tokenizer, prefix="", test=False):
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs['labels'].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)
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=1)
......@@ -273,8 +306,14 @@ def evaluate(args, model, tokenizer, prefix="", test=False):
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(str(prefix) + " is test:" + str(test)))
writer.write("model =%s\n" % str(args.model_name_or_path))
writer.write("total batch size=%d\n" % (args.per_gpu_train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1)))
writer.write(
"total batch size=%d\n"
% (
args.per_gpu_train_batch_size
* args.gradient_accumulation_steps
* (torch.distributed.get_world_size() if args.local_rank != -1 else 1)
)
)
writer.write("train num epochs=%d\n" % args.num_train_epochs)
writer.write("fp16 =%s\n" % args.fp16)
writer.write("max seq length =%d\n" % args.max_seq_length)
......@@ -291,17 +330,21 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
processor = processors[task]()
# Load data features from cache or dataset file
if evaluate:
cached_mode = 'dev'
cached_mode = "dev"
elif test:
cached_mode = 'test'
cached_mode = "test"
else:
cached_mode = 'train'
cached_mode = "train"
assert (evaluate == True and test == True) == False
cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format(
cached_features_file = os.path.join(
args.data_dir,
"cached_{}_{}_{}_{}".format(
cached_mode,
list(filter(None, args.model_name_or_path.split('/'))).pop(),
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
str(task)))
str(task),
),
)
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)
......@@ -320,8 +363,8 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
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_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,
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s", cached_features_file)
......@@ -331,9 +374,9 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
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(select_field(features, 'input_ids'), dtype=torch.long)
all_input_mask = torch.tensor(select_field(features, 'input_mask'), dtype=torch.long)
all_segment_ids = torch.tensor(select_field(features, 'segment_ids'), dtype=torch.long)
all_input_ids = torch.tensor(select_field(features, "input_ids"), dtype=torch.long)
all_input_mask = torch.tensor(select_field(features, "input_mask"), dtype=torch.long)
all_segment_ids = torch.tensor(select_field(features, "segment_ids"), dtype=torch.long)
all_label_ids = torch.tensor([f.label for f in features], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
......@@ -344,91 +387,150 @@ 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 .tsv files (or other data files) for the 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("--task_name", default=None, type=str, required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model predictions and checkpoints will be written.")
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=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(
"--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
)
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("--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,
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_test", action='store_true', help='Whether to run test on the test set')
parser.add_argument("--evaluate_during_training", action='store_true',
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.")
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 deay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs", default=3.0, 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',
"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_test", action="store_true", help="Whether to run test on the test set")
parser.add_argument(
"--evaluate_during_training", action="store_true", 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."
)
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 deay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs", default=3.0, 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.")
"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))
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()
......@@ -440,16 +542,24 @@ def main():
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')
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)
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)
......@@ -468,17 +578,23 @@ def main():
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,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
......@@ -494,7 +610,6 @@ def main():
global_step, tr_loss, best_steps = train(args, train_dataset, model, tokenizer)
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
......@@ -504,19 +619,20 @@ def main():
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 = (
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'))
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
model.to(args.device)
# Evaluation
results = {}
if args.do_eval and args.local_rank in [-1, 0]:
......@@ -524,17 +640,19 @@ def main():
args.output_dir = args.model_name_or_path
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)))
checkpoints = list(
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
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=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
if args.do_test and args.local_rank in [-1, 0]:
......@@ -546,13 +664,13 @@ def main():
# 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 ""
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=prefix, test=True)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
if best_steps:
logger.info("best steps of eval acc is the following checkpoints: %s", best_steps)
......
examples/run_ner.py
View file @ fa84ae26
......@@ -43,9 +43,12 @@ from transformers import XLMRobertaConfig, XLMRobertaForTokenClassification, XLM
logger = logging.getLogger(__name__)
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig,
CamembertConfig, XLMRobertaConfig)),
())
(
tuple(conf.pretrained_config_archive_map.keys())
for conf in (BertConfig, RobertaConfig, DistilBertConfig, CamembertConfig, XLMRobertaConfig)
),
(),
)
MODEL_CLASSES = {
"bert": (BertConfig, BertForTokenClassification, BertTokenizer),
......@@ -82,18 +85,24 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
# 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}
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
......@@ -108,18 +117,21 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
# 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)
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(
" 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)
......@@ -129,7 +141,7 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
......@@ -140,7 +152,9 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
train_iterator = trange(
epochs_trained, 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])
......@@ -153,11 +167,11 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3]}
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
if args.model_type != "distilbert":
inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None # XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert", "xlnet"] else None
) # XLM and RoBERTa don"t use segment_ids
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
......@@ -187,7 +201,9 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
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
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, mode="dev")
for key, value in results.items():
tb_writer.add_scalar("eval_{}".format(key), value, global_step)
......@@ -200,15 +216,17 @@ def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
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 = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
......@@ -249,11 +267,11 @@ def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {"input_ids": batch[0],
"attention_mask": batch[1],
"labels": batch[3]}
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
if args.model_type != "distilbert":
inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None # XLM and RoBERTa don"t use segment_ids
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert", "xlnet"] else None
) # XLM and RoBERTa don"t use segment_ids
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
......@@ -287,7 +305,7 @@ def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""
"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)
"f1": f1_score(out_label_list, preds_list),
}
logger.info("***** Eval results %s *****", prefix)
......@@ -302,16 +320,23 @@ def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
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)))
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,
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,
......@@ -323,7 +348,7 @@ def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
# 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
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)
......@@ -346,95 +371,151 @@ 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.")
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,
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",
"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")
"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:
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))
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()
......@@ -451,11 +532,19 @@ def main():
args.device = device
# Setup logging
logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
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)
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)
......@@ -472,16 +561,22 @@ def main():
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,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
......@@ -505,7 +600,9 @@ def main():
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 = (
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)
......@@ -518,7 +615,9 @@ def main():
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)))
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:
......@@ -565,4 +664,3 @@ def main():
if __name__ == "__main__":
main()
examples/run_squad.py
View file @ fa84ae26
......@@ -17,7 +17,11 @@
from __future__ import absolute_import, division, print_function
from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor, SquadResult
from transformers.data.metrics.squad_metrics import compute_predictions_logits, compute_predictions_log_probs, squad_evaluate
from transformers.data.metrics.squad_metrics import (
compute_predictions_logits,
compute_predictions_log_probs,
squad_evaluate,
)
import argparse
import logging
......@@ -27,8 +31,7 @@ import glob
import timeit
import numpy as np
import torch
from torch.utils.data import (
DataLoader, RandomSampler, SequentialSampler, TensorDataset)
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
try:
......@@ -38,32 +41,47 @@ except:
from tqdm import tqdm, trange
from transformers import (WEIGHTS_NAME, BertConfig,
BertForQuestionAnswering, BertTokenizer,
RobertaForQuestionAnswering, RobertaTokenizer, RobertaConfig,
XLMConfig, XLMForQuestionAnswering,
XLMTokenizer, XLNetConfig,
from transformers import (
WEIGHTS_NAME,
BertConfig,
BertForQuestionAnswering,
BertTokenizer,
RobertaForQuestionAnswering,
RobertaTokenizer,
RobertaConfig,
XLMConfig,
XLMForQuestionAnswering,
XLMTokenizer,
XLNetConfig,
XLNetForQuestionAnswering,
XLNetTokenizer,
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer,
XLMConfig, XLMForQuestionAnswering, XLMTokenizer,
)
DistilBertConfig,
DistilBertForQuestionAnswering,
DistilBertTokenizer,
AlbertConfig,
AlbertForQuestionAnswering,
AlbertTokenizer,
XLMConfig,
XLMForQuestionAnswering,
XLMTokenizer,
)
from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
logger = logging.getLogger(__name__)
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
for conf in (BertConfig, RobertaConfig, XLNetConfig, XLMConfig)), ())
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, XLNetConfig, XLMConfig)),
(),
)
MODEL_CLASSES = {
'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
'roberta': (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
"bert": (BertConfig, BertForQuestionAnswering, BertTokenizer),
"roberta": (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
"xlnet": (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
"xlm": (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
"distilbert": (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
"albert": (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
}
......@@ -85,49 +103,44 @@ def train(args, train_dataset, model, tokenizer):
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)
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
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
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']
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}
{
"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)
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(
os.path.join(args.model_name_or_path, 'optimizer.pt')))
scheduler.load_state_dict(torch.load(
os.path.join(args.model_name_or_path, 'scheduler.pt')))
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
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.")
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)
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:
......@@ -135,20 +148,22 @@ def train(args, train_dataset, model, tokenizer):
# 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)
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(" 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 = 1
......@@ -157,29 +172,25 @@ def train(args, train_dataset, model, tokenizer):
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
epochs_trained = global_step // (len(train_dataloader) //
args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (
len(train_dataloader) // args.gradient_accumulation_steps)
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
logger.info(
" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
logger.info(" Continuing training from epoch %d", epochs_trained)
logger.info(" Continuing training from global step %d", global_step)
logger.info(" Will skip the first %d steps in the first epoch",
steps_trained_in_current_epoch)
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(epochs_trained, int(
args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
train_iterator = trange(
epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
)
# Added here for reproductibility (even between python 2 and 3)
set_seed(args)
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader, desc="Iteration",
disable=args.local_rank not in [-1, 0])
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
......@@ -191,18 +202,17 @@ 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 in ['xlm', 'roberta', 'distilbert'] else batch[2],
'start_positions': batch[3],
'end_positions': batch[4],
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": None if args.model_type in ["xlm", "roberta", "distilbert"] else batch[2],
"start_positions": batch[3],
"end_positions": batch[4],
}
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[5],
'p_mask': batch[6]})
if args.model_type in ["xlnet", "xlm"]:
inputs.update({"cls_index": batch[5], "p_mask": batch[6]})
if args.version_2_with_negative:
inputs.update({'is_impossible': batch[7]})
inputs.update({"is_impossible": batch[7]})
outputs = model(**inputs)
# model outputs are always tuple in transformers (see doc)
loss = outputs[0]
......@@ -221,11 +231,9 @@ def train(args, train_dataset, model, tokenizer):
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)
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)
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
......@@ -238,36 +246,27 @@ def train(args, train_dataset, model, tokenizer):
if args.local_rank == -1 and args.evaluate_during_training:
results = evaluate(args, model, tokenizer)
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)
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
# Save model checkpoint
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
output_dir = os.path.join(
args.output_dir, 'checkpoint-{}'.format(global_step))
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(
model, 'module') else model
model_to_save = model.module if hasattr(model, "module") else model
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(
output_dir, 'training_args.bin'))
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(
output_dir, 'optimizer.pt'))
torch.save(scheduler.state_dict(), os.path.join(
output_dir, 'scheduler.pt'))
logger.info(
"Saving optimizer and scheduler states to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
......@@ -283,8 +282,7 @@ def train(args, train_dataset, model, tokenizer):
def evaluate(args, model, tokenizer, prefix=""):
dataset, examples, features = load_and_cache_examples(
args, tokenizer, evaluate=True, output_examples=True)
dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
......@@ -293,8 +291,7 @@ def evaluate(args, model, tokenizer, prefix=""):
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(dataset)
eval_dataloader = DataLoader(
dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
# multi-gpu evaluate
if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
......@@ -314,15 +311,15 @@ def evaluate(args, model, tokenizer, prefix=""):
with torch.no_grad():
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
'token_type_ids': None if args.model_type in ['xlm', 'roberta', 'distilbert'] else batch[2],
"input_ids": batch[0],
"attention_mask": batch[1],
"token_type_ids": None if args.model_type in ["xlm", "roberta", "distilbert"] else batch[2],
}
example_indices = batch[3]
# XLNet and XLM use more arguments for their predictions
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
if args.model_type in ["xlnet", "xlm"]:
inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
outputs = model(**inputs)
......@@ -342,53 +339,68 @@ def evaluate(args, model, tokenizer, prefix=""):
cls_logits = output[4]
result = SquadResult(
unique_id, start_logits, end_logits,
unique_id,
start_logits,
end_logits,
start_top_index=start_top_index,
end_top_index=end_top_index,
cls_logits=cls_logits
cls_logits=cls_logits,
)
else:
start_logits, end_logits = output
result = SquadResult(
unique_id, start_logits, end_logits
)
result = SquadResult(unique_id, start_logits, end_logits)
all_results.append(result)
evalTime = timeit.default_timer() - start_time
logger.info(" Evaluation done in total %f secs (%f sec per example)",
evalTime, evalTime / len(dataset))
logger.info(" Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
# Compute predictions
output_prediction_file = os.path.join(
args.output_dir, "predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(
args.output_dir, "nbest_predictions_{}.json".format(prefix))
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
if args.version_2_with_negative:
output_null_log_odds_file = os.path.join(
args.output_dir, "null_odds_{}.json".format(prefix))
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
else:
output_null_log_odds_file = None
# XLNet and XLM use a more complex post-processing procedure
if args.model_type in ['xlnet', 'xlm']:
start_n_top = model.config.start_n_top if hasattr(
model, "config") else model.module.config.start_n_top
end_n_top = model.config.end_n_top if hasattr(
model, "config") else model.module.config.end_n_top
predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
args.max_answer_length, output_prediction_file,
output_nbest_file, output_null_log_odds_file,
start_n_top, end_n_top,
args.version_2_with_negative, tokenizer, args.verbose_logging)
if args.model_type in ["xlnet", "xlm"]:
start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
predictions = compute_predictions_log_probs(
examples,
features,
all_results,
args.n_best_size,
args.max_answer_length,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
start_n_top,
end_n_top,
args.version_2_with_negative,
tokenizer,
args.verbose_logging,
)
else:
predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
args.max_answer_length, args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative, args.null_score_diff_threshold, tokenizer)
predictions = compute_predictions_logits(
examples,
features,
all_results,
args.n_best_size,
args.max_answer_length,
args.do_lower_case,
output_prediction_file,
output_nbest_file,
output_null_log_odds_file,
args.verbose_logging,
args.version_2_with_negative,
args.null_score_diff_threshold,
tokenizer,
)
# Compute the F1 and exact scores.
results = squad_evaluate(examples, predictions)
......@@ -402,16 +414,18 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
# Load data features from cache or dataset file
input_dir = args.data_dir if args.data_dir else "."
cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
'dev' if evaluate else 'train',
list(filter(None, args.model_name_or_path.split('/'))).pop(),
str(args.max_seq_length))
cached_features_file = os.path.join(
input_dir,
"cached_{}_{}_{}".format(
"dev" if evaluate else "train",
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
),
)
# Init features and dataset from cache if it exists
if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
logger.info("Loading features from cached file %s",
cached_features_file)
logger.info("Loading features from cached file %s", cached_features_file)
features_and_dataset = torch.load(cached_features_file)
features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
else:
......@@ -421,16 +435,13 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
try:
import tensorflow_datasets as tfds
except ImportError:
raise ImportError(
"If not data_dir is specified, tensorflow_datasets needs to be installed.")
raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
if args.version_2_with_negative:
logger.warn(
"tensorflow_datasets does not handle version 2 of SQuAD.")
logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")
tfds_examples = tfds.load("squad")
examples = SquadV1Processor().get_examples_from_dataset(
tfds_examples, evaluate=evaluate)
examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
else:
processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
if evaluate:
......@@ -445,15 +456,13 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=not evaluate,
return_dataset='pt',
return_dataset="pt",
threads=args.threads,
)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s",
cached_features_file)
torch.save({"features": features, "dataset": dataset},
cached_features_file)
logger.info("Saving features into cached file %s", cached_features_file)
torch.save({"features": features, "dataset": dataset}, cached_features_file)
if args.local_rank == 0 and not evaluate:
# Make sure only the first process in distributed training process the dataset, and the others will use the cache
......@@ -468,140 +477,232 @@ def main():
parser = argparse.ArgumentParser()
# Required parameters
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 checkpoints and predictions will be written.")
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 checkpoints and predictions will be written.",
)
# Other parameters
parser.add_argument("--data_dir", default=None, type=str,
help="The input data dir. Should contain the .json files for the task." +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--train_file", default=None, type=str,
help="The input training file. If a data dir is specified, will look for the file there" +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--predict_file", default=None, type=str,
help="The input evaluation file. If a data dir is specified, will look for the file there" +
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
parser.add_argument("--config_name", default="", type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument("--tokenizer_name", default="", type=str,
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('--version_2_with_negative', action='store_true',
help='If true, the SQuAD examples contain some that do not have an answer.')
parser.add_argument('--null_score_diff_threshold', type=float, default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.")
parser.add_argument("--max_seq_length", default=384, type=int,
parser.add_argument(
"--data_dir",
default=None,
type=str,
help="The input data dir. Should contain the .json files for the task."
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--train_file",
default=None,
type=str,
help="The input training file. If a data dir is specified, will look for the file there"
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--predict_file",
default=None,
type=str,
help="The input evaluation file. If a data dir is specified, will look for the file there"
+ "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
)
parser.add_argument(
"--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
)
parser.add_argument(
"--tokenizer_name",
default="",
type=str,
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(
"--version_2_with_negative",
action="store_true",
help="If true, the SQuAD examples contain some that do not have an answer.",
)
parser.add_argument(
"--null_score_diff_threshold",
type=float,
default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.",
)
parser.add_argument(
"--max_seq_length",
default=384,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded.")
parser.add_argument("--doc_stride", default=128, type=int,
help="When splitting up a long document into chunks, how much stride to take between chunks.")
parser.add_argument("--max_query_length", default=64, type=int,
"longer than this will be truncated, and sequences shorter than this will be padded.",
)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help="When splitting up a long document into chunks, how much stride to take between chunks.",
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help="The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument("--do_train", action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval", action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--evaluate_during_training", action='store_true',
help="Rul evaluation during training at each logging step.")
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--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("--learning_rate", default=5e-5, type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--weight_decay", default=0.0, type=float,
help="Weight 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("--n_best_size", default=20, type=int,
help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
parser.add_argument("--max_answer_length", default=30, type=int,
"be truncated to this length.",
)
parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
parser.add_argument(
"--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
)
parser.add_argument(
"--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
)
parser.add_argument("--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("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight 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(
"--n_best_size",
default=20,
type=int,
help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
)
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument("--verbose_logging", action='store_true',
"and end predictions are not conditioned on one another.",
)
parser.add_argument(
"--verbose_logging",
action="store_true",
help="If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
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="Whether not to use 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("--local_rank", type=int, default=-1,
help="local_rank for distributed training on gpus")
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',
"A number of warnings are expected for a normal SQuAD evaluation.",
)
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="Whether not to use 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("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
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('--server_ip', type=str, default='', help="Can be used for distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
"See details at https://nvidia.github.io/apex/amp.html",
)
parser.add_argument("--server_ip", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
parser.add_argument('--threads', type=int, default=1, help='multiple threads for converting example to features')
parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
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:
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))
"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.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")
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')
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)
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)
......@@ -613,16 +714,21 @@ def main():
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,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=bool(
'.ckpt' in args.model_name_or_path),
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
# Make sure only the first process in distributed training will download model & vocab
......@@ -638,18 +744,16 @@ def main():
if args.fp16:
try:
import apex
apex.amp.register_half_function(torch, 'einsum')
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.")
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)
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s",
global_step, tr_loss)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Save the trained model and the tokenizer
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
......@@ -661,18 +765,16 @@ def main():
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
# They can then be reloaded using `from_pretrained()`
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(model, 'module') else model
model_to_save = model.module if hasattr(model, "module") else model
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'))
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(
args.output_dir, force_download=True)
tokenizer = tokenizer_class.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
model = model_class.from_pretrained(args.output_dir, force_download=True)
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
model.to(args.device)
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
......@@ -682,7 +784,10 @@ def main():
logger.info("Loading checkpoints saved during training for evaluation")
checkpoints = [args.output_dir]
if args.eval_all_checkpoints:
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
checkpoints = list(
os.path.dirname(c)
for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
else:
logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
......@@ -692,17 +797,14 @@ def main():
for checkpoint in checkpoints:
# Reload the model
global_step = checkpoint.split(
'-')[-1] if len(checkpoints) > 1 else ""
model = model_class.from_pretrained(
checkpoint, force_download=True)
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
model = model_class.from_pretrained(checkpoint, force_download=True)
model.to(args.device)
# Evaluate
result = evaluate(args, model, tokenizer, prefix=global_step)
result = dict((k + ('_{}'.format(global_step) if global_step else ''), v)
for k, v in result.items())
result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
results.update(result)
logger.info("Results: {}".format(results))
......
examples/run_tf_glue.py
View file @ fa84ae26
import os
import tensorflow as tf
import tensorflow_datasets
from transformers import BertTokenizer, TFBertForSequenceClassification, BertConfig, glue_convert_examples_to_features, BertForSequenceClassification, glue_processors
from transformers import (
BertTokenizer,
TFBertForSequenceClassification,
BertConfig,
glue_convert_examples_to_features,
BertForSequenceClassification,
glue_processors,
)
# script parameters
BATCH_SIZE = 32
......@@ -27,21 +34,21 @@ 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)
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
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
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)
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)
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)
......@@ -49,7 +56,7 @@ valid_dataset = valid_dataset.batch(EVAL_BATCH_SIZE)
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')
opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, "dynamic")
if num_labels == 1:
......@@ -57,37 +64,42 @@ if num_labels == 1:
else:
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
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
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)
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/')
os.makedirs("./save/", exist_ok=True)
model.save_pretrained("./save/")
if TASK == "mrpc":
# Load the TensorFlow model in PyTorch for inspection
# This is to demo the interoperability between the two frameworks, you don't have to
# do this in real life (you can run the inference on the TF model).
pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
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')
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')
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/run_tf_ner.py
View file @ fa84ae26
......@@ -21,189 +21,156 @@ from absl import app
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
())
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)), ()
)
MODEL_CLASSES = {
"bert": (BertConfig, TFBertForTokenClassification, BertTokenizer),
"roberta": (RobertaConfig, TFRobertaForTokenClassification, RobertaTokenizer),
"distilbert": (DistilBertConfig, TFDistilBertForTokenClassification, DistilBertTokenizer)
"distilbert": (DistilBertConfig, TFDistilBertForTokenClassification, DistilBertTokenizer),
}
flags.DEFINE_string(
"data_dir", None,
"The input data dir. Should contain the .conll files (or other data files) "
"for the task.")
"data_dir", None, "The input data dir. Should contain the .conll files (or other data files) " "for the task."
)
flags.DEFINE_string(
"model_type", None,
"Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
flags.DEFINE_string("model_type", None, "Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
flags.DEFINE_string(
"model_name_or_path", None,
"Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
"model_name_or_path",
None,
"Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
)
flags.DEFINE_string(
"output_dir", None,
"The output directory where the model checkpoints will be written.")
flags.DEFINE_string("output_dir", None, "The output directory where the model checkpoints will be written.")
flags.DEFINE_string(
"labels", "",
"Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
"labels", "", "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used."
)
flags.DEFINE_string(
"config_name", "",
"Pretrained config name or path if not the same as model_name")
flags.DEFINE_string("config_name", "", "Pretrained config name or path if not the same as model_name")
flags.DEFINE_string(
"tokenizer_name", "",
"Pretrained tokenizer name or path if not the same as model_name")
flags.DEFINE_string("tokenizer_name", "", "Pretrained tokenizer name or path if not the same as model_name")
flags.DEFINE_string(
"cache_dir", "",
"Where do you want to store the pre-trained models downloaded from s3")
flags.DEFINE_string("cache_dir", "", "Where do you want to store the pre-trained models downloaded from s3")
flags.DEFINE_integer(
"max_seq_length", 128,
"max_seq_length",
128,
"The maximum total input sentence length after tokenization. "
"Sequences longer than this will be truncated, sequences shorter "
"will be padded.")
"will be padded.",
)
flags.DEFINE_string(
"tpu", None,
"tpu",
None,
"The Cloud TPU to use for training. This should be either the name "
"used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 "
"url.")
"url.",
)
flags.DEFINE_integer(
"num_tpu_cores", 8,
"Total number of TPU cores to use.")
flags.DEFINE_integer("num_tpu_cores", 8, "Total number of TPU cores to use.")
flags.DEFINE_boolean(
"do_train", False,
"Whether to run training.")
flags.DEFINE_boolean("do_train", False, "Whether to run training.")
flags.DEFINE_boolean(
"do_eval", False,
"Whether to run eval on the dev set.")
flags.DEFINE_boolean("do_eval", False, "Whether to run eval on the dev set.")
flags.DEFINE_boolean(
"do_predict", False,
"Whether to run predictions on the test set.")
flags.DEFINE_boolean("do_predict", False, "Whether to run predictions on the test set.")
flags.DEFINE_boolean(
"evaluate_during_training", False,
"Whether to run evaluation during training at each logging step.")
"evaluate_during_training", False, "Whether to run evaluation during training at each logging step."
)
flags.DEFINE_boolean(
"do_lower_case", False,
"Set this flag if you are using an uncased model.")
flags.DEFINE_boolean("do_lower_case", False, "Set this flag if you are using an uncased model.")
flags.DEFINE_integer(
"per_device_train_batch_size", 8,
"Batch size per GPU/CPU/TPU for training.")
flags.DEFINE_integer("per_device_train_batch_size", 8, "Batch size per GPU/CPU/TPU for training.")
flags.DEFINE_integer(
"per_device_eval_batch_size", 8,
"Batch size per GPU/CPU/TPU for evaluation.")
flags.DEFINE_integer("per_device_eval_batch_size", 8, "Batch size per GPU/CPU/TPU for evaluation.")
flags.DEFINE_integer(
"gradient_accumulation_steps", 1,
"Number of updates steps to accumulate before performing a backward/update pass.")
"gradient_accumulation_steps", 1, "Number of updates steps to accumulate before performing a backward/update pass."
)
flags.DEFINE_float(
"learning_rate", 5e-5,
"The initial learning rate for Adam.")
flags.DEFINE_float("learning_rate", 5e-5, "The initial learning rate for Adam.")
flags.DEFINE_float(
"weight_decay", 0.0,
"Weight decay if we apply some.")
flags.DEFINE_float("weight_decay", 0.0, "Weight decay if we apply some.")
flags.DEFINE_float(
"adam_epsilon", 1e-8,
"Epsilon for Adam optimizer.")
flags.DEFINE_float("adam_epsilon", 1e-8, "Epsilon for Adam optimizer.")
flags.DEFINE_float(
"max_grad_norm", 1.0,
"Max gradient norm.")
flags.DEFINE_float("max_grad_norm", 1.0, "Max gradient norm.")
flags.DEFINE_integer(
"num_train_epochs", 3,
"Total number of training epochs to perform.")
flags.DEFINE_integer("num_train_epochs", 3, "Total number of training epochs to perform.")
flags.DEFINE_integer(
"max_steps", -1,
"If > 0: set total number of training steps to perform. Override num_train_epochs.")
"max_steps", -1, "If > 0: set total number of training steps to perform. Override num_train_epochs."
)
flags.DEFINE_integer(
"warmup_steps", 0,
"Linear warmup over warmup_steps.")
flags.DEFINE_integer("warmup_steps", 0, "Linear warmup over warmup_steps.")
flags.DEFINE_integer(
"logging_steps", 50,
"Log every X updates steps.")
flags.DEFINE_integer("logging_steps", 50, "Log every X updates steps.")
flags.DEFINE_integer(
"save_steps", 50,
"Save checkpoint every X updates steps.")
flags.DEFINE_integer("save_steps", 50, "Save checkpoint every X updates steps.")
flags.DEFINE_boolean(
"eval_all_checkpoints", False,
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
"eval_all_checkpoints",
False,
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
)
flags.DEFINE_boolean(
"no_cuda", False,
"Avoid using CUDA when available")
flags.DEFINE_boolean("no_cuda", False, "Avoid using CUDA when available")
flags.DEFINE_boolean(
"overwrite_output_dir", False,
"Overwrite the content of the output directory")
flags.DEFINE_boolean("overwrite_output_dir", False, "Overwrite the content of the output directory")
flags.DEFINE_boolean(
"overwrite_cache", False,
"Overwrite the cached training and evaluation sets")
flags.DEFINE_boolean("overwrite_cache", False, "Overwrite the cached training and evaluation sets")
flags.DEFINE_integer(
"seed", 42,
"random seed for initialization")
flags.DEFINE_integer("seed", 42, "random seed for initialization")
flags.DEFINE_boolean(
"fp16", False,
"Whether to use 16-bit (mixed) precision instead of 32-bit")
flags.DEFINE_boolean("fp16", False, "Whether to use 16-bit (mixed) precision instead of 32-bit")
flags.DEFINE_string(
"gpus", "0",
"gpus",
"0",
"Comma separated list of gpus devices. If only one, switch to single "
"gpu strategy, if None takes all the gpus available.")
"gpu strategy, if None takes all the gpus available.",
)
def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id):
if args['max_steps'] > 0:
num_train_steps = args['max_steps'] * args['gradient_accumulation_steps']
args['num_train_epochs'] = 1
def train(
args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id
):
if args["max_steps"] > 0:
num_train_steps = args["max_steps"] * args["gradient_accumulation_steps"]
args["num_train_epochs"] = 1
else:
num_train_steps = math.ceil(num_train_examples / train_batch_size) // args['gradient_accumulation_steps'] * args['num_train_epochs']
num_train_steps = (
math.ceil(num_train_examples / train_batch_size)
// args["gradient_accumulation_steps"]
* args["num_train_epochs"]
)
writer = tf.summary.create_file_writer("/tmp/mylogs")
with strategy.scope():
loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
optimizer = create_optimizer(args['learning_rate'], num_train_steps, args['warmup_steps'])
optimizer = create_optimizer(args["learning_rate"], num_train_steps, args["warmup_steps"])
if args['fp16']:
optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(optimizer, 'dynamic')
if args["fp16"]:
optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(optimizer, "dynamic")
loss_metric = tf.keras.metrics.Mean(name='loss', dtype=tf.float32)
loss_metric = tf.keras.metrics.Mean(name="loss", dtype=tf.float32)
gradient_accumulator = GradientAccumulator()
logging.info("***** Running training *****")
logging.info(" Num examples = %d", num_train_examples)
logging.info(" Num Epochs = %d", args['num_train_epochs'])
logging.info(" Instantaneous batch size per device = %d", args['per_device_train_batch_size'])
logging.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
train_batch_size * args['gradient_accumulation_steps'])
logging.info(" Gradient Accumulation steps = %d", args['gradient_accumulation_steps'])
logging.info(" Num Epochs = %d", args["num_train_epochs"])
logging.info(" Instantaneous batch size per device = %d", args["per_device_train_batch_size"])
logging.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
train_batch_size * args["gradient_accumulation_steps"],
)
logging.info(" Gradient Accumulation steps = %d", args["gradient_accumulation_steps"])
logging.info(" Total training steps = %d", num_train_steps)
model.summary()
......@@ -214,26 +181,28 @@ def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, l
for gradient, variable in zip(gradient_accumulator.gradients, model.trainable_variables):
if gradient is not None:
scaled_gradient = gradient / (args['n_device'] * args['gradient_accumulation_steps'])
scaled_gradient = gradient / (args["n_device"] * args["gradient_accumulation_steps"])
grads_and_vars.append((scaled_gradient, variable))
else:
grads_and_vars.append((gradient, variable))
optimizer.apply_gradients(grads_and_vars, args['max_grad_norm'])
optimizer.apply_gradients(grads_and_vars, args["max_grad_norm"])
gradient_accumulator.reset()
@tf.function
def train_step(train_features, train_labels):
def step_fn(train_features, train_labels):
inputs = {'attention_mask': train_features['input_mask'], 'training': True}
inputs = {"attention_mask": train_features["input_mask"], "training": True}
if args['model_type'] != "distilbert":
inputs["token_type_ids"] = train_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
if args["model_type"] != "distilbert":
inputs["token_type_ids"] = (
train_features["segment_ids"] if args["model_type"] in ["bert", "xlnet"] else None
)
with tf.GradientTape() as tape:
logits = model(train_features['input_ids'], **inputs)[0]
logits = model(train_features["input_ids"], **inputs)[0]
logits = tf.reshape(logits, (-1, len(labels) + 1))
active_loss = tf.reshape(train_features['input_mask'], (-1,))
active_loss = tf.reshape(train_features["input_mask"], (-1,))
active_logits = tf.boolean_mask(logits, active_loss)
train_labels = tf.reshape(train_labels, (-1,))
active_labels = tf.boolean_mask(train_labels, active_loss)
......@@ -251,29 +220,35 @@ def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, l
return mean_loss
current_time = datetime.datetime.now()
train_iterator = master_bar(range(args['num_train_epochs']))
train_iterator = master_bar(range(args["num_train_epochs"]))
global_step = 0
logging_loss = 0.0
for epoch in train_iterator:
epoch_iterator = progress_bar(train_dataset, total=num_train_steps, parent=train_iterator, display=args['n_device'] > 1)
epoch_iterator = progress_bar(
train_dataset, total=num_train_steps, parent=train_iterator, display=args["n_device"] > 1
)
step = 1
with strategy.scope():
for train_features, train_labels in epoch_iterator:
loss = train_step(train_features, train_labels)
if step % args['gradient_accumulation_steps'] == 0:
if step % args["gradient_accumulation_steps"] == 0:
strategy.experimental_run_v2(apply_gradients)
loss_metric(loss)
global_step += 1
if args['logging_steps'] > 0 and global_step % args['logging_steps'] == 0:
if args["logging_steps"] > 0 and global_step % args["logging_steps"] == 0:
# Log metrics
if args['n_device'] == 1 and args['evaluate_during_training']: # Only evaluate when single GPU otherwise metrics may not average well
y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
if (
args["n_device"] == 1 and args["evaluate_during_training"]
): # Only evaluate when single GPU otherwise metrics may not average well
y_true, y_pred, eval_loss = evaluate(
args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev"
)
report = metrics.classification_report(y_true, y_pred, digits=4)
logging.info("Eval at step " + str(global_step) + "\n" + report)
......@@ -294,16 +269,18 @@ def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, l
with writer.as_default():
tf.summary.scalar("lr", learning_rate, global_step)
tf.summary.scalar("loss", (loss_metric.result() - logging_loss) / args['logging_steps'], global_step)
tf.summary.scalar(
"loss", (loss_metric.result() - logging_loss) / args["logging_steps"], global_step
)
logging_loss = loss_metric.result()
with writer.as_default():
tf.summary.scalar("loss", loss_metric.result(), step=step)
if args['save_steps'] > 0 and global_step % args['save_steps'] == 0:
if 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))
output_dir = os.path.join(args["output_dir"], "checkpoint-{}".format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
......@@ -311,10 +288,10 @@ def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, l
model.save_pretrained(output_dir)
logging.info("Saving model checkpoint to %s", output_dir)
train_iterator.child.comment = f'loss : {loss_metric.result()}'
train_iterator.child.comment = f"loss : {loss_metric.result()}"
step += 1
train_iterator.write(f'loss epoch {epoch + 1}: {loss_metric.result()}')
train_iterator.write(f"loss epoch {epoch + 1}: {loss_metric.result()}")
loss_metric.reset_states()
......@@ -322,13 +299,15 @@ def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, l
def evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode):
eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
eval_dataset, size = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode=mode)
eval_batch_size = args["per_device_eval_batch_size"] * args["n_device"]
eval_dataset, size = load_and_cache_examples(
args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode=mode
)
eval_dataset = strategy.experimental_distribute_dataset(eval_dataset)
preds = None
num_eval_steps = math.ceil(size / eval_batch_size)
master = master_bar(range(1))
eval_iterator = progress_bar(eval_dataset, total=num_eval_steps, parent=master, display=args['n_device'] > 1)
eval_iterator = progress_bar(eval_dataset, total=num_eval_steps, parent=master, display=args["n_device"] > 1)
loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
loss = 0.0
......@@ -337,15 +316,17 @@ def evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode)
logging.info(" Batch size = %d", eval_batch_size)
for eval_features, eval_labels in eval_iterator:
inputs = {'attention_mask': eval_features['input_mask'], 'training': False}
inputs = {"attention_mask": eval_features["input_mask"], "training": False}
if args['model_type'] != "distilbert":
inputs["token_type_ids"] = eval_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
if args["model_type"] != "distilbert":
inputs["token_type_ids"] = (
eval_features["segment_ids"] if args["model_type"] in ["bert", "xlnet"] else None
)
with strategy.scope():
logits = model(eval_features['input_ids'], **inputs)[0]
logits = model(eval_features["input_ids"], **inputs)[0]
tmp_logits = tf.reshape(logits, (-1, len(labels) + 1))
active_loss = tf.reshape(eval_features['input_mask'], (-1,))
active_loss = tf.reshape(eval_features["input_mask"], (-1,))
active_logits = tf.boolean_mask(tmp_logits, active_loss)
tmp_eval_labels = tf.reshape(eval_labels, (-1,))
active_labels = tf.boolean_mask(tmp_eval_labels, active_loss)
......@@ -384,11 +365,11 @@ def load_cache(cached_file, max_seq_length):
def _decode_record(record):
example = tf.io.parse_single_example(record, name_to_features)
features = {}
features['input_ids'] = example['input_ids']
features['input_mask'] = example['input_mask']
features['segment_ids'] = example['segment_ids']
features["input_ids"] = example["input_ids"]
features["input_mask"] = example["input_mask"]
features["segment_ids"] = example["segment_ids"]
return features, example['label_ids']
return features, example["label_ids"]
d = tf.data.TFRecordDataset(cached_file)
d = d.map(_decode_record, num_parallel_calls=4)
......@@ -422,39 +403,46 @@ def save_cache(features, cached_features_file):
def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, batch_size, mode):
drop_remainder = True if args['tpu'] or mode == 'train' else False
drop_remainder = True if args["tpu"] or mode == "train" else False
# Load data features from cache or dataset file
cached_features_file = os.path.join(args['data_dir'], "cached_{}_{}_{}.tf_record".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']:
cached_features_file = os.path.join(
args["data_dir"],
"cached_{}_{}_{}.tf_record".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"]:
logging.info("Loading features from cached file %s", cached_features_file)
dataset, size = load_cache(cached_features_file, args['max_seq_length'])
dataset, size = load_cache(cached_features_file, args["max_seq_length"])
else:
logging.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"]),
logging.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,
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"]),
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_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
pad_token_segment_id=4 if args["model_type"] in ["xlnet"] else 0,
pad_token_label_id=pad_token_label_id,
)
logging.info("Saving features into cached file %s", cached_features_file)
save_cache(features, cached_features_file)
dataset, size = load_cache(cached_features_file, args['max_seq_length'])
dataset, size = load_cache(cached_features_file, args["max_seq_length"])
if mode == 'train':
if mode == "train":
dataset = dataset.repeat()
dataset = dataset.shuffle(buffer_size=8192, seed=args['seed'])
dataset = dataset.shuffle(buffer_size=8192, seed=args["seed"])
dataset = dataset.batch(batch_size, drop_remainder)
dataset = dataset.prefetch(buffer_size=batch_size)
......@@ -466,83 +454,117 @@ def main(_):
logging.set_verbosity(logging.INFO)
args = flags.FLAGS.flag_values_dict()
if os.path.exists(args['output_dir']) and os.listdir(
args['output_dir']) and args['do_train'] and not args['overwrite_output_dir']:
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']))
args["output_dir"]
)
)
if args['fp16']:
if args["fp16"]:
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
if args['tpu']:
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=args['tpu'])
if args["tpu"]:
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=args["tpu"])
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.experimental.TPUStrategy(resolver)
args['n_device'] = args['num_tpu_cores']
elif len(args['gpus'].split(',')) > 1:
args['n_device'] = len([f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
strategy = tf.distribute.MirroredStrategy(devices=[f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
elif args['no_cuda']:
args['n_device'] = 1
args["n_device"] = args["num_tpu_cores"]
elif len(args["gpus"].split(",")) > 1:
args["n_device"] = len([f"/gpu:{gpu}" for gpu in args["gpus"].split(",")])
strategy = tf.distribute.MirroredStrategy(devices=[f"/gpu:{gpu}" for gpu in args["gpus"].split(",")])
elif args["no_cuda"]:
args["n_device"] = 1
strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
else:
args['n_device'] = len(args['gpus'].split(','))
strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" + args['gpus'].split(',')[0])
logging.warning("n_device: %s, distributed training: %s, 16-bits training: %s",
args['n_device'], bool(args['n_device'] > 1), args['fp16'])
args["n_device"] = len(args["gpus"].split(","))
strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" + args["gpus"].split(",")[0])
logging.warning(
"n_device: %s, distributed training: %s, 16-bits training: %s",
args["n_device"],
bool(args["n_device"] > 1),
args["fp16"],
)
labels = get_labels(args['labels'])
labels = get_labels(args["labels"])
num_labels = len(labels) + 1
pad_token_label_id = 0
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'],
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,
cache_dir=args['cache_dir'] if args['cache_dir'] else None)
cache_dir=args["cache_dir"] if args["cache_dir"] else None,
)
logging.info("Training/evaluation parameters %s", args)
# Training
if args['do_train']:
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'],
cache_dir=args['cache_dir'] if args['cache_dir'] else None)
if args["do_train"]:
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"],
cache_dir=args["cache_dir"] if args["cache_dir"] else None,
)
with strategy.scope():
model = model_class.from_pretrained(args['model_name_or_path'],
from_pt=bool(".bin" in args['model_name_or_path']),
model = model_class.from_pretrained(
args["model_name_or_path"],
from_pt=bool(".bin" in args["model_name_or_path"]),
config=config,
cache_dir=args['cache_dir'] if args['cache_dir'] else None)
cache_dir=args["cache_dir"] if args["cache_dir"] else None,
)
model.layers[-1].activation = tf.keras.activations.softmax
train_batch_size = args['per_device_train_batch_size'] * args['n_device']
train_dataset, num_train_examples = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, train_batch_size, mode="train")
train_batch_size = args["per_device_train_batch_size"] * args["n_device"]
train_dataset, num_train_examples = load_and_cache_examples(
args, tokenizer, labels, pad_token_label_id, train_batch_size, mode="train"
)
train_dataset = strategy.experimental_distribute_dataset(train_dataset)
train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id)
train(
args,
strategy,
train_dataset,
tokenizer,
model,
num_train_examples,
labels,
train_batch_size,
pad_token_label_id,
)
if not os.path.exists(args['output_dir']):
os.makedirs(args['output_dir'])
if not os.path.exists(args["output_dir"]):
os.makedirs(args["output_dir"])
logging.info("Saving model to %s", args['output_dir'])
logging.info("Saving model to %s", args["output_dir"])
model.save_pretrained(args['output_dir'])
tokenizer.save_pretrained(args['output_dir'])
model.save_pretrained(args["output_dir"])
tokenizer.save_pretrained(args["output_dir"])
# Evaluation
if args['do_eval']:
tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
if args["do_eval"]:
tokenizer = tokenizer_class.from_pretrained(args["output_dir"], do_lower_case=args["do_lower_case"])
checkpoints = []
results = []
if args['eval_all_checkpoints']:
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args['output_dir'] + "/**/" + TF2_WEIGHTS_NAME, recursive=True), key=lambda f: int(''.join(filter(str.isdigit, f)) or -1)))
if args["eval_all_checkpoints"]:
checkpoints = list(
os.path.dirname(c)
for c in sorted(
glob.glob(args["output_dir"] + "/**/" + TF2_WEIGHTS_NAME, recursive=True),
key=lambda f: int("".join(filter(str.isdigit, f)) or -1),
)
)
logging.info("Evaluate the following checkpoints: %s", checkpoints)
if len(checkpoints) == 0:
checkpoints.append(args['output_dir'])
checkpoints.append(args["output_dir"])
for checkpoint in checkpoints:
global_step = checkpoint.split("-")[-1] if re.match(".*checkpoint-[0-9]", checkpoint) else "final"
......@@ -550,13 +572,15 @@ def main(_):
with strategy.scope():
model = model_class.from_pretrained(checkpoint)
y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
y_true, y_pred, eval_loss = evaluate(
args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev"
)
report = metrics.classification_report(y_true, y_pred, digits=4)
if global_step:
results.append({global_step + "_report": report, global_step + "_loss": eval_loss})
output_eval_file = os.path.join(args['output_dir'], "eval_results.txt")
output_eval_file = os.path.join(args["output_dir"], "eval_results.txt")
with tf.io.gfile.GFile(output_eval_file, "w") as writer:
for res in results:
......@@ -572,14 +596,16 @@ def main(_):
writer.write(report)
writer.write("\n")
if args['do_predict']:
tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
model = model_class.from_pretrained(args['output_dir'])
eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
predict_dataset, _ = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode="test")
if args["do_predict"]:
tokenizer = tokenizer_class.from_pretrained(args["output_dir"], do_lower_case=args["do_lower_case"])
model = model_class.from_pretrained(args["output_dir"])
eval_batch_size = args["per_device_eval_batch_size"] * args["n_device"]
predict_dataset, _ = load_and_cache_examples(
args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode="test"
)
y_true, y_pred, pred_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="test")
output_test_results_file = os.path.join(args['output_dir'], "test_results.txt")
output_test_predictions_file = os.path.join(args['output_dir'], "test_predictions.txt")
output_test_results_file = os.path.join(args["output_dir"], "test_results.txt")
output_test_predictions_file = os.path.join(args["output_dir"], "test_predictions.txt")
report = metrics.classification_report(y_true, y_pred, digits=4)
with tf.io.gfile.GFile(output_test_results_file, "w") as writer:
......@@ -591,7 +617,7 @@ def main(_):
writer.write("\n\nloss = " + str(pred_loss))
with tf.io.gfile.GFile(output_test_predictions_file, "w") as writer:
with tf.io.gfile.GFile(os.path.join(args['data_dir'], "test.txt"), "r") as f:
with tf.io.gfile.GFile(os.path.join(args["data_dir"], "test.txt"), "r") as f:
example_id = 0
for line in f:
......
examples/run_xnli.py
View file @ fa84ae26
......@@ -26,8 +26,7 @@ import random
import numpy as np
import torch
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
TensorDataset)
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
from torch.utils.data.distributed import DistributedSampler
try:
......@@ -37,10 +36,18 @@ except:
from tqdm import tqdm, trange
from transformers import (WEIGHTS_NAME,
BertConfig, BertForSequenceClassification, BertTokenizer,
XLMConfig, XLMForSequenceClassification, XLMTokenizer,
DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
from transformers import (
WEIGHTS_NAME,
BertConfig,
BertForSequenceClassification,
BertTokenizer,
XLMConfig,
XLMForSequenceClassification,
XLMTokenizer,
DistilBertConfig,
DistilBertForSequenceClassification,
DistilBertTokenizer,
)
from transformers import AdamW, get_linear_schedule_with_warmup
......@@ -52,12 +59,14 @@ from transformers import glue_convert_examples_to_features as convert_examples_t
logger = logging.getLogger(__name__)
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, DistilBertConfig, XLMConfig)), ())
ALL_MODELS = sum(
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, DistilBertConfig, XLMConfig)), ()
)
MODEL_CLASSES = {
'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
"bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
"xlm": (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
"distilbert": (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
}
......@@ -85,19 +94,26 @@ def train(args, train_dataset, model, tokenizer):
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']
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}
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0},
]
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
)
# Check if saved optimizer or scheduler states exist
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
os.path.join(args.model_name_or_path, "scheduler.pt")
):
# Load in optimizer and scheduler states
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
if args.fp16:
try:
......@@ -112,17 +128,21 @@ def train(args, train_dataset, model, tokenizer):
# 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)
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(
" 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)
......@@ -132,7 +152,7 @@ def train(args, train_dataset, model, tokenizer):
# Check if continuing training from a checkpoint
if os.path.exists(args.model_name_or_path):
# set global_step to gobal_step of last saved checkpoint from model path
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
......@@ -143,7 +163,9 @@ def train(args, train_dataset, model, tokenizer):
tr_loss, logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
train_iterator = trange(
epochs_trained, 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])
......@@ -155,11 +177,11 @@ def train(args, train_dataset, model, tokenizer):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {'input_ids': batch[0],
'attention_mask': batch[1],
'labels': batch[3]}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert'] else None # XLM and DistilBERT don't use segment_ids
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
if args.model_type != "distilbert":
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert"] else None
) # XLM and DistilBERT don't use segment_ids
outputs = model(**inputs)
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
......@@ -188,28 +210,32 @@ def train(args, train_dataset, model, tokenizer):
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
if (
args.local_rank == -1 and args.evaluate_during_training
): # Only evaluate when single GPU otherwise metrics may not average well
results = evaluate(args, model, tokenizer)
for key, value in results.items():
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)
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))
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 = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
torch.save(args, os.path.join(output_dir, "training_args.bin"))
logger.info("Saving model checkpoint to %s", output_dir)
torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
logger.info("Saving optimizer and scheduler states to %s", output_dir)
if args.max_steps > 0 and global_step > args.max_steps:
......@@ -258,11 +284,11 @@ 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],
'labels': batch[3]}
if args.model_type != 'distilbert':
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert'] else None # XLM and DistilBERT don't use segment_ids
inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
if args.model_type != "distilbert":
inputs["token_type_ids"] = (
batch[2] if args.model_type in ["bert"] else None
) # XLM and DistilBERT don't use segment_ids
outputs = model(**inputs)
tmp_eval_loss, logits = outputs[:2]
......@@ -270,16 +296,16 @@ def evaluate(args, model, tokenizer, prefix=""):
nb_eval_steps += 1
if preds is None:
preds = logits.detach().cpu().numpy()
out_label_ids = inputs['labels'].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)
out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
eval_loss = eval_loss / nb_eval_steps
if args.output_mode == "classification":
preds = np.argmax(preds, axis=1)
else:
raise ValueError('No other `output_mode` for XNLI.')
raise ValueError("No other `output_mode` for XNLI.")
result = compute_metrics(eval_task, preds, out_label_ids)
results.update(result)
......@@ -300,20 +326,27 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
processor = processors[task](language=args.language, train_language=args.train_language)
output_mode = output_modes[task]
# Load data features from cache or dataset file
cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}_{}'.format(
'test' if evaluate else 'train',
list(filter(None, args.model_name_or_path.split('/'))).pop(),
cached_features_file = os.path.join(
args.data_dir,
"cached_{}_{}_{}_{}_{}".format(
"test" if evaluate else "train",
list(filter(None, args.model_name_or_path.split("/"))).pop(),
str(args.max_seq_length),
str(task),
str(args.train_language if (not evaluate and args.train_language is not None) else args.language)))
str(args.train_language if (not evaluate and args.train_language is not None) else args.language),
),
)
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)
label_list = processor.get_labels()
examples = processor.get_test_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
features = convert_examples_to_features(examples,
examples = (
processor.get_test_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
)
features = convert_examples_to_features(
examples,
tokenizer,
label_list=label_list,
max_length=args.max_seq_length,
......@@ -336,7 +369,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
if output_mode == "classification":
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
else:
raise ValueError('No other `output_mode` for XNLI.')
raise ValueError("No other `output_mode` for XNLI.")
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
return dataset
......@@ -346,92 +379,152 @@ 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 .tsv files (or other data files) for the 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("--language", default=None, type=str, required=True,
help="Evaluation language. Also train language if `train_language` is set to None.")
parser.add_argument("--train_language", default=None, type=str,
help="Train language if is different of the evaluation language.")
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model predictions and checkpoints will be written.")
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
)
parser.add_argument(
"--model_type",
default=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(
"--language",
default=None,
type=str,
required=True,
help="Evaluation language. Also train language if `train_language` is set to None.",
)
parser.add_argument(
"--train_language", default=None, type=str, help="Train language if is different of the evaluation language."
)
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("--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,
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 test set.")
parser.add_argument("--evaluate_during_training", action='store_true',
help="Rul evaluation during training at each logging step.")
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--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 deay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float,
help="Max gradient norm.")
parser.add_argument("--num_train_epochs", default=3.0, 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',
"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 test set.")
parser.add_argument(
"--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
)
parser.add_argument(
"--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
)
parser.add_argument("--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 deay if we apply some.")
parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
parser.add_argument(
"--num_train_epochs", default=3.0, 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.")
"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))
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()
......@@ -443,22 +536,30 @@ def main():
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')
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)
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 XNLI task
args.task_name = 'xnli'
args.task_name = "xnli"
if args.task_name not in processors:
raise ValueError("Task not found: %s" % (args.task_name))
processor = processors[args.task_name](language=args.language, train_language=args.train_language)
......@@ -472,17 +573,23 @@ def main():
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,
config = config_class.from_pretrained(
args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_labels,
finetuning_task=args.task_name,
cache_dir=args.cache_dir if args.cache_dir else None)
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
cache_dir=args.cache_dir if args.cache_dir else None,
)
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,
cache_dir=args.cache_dir if args.cache_dir else None)
model = model_class.from_pretrained(args.model_name_or_path,
from_tf=bool('.ckpt' in args.model_name_or_path),
cache_dir=args.cache_dir if args.cache_dir else None,
)
model = model_class.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
cache_dir=args.cache_dir if args.cache_dir else None)
cache_dir=args.cache_dir if args.cache_dir else None,
)
if args.local_rank == 0:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
......@@ -491,14 +598,12 @@ def main():
logger.info("Training/evaluation parameters %s", args)
# Training
if args.do_train:
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
# Saving best-practices: if you use 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
......@@ -508,36 +613,39 @@ def main():
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 = (
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'))
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
# Load a trained model and vocabulary that you have fine-tuned
model = model_class.from_pretrained(args.output_dir)
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
model.to(args.device)
# 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)))
checkpoints = list(
os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
)
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
logger.info("Evaluate the following checkpoints: %s", checkpoints)
for checkpoint in checkpoints:
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
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=prefix)
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
results.update(result)
return results
......
examples/summarization/convert_bertabs_original_pytorch_checkpoint.py
View file @ fa84ae26
......@@ -34,12 +34,30 @@ logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
SAMPLE_TEXT = 'Hello world! cécé herlolip'
SAMPLE_TEXT = "Hello world! cécé herlolip"
BertAbsConfig = namedtuple(
"BertAbsConfig",
["temp_dir", "large", "use_bert_emb", "finetune_bert", "encoder", "share_emb", "max_pos", "enc_layers", "enc_hidden_size", "enc_heads", "enc_ff_size", "enc_dropout", "dec_layers", "dec_hidden_size", "dec_heads", "dec_ff_size", "dec_dropout"],
[
"temp_dir",
"large",
"use_bert_emb",
"finetune_bert",
"encoder",
"share_emb",
"max_pos",
"enc_layers",
"enc_hidden_size",
"enc_heads",
"enc_ff_size",
"enc_dropout",
"dec_layers",
"dec_hidden_size",
"dec_heads",
"dec_ff_size",
"dec_dropout",
],
)
......@@ -119,7 +137,9 @@ def convert_bertabs_checkpoints(path_to_checkpoints, dump_path):
output_original_model = original(src, tgt, segs, clss, mask_src, mask_tgt, mask_cls)[0]
output_original_generator = original.generator(output_original_model)
output_converted_model = new_model(encoder_input_ids, decoder_input_ids, token_type_ids, encoder_attention_mask, decoder_attention_mask)[0]
output_converted_model = new_model(
encoder_input_ids, decoder_input_ids, token_type_ids, encoder_attention_mask, decoder_attention_mask
)[0]
output_converted_generator = new_model.generator(output_converted_model)
maximum_absolute_difference = torch.max(torch.abs(output_converted_model - output_original_model)).item()
......@@ -136,28 +156,21 @@ def convert_bertabs_checkpoints(path_to_checkpoints, dump_path):
# The model has been saved with torch.save(model) and this is bound to the exact
# directory structure. We save the state_dict instead.
logging.info("saving the model's state dictionary")
torch.save(new_model.state_dict(), "bertabs-finetuned-cnndm-extractive-abstractive-summarization-pytorch_model.bin")
torch.save(
new_model.state_dict(), "bertabs-finetuned-cnndm-extractive-abstractive-summarization-pytorch_model.bin"
)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--bertabs_checkpoint_path",
default=None,
type=str,
required=True,
help="Path the official PyTorch dump.",
"--bertabs_checkpoint_path", default=None, type=str, required=True, help="Path the official PyTorch dump.",
)
parser.add_argument(
"--pytorch_dump_folder_path",
default=None,
type=str,
required=True,
help="Path to the output PyTorch model.",
"--pytorch_dump_folder_path", default=None, type=str, required=True, help="Path to the output PyTorch model.",
)
args = parser.parse_args()
convert_bertabs_checkpoints(
args.bertabs_checkpoint_path,
args.pytorch_dump_folder_path,
args.bertabs_checkpoint_path, args.pytorch_dump_folder_path,
)
examples/summarization/modeling_bertabs.py
View file @ fa84ae26
......@@ -56,40 +56,22 @@ class BertAbs(BertAbsPreTrainedModel):
load_bert_pretrained_extractive = True if bert_extractive_checkpoint else False
if load_bert_pretrained_extractive:
self.bert.model.load_state_dict(
dict(
[
(n[11:], p)
for n, p in bert_extractive_checkpoint.items()
if n.startswith("bert.model")
]
),
dict([(n[11:], p) for n, p in bert_extractive_checkpoint.items() if n.startswith("bert.model")]),
strict=True,
)
self.vocab_size = self.bert.model.config.vocab_size
if args.max_pos > 512:
my_pos_embeddings = nn.Embedding(
args.max_pos, self.bert.model.config.hidden_size
)
my_pos_embeddings.weight.data[
:512
] = self.bert.model.embeddings.position_embeddings.weight.data
my_pos_embeddings.weight.data[
512:
] = self.bert.model.embeddings.position_embeddings.weight.data[-1][
my_pos_embeddings = nn.Embedding(args.max_pos, self.bert.model.config.hidden_size)
my_pos_embeddings.weight.data[:512] = self.bert.model.embeddings.position_embeddings.weight.data
my_pos_embeddings.weight.data[512:] = self.bert.model.embeddings.position_embeddings.weight.data[-1][
None, :
].repeat(
args.max_pos - 512, 1
)
].repeat(args.max_pos - 512, 1)
self.bert.model.embeddings.position_embeddings = my_pos_embeddings
tgt_embeddings = nn.Embedding(
self.vocab_size, self.bert.model.config.hidden_size, padding_idx=0
)
tgt_embeddings = nn.Embedding(self.vocab_size, self.bert.model.config.hidden_size, padding_idx=0)
tgt_embeddings.weight = copy.deepcopy(
self.bert.model.embeddings.word_embeddings.weight
)
tgt_embeddings.weight = copy.deepcopy(self.bert.model.embeddings.word_embeddings.weight)
self.decoder = TransformerDecoder(
self.args.dec_layers,
......@@ -102,9 +84,7 @@ class BertAbs(BertAbsPreTrainedModel):
)
gen_func = nn.LogSoftmax(dim=-1)
self.generator = nn.Sequential(
nn.Linear(args.dec_hidden_size, args.vocab_size), gen_func
)
self.generator = nn.Sequential(nn.Linear(args.dec_hidden_size, args.vocab_size), gen_func)
self.generator[0].weight = self.decoder.embeddings.weight
load_from_checkpoints = False if checkpoint is None else True
......@@ -127,25 +107,14 @@ class BertAbs(BertAbsPreTrainedModel):
p.data.zero_()
def forward(
self,
encoder_input_ids,
decoder_input_ids,
token_type_ids,
encoder_attention_mask,
decoder_attention_mask,
self, encoder_input_ids, decoder_input_ids, token_type_ids, encoder_attention_mask, decoder_attention_mask,
):
encoder_output = self.bert(
input_ids=encoder_input_ids,
token_type_ids=token_type_ids,
attention_mask=encoder_attention_mask,
input_ids=encoder_input_ids, token_type_ids=token_type_ids, attention_mask=encoder_attention_mask,
)
encoder_hidden_states = encoder_output[0]
dec_state = self.decoder.init_decoder_state(
encoder_input_ids, encoder_hidden_states
)
decoder_outputs, _ = self.decoder(
decoder_input_ids[:, :-1], encoder_hidden_states, dec_state
)
dec_state = self.decoder.init_decoder_state(encoder_input_ids, encoder_hidden_states)
decoder_outputs, _ = self.decoder(decoder_input_ids[:, :-1], encoder_hidden_states, dec_state)
return decoder_outputs
......@@ -162,10 +131,7 @@ class Bert(nn.Module):
self.eval()
with torch.no_grad():
encoder_outputs, _ = self.model(
input_ids,
token_type_ids=token_type_ids,
attention_mask=attention_mask,
**kwargs
input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask, **kwargs
)
return encoder_outputs
......@@ -196,10 +162,7 @@ class TransformerDecoder(nn.Module):
# Build TransformerDecoder.
self.transformer_layers = nn.ModuleList(
[
TransformerDecoderLayer(d_model, heads, d_ff, dropout)
for _ in range(num_layers)
]
[TransformerDecoderLayer(d_model, heads, d_ff, dropout) for _ in range(num_layers)]
)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
......@@ -236,20 +199,14 @@ class TransformerDecoder(nn.Module):
# Decoder padding mask
tgt_words = tgt
tgt_batch, tgt_len = tgt_words.size()
tgt_pad_mask = (
tgt_words.data.eq(padding_idx).unsqueeze(1).expand(tgt_batch, tgt_len, tgt_len)
)
tgt_pad_mask = tgt_words.data.eq(padding_idx).unsqueeze(1).expand(tgt_batch, tgt_len, tgt_len)
# Encoder padding mask
if memory_mask is not None:
src_len = memory_mask.size(-1)
src_pad_mask = memory_mask.expand(src_batch, tgt_len, src_len)
else:
src_pad_mask = (
src_words.data.eq(padding_idx)
.unsqueeze(1)
.expand(src_batch, tgt_len, src_len)
)
src_pad_mask = src_words.data.eq(padding_idx).unsqueeze(1).expand(src_batch, tgt_len, src_len)
# Pass through the embeddings
emb = self.embeddings(input_ids)
......@@ -271,9 +228,7 @@ class TransformerDecoder(nn.Module):
src_pad_mask,
tgt_pad_mask,
previous_input=prev_layer_input,
layer_cache=state.cache["layer_{}".format(i)]
if state.cache is not None
else None,
layer_cache=state.cache["layer_{}".format(i)] if state.cache is not None else None,
step=step,
)
if state.cache is None:
......@@ -303,9 +258,7 @@ class PositionalEncoding(nn.Module):
def __init__(self, dropout, dim, max_len=5000):
pe = torch.zeros(max_len, dim)
position = torch.arange(0, max_len).unsqueeze(1)
div_term = torch.exp(
(torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(10000.0) / dim))
)
div_term = torch.exp((torch.arange(0, dim, 2, dtype=torch.float) * -(math.log(10000.0) / dim)))
pe[:, 0::2] = torch.sin(position.float() * div_term)
pe[:, 1::2] = torch.cos(position.float() * div_term)
pe = pe.unsqueeze(0)
......@@ -356,14 +309,7 @@ class TransformerDecoderLayer(nn.Module):
self.register_buffer("mask", mask)
def forward(
self,
inputs,
memory_bank,
src_pad_mask,
tgt_pad_mask,
previous_input=None,
layer_cache=None,
step=None,
self, inputs, memory_bank, src_pad_mask, tgt_pad_mask, previous_input=None, layer_cache=None, step=None,
):
"""
Args:
......@@ -380,34 +326,20 @@ class TransformerDecoderLayer(nn.Module):
* all_input `[batch_size x current_step x model_dim]`
"""
dec_mask = torch.gt(
tgt_pad_mask + self.mask[:, : tgt_pad_mask.size(1), : tgt_pad_mask.size(1)], 0
)
dec_mask = torch.gt(tgt_pad_mask + self.mask[:, : tgt_pad_mask.size(1), : tgt_pad_mask.size(1)], 0)
input_norm = self.layer_norm_1(inputs)
all_input = input_norm
if previous_input is not None:
all_input = torch.cat((previous_input, input_norm), dim=1)
dec_mask = None
query = self.self_attn(
all_input,
all_input,
input_norm,
mask=dec_mask,
layer_cache=layer_cache,
type="self",
)
query = self.self_attn(all_input, all_input, input_norm, mask=dec_mask, layer_cache=layer_cache, type="self",)
query = self.drop(query) + inputs
query_norm = self.layer_norm_2(query)
mid = self.context_attn(
memory_bank,
memory_bank,
query_norm,
mask=src_pad_mask,
layer_cache=layer_cache,
type="context",
memory_bank, memory_bank, query_norm, mask=src_pad_mask, layer_cache=layer_cache, type="context",
)
output = self.feed_forward(self.drop(mid) + query)
......@@ -492,14 +424,7 @@ class MultiHeadedAttention(nn.Module):
self.final_linear = nn.Linear(model_dim, model_dim)
def forward(
self,
key,
value,
query,
mask=None,
layer_cache=None,
type=None,
predefined_graph_1=None,
self, key, value, query, mask=None, layer_cache=None, type=None, predefined_graph_1=None,
):
"""
Compute the context vector and the attention vectors.
......@@ -531,11 +456,7 @@ class MultiHeadedAttention(nn.Module):
def unshape(x):
""" compute context """
return (
x.transpose(1, 2)
.contiguous()
.view(batch_size, -1, head_count * dim_per_head)
)
return x.transpose(1, 2).contiguous().view(batch_size, -1, head_count * dim_per_head)
# 1) Project key, value, and query.
if layer_cache is not None:
......@@ -554,9 +475,7 @@ class MultiHeadedAttention(nn.Module):
if layer_cache["self_keys"] is not None:
key = torch.cat((layer_cache["self_keys"].to(device), key), dim=2)
if layer_cache["self_values"] is not None:
value = torch.cat(
(layer_cache["self_values"].to(device), value), dim=2
)
value = torch.cat((layer_cache["self_values"].to(device), value), dim=2)
layer_cache["self_keys"] = key
layer_cache["self_values"] = value
elif type == "context":
......@@ -637,13 +556,9 @@ class DecoderState(object):
sizes = e.size()
br = sizes[1]
if len(sizes) == 3:
sent_states = e.view(sizes[0], beam_size, br // beam_size, sizes[2])[
:, :, idx
]
sent_states = e.view(sizes[0], beam_size, br // beam_size, sizes[2])[:, :, idx]
else:
sent_states = e.view(
sizes[0], beam_size, br // beam_size, sizes[2], sizes[3]
)[:, :, idx]
sent_states = e.view(sizes[0], beam_size, br // beam_size, sizes[2], sizes[3])[:, :, idx]
sent_states.data.copy_(sent_states.data.index_select(1, positions))
......@@ -716,11 +631,7 @@ class TransformerDecoderState(DecoderState):
def gelu(x):
return (
0.5
* x
* (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
)
return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
class PositionwiseFeedForward(nn.Module):
......@@ -758,9 +669,7 @@ class PositionwiseFeedForward(nn.Module):
def build_predictor(args, tokenizer, symbols, model, logger=None):
# we should be able to refactor the global scorer a lot
scorer = GNMTGlobalScorer(args.alpha, length_penalty="wu")
translator = Translator(
args, model, tokenizer, symbols, global_scorer=scorer, logger=logger
)
translator = Translator(args, model, tokenizer, symbols, global_scorer=scorer, logger=logger)
return translator
......@@ -891,9 +800,7 @@ class Translator(object):
Shouldn't need the original dataset.
"""
with torch.no_grad():
return self._fast_translate_batch(
batch, self.max_length, min_length=self.min_length
)
return self._fast_translate_batch(batch, self.max_length, min_length=self.min_length)
# Where the beam search lives
# I have no idea why it is being called from the method above
......@@ -912,26 +819,18 @@ class Translator(object):
mask_src = batch.mask_src
src_features = self.model.bert(src, segs, mask_src)
dec_states = self.model.decoder.init_decoder_state(
src, src_features, with_cache=True
)
dec_states = self.model.decoder.init_decoder_state(src, src_features, with_cache=True)
device = src_features.device
# Tile states and memory beam_size times.
dec_states.map_batch_fn(lambda state, dim: tile(state, beam_size, dim=dim))
src_features = tile(src_features, beam_size, dim=0)
batch_offset = torch.arange(batch_size, dtype=torch.long, device=device)
beam_offset = torch.arange(
0, batch_size * beam_size, step=beam_size, dtype=torch.long, device=device
)
alive_seq = torch.full(
[batch_size * beam_size, 1], self.start_token, dtype=torch.long, device=device
)
beam_offset = torch.arange(0, batch_size * beam_size, step=beam_size, dtype=torch.long, device=device)
alive_seq = torch.full([batch_size * beam_size, 1], self.start_token, dtype=torch.long, device=device)
# Give full probability to the first beam on the first step.
topk_log_probs = torch.tensor(
[0.0] + [float("-inf")] * (beam_size - 1), device=device
).repeat(batch_size)
topk_log_probs = torch.tensor([0.0] + [float("-inf")] * (beam_size - 1), device=device).repeat(batch_size)
# Structure that holds finished hypotheses.
hypotheses = [[] for _ in range(batch_size)] # noqa: F812
......@@ -948,9 +847,7 @@ class Translator(object):
# Decoder forward.
decoder_input = decoder_input.transpose(0, 1)
dec_out, dec_states = self.model.decoder(
decoder_input, src_features, dec_states, step=step
)
dec_out, dec_states = self.model.decoder(decoder_input, src_features, dec_states, step=step)
# Generator forward.
log_probs = self.generator.forward(dec_out.transpose(0, 1).squeeze(0))
......@@ -978,10 +875,7 @@ class Translator(object):
words = " ".join(words).replace(" ##", "").split()
if len(words) <= 3:
continue
trigrams = [
(words[i - 1], words[i], words[i + 1])
for i in range(1, len(words) - 1)
]
trigrams = [(words[i - 1], words[i], words[i + 1]) for i in range(1, len(words) - 1)]
trigram = tuple(trigrams[-1])
if trigram in trigrams[:-1]:
fail = True
......@@ -999,15 +893,11 @@ class Translator(object):
topk_ids = topk_ids.fmod(vocab_size)
# Map beam_index to batch_index in the flat representation.
batch_index = topk_beam_index + beam_offset[
: topk_beam_index.size(0)
].unsqueeze(1)
batch_index = topk_beam_index + beam_offset[: topk_beam_index.size(0)].unsqueeze(1)
select_indices = batch_index.view(-1)
# Append last prediction.
alive_seq = torch.cat(
[alive_seq.index_select(0, select_indices), topk_ids.view(-1, 1)], -1
)
alive_seq = torch.cat([alive_seq.index_select(0, select_indices), topk_ids.view(-1, 1)], -1)
is_finished = topk_ids.eq(self.end_token)
if step + 1 == max_length:
......@@ -1040,15 +930,11 @@ class Translator(object):
topk_log_probs = topk_log_probs.index_select(0, non_finished)
batch_index = batch_index.index_select(0, non_finished)
batch_offset = batch_offset.index_select(0, non_finished)
alive_seq = predictions.index_select(0, non_finished).view(
-1, alive_seq.size(-1)
)
alive_seq = predictions.index_select(0, non_finished).view(-1, alive_seq.size(-1))
# Reorder states.
select_indices = batch_index.view(-1)
src_features = src_features.index_select(0, select_indices)
dec_states.map_batch_fn(
lambda state, dim: state.index_select(dim, select_indices)
)
dec_states.map_batch_fn(lambda state, dim: state.index_select(dim, select_indices))
return results
......@@ -1089,14 +975,7 @@ def tile(x, count, dim=0):
out_size = list(x.size())
out_size[0] *= count
batch = x.size(0)
x = (
x.view(batch, -1)
.transpose(0, 1)
.repeat(count, 1)
.transpose(0, 1)
.contiguous()
.view(*out_size)
)
x = x.view(batch, -1).transpose(0, 1).repeat(count, 1).transpose(0, 1).contiguous().view(*out_size)
if dim != 0:
x = x.permute(perm).contiguous()
return x
......@@ -1107,6 +986,7 @@ def tile(x, count, dim=0):
# a finetuning script.
#
class BertSumOptimizer(object):
""" Specific optimizer for BertSum.
......@@ -1126,16 +1006,10 @@ class BertSumOptimizer(object):
self.optimizers = {
"encoder": torch.optim.Adam(
model.encoder.parameters(),
lr=lr["encoder"],
betas=(beta_1, beta_2),
eps=eps,
model.encoder.parameters(), lr=lr["encoder"], betas=(beta_1, beta_2), eps=eps,
),
"decoder": torch.optim.Adam(
model.decoder.parameters(),
lr=lr["decoder"],
betas=(beta_1, beta_2),
eps=eps,
model.decoder.parameters(), lr=lr["decoder"], betas=(beta_1, beta_2), eps=eps,
),
}
......@@ -1143,9 +1017,7 @@ class BertSumOptimizer(object):
self.current_learning_rates = {}
def _update_rate(self, stack):
return self.lr[stack] * min(
self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-1.5)
)
return self.lr[stack] * min(self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-1.5))
def zero_grad(self):
self.optimizer_decoder.zero_grad()
......
examples/summarization/run_summarization.py
View file @ fa84ae26
......@@ -25,9 +25,7 @@ logger = logging.getLogger(__name__)
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
Batch = namedtuple(
"Batch", ["document_names", "batch_size", "src", "segs", "mask_src", "tgt_str"]
)
Batch = namedtuple("Batch", ["document_names", "batch_size", "src", "segs", "mask_src", "tgt_str"])
def evaluate(args):
......@@ -48,13 +46,14 @@ def evaluate(args):
import rouge
import nltk
nltk.download('punkt')
nltk.download("punkt")
rouge_evaluator = rouge.Rouge(
metrics=['rouge-n', 'rouge-l'],
metrics=["rouge-n", "rouge-l"],
max_n=2,
limit_length=True,
length_limit=args.beam_size,
length_limit_type='words',
length_limit_type="words",
apply_avg=True,
apply_best=False,
alpha=0.5, # Default F1_score
......@@ -161,15 +160,15 @@ Recall >> {:.3f}
F1 >> {:.3f}
Precision >> {:.3f}
Recall >> {:.3f}""".format(
scores['rouge-1']['f'],
scores['rouge-1']['p'],
scores['rouge-1']['r'],
scores['rouge-2']['f'],
scores['rouge-2']['p'],
scores['rouge-2']['r'],
scores['rouge-l']['f'],
scores['rouge-l']['p'],
scores['rouge-l']['r'],
scores["rouge-1"]["f"],
scores["rouge-1"]["p"],
scores["rouge-1"]["r"],
scores["rouge-2"]["f"],
scores["rouge-2"]["p"],
scores["rouge-2"]["r"],
scores["rouge-l"]["f"],
scores["rouge-l"]["p"],
scores["rouge-l"]["r"],
)
......@@ -187,9 +186,7 @@ def build_data_iterator(args, tokenizer):
dataset = load_and_cache_examples(args, tokenizer)
sampler = SequentialSampler(dataset)
collate_fn = lambda data: collate(data, tokenizer, block_size=512, device=args.device)
iterator = DataLoader(
dataset, sampler=sampler, batch_size=args.batch_size, collate_fn=collate_fn,
)
iterator = DataLoader(dataset, sampler=sampler, batch_size=args.batch_size, collate_fn=collate_fn,)
return iterator
......@@ -210,14 +207,9 @@ def collate(data, tokenizer, block_size, device):
names = [name for name, _, _ in data]
summaries = [" ".join(summary_list) for _, _, summary_list in data]
encoded_text = [
encode_for_summarization(story, summary, tokenizer) for _, story, summary in data
]
encoded_text = [encode_for_summarization(story, summary, tokenizer) for _, story, summary in data]
encoded_stories = torch.tensor(
[
fit_to_block_size(story, block_size, tokenizer.pad_token_id)
for story, _ in encoded_text
]
[fit_to_block_size(story, block_size, tokenizer.pad_token_id) for story, _ in encoded_text]
)
encoder_token_type_ids = compute_token_type_ids(encoded_stories, tokenizer.cls_token_id)
encoder_mask = build_mask(encoded_stories, tokenizer.pad_token_id)
......@@ -272,38 +264,23 @@ def main():
)
# EVALUATION options
parser.add_argument(
"--no_cuda",
default=False,
type=bool,
help="Whether to force the execution on CPU.",
"--no_cuda", default=False, type=bool, help="Whether to force the execution on CPU.",
)
parser.add_argument(
"--batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.",
)
# BEAM SEARCH arguments
parser.add_argument(
"--min_length",
default=50,
type=int,
help="Minimum number of tokens for the summaries.",
"--min_length", default=50, type=int, help="Minimum number of tokens for the summaries.",
)
parser.add_argument(
"--max_length",
default=200,
type=int,
help="Maixmum number of tokens for the summaries.",
"--max_length", default=200, type=int, help="Maixmum number of tokens for the summaries.",
)
parser.add_argument(
"--beam_size",
default=5,
type=int,
help="The number of beams to start with for each example.",
"--beam_size", default=5, type=int, help="The number of beams to start with for each example.",
)
parser.add_argument(
"--alpha",
default=0.95,
type=float,
help="The value of alpha for the length penalty in the beam search.",
"--alpha", default=0.95, type=float, help="The value of alpha for the length penalty in the beam search.",
)
parser.add_argument(
"--block_trigram",
......
examples/summarization/utils_summarization.py
View file @ fa84ae26
......@@ -68,9 +68,7 @@ def process_story(raw_story):
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")])
)
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]
......@@ -135,13 +133,9 @@ def encode_for_summarization(story_lines, summary_lines, tokenizer):
sentences.
"""
story_lines_token_ids = [tokenizer.encode(line) for line in story_lines]
story_token_ids = [
token for sentence in story_lines_token_ids for token in sentence
]
story_token_ids = [token for sentence in story_lines_token_ids for token in sentence]
summary_lines_token_ids = [tokenizer.encode(line) for line in summary_lines]
summary_token_ids = [
token for sentence in summary_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
......
examples/summarization/utils_summarization_test.py
View file @ fa84ae26
......@@ -33,25 +33,19 @@ class SummarizationDataProcessingTest(unittest.TestCase):
""" 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
)
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
)
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
)
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.
......@@ -95,9 +89,7 @@ class SummarizationDataProcessingTest(unittest.TestCase):
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()
)
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])
......@@ -106,12 +98,8 @@ class SummarizationDataProcessingTest(unittest.TestCase):
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(
[[1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0], [1, 0, 0, 0, 1, 1]]
)
batch = torch.tensor([[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]])
expected = torch.tensor([[1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0], [1, 0, 0, 0, 1, 1]])
result = compute_token_type_ids(batch, separator)
np.testing.assert_array_equal(result, expected)
......
examples/test_examples.py
View file @ fa84ae26
......@@ -35,19 +35,21 @@ logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger()
def get_setup_file():
parser = argparse.ArgumentParser()
parser.add_argument('-f')
parser.add_argument("-f")
args = parser.parse_args()
return args.f
class ExamplesTests(unittest.TestCase):
class ExamplesTests(unittest.TestCase):
def test_run_glue(self):
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)
testargs = ["run_glue.py",
testargs = [
"run_glue.py",
"--data_dir=./examples/tests_samples/MRPC/",
"--task_name=mrpc",
"--do_train",
......@@ -59,10 +61,10 @@ class ExamplesTests(unittest.TestCase):
"--max_steps=10",
"--warmup_steps=2",
"--overwrite_output_dir",
"--seed=42"]
model_type, model_name = ("--model_type=bert",
"--model_name_or_path=bert-base-uncased")
with patch.object(sys, 'argv', testargs + [model_type, model_name]):
"--seed=42",
]
model_type, model_name = ("--model_type=bert", "--model_name_or_path=bert-base-uncased")
with patch.object(sys, "argv", testargs + [model_type, model_name]):
result = run_glue.main()
for value in result.values():
self.assertGreaterEqual(value, 0.75)
......@@ -71,7 +73,8 @@ class ExamplesTests(unittest.TestCase):
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)
testargs = ["run_squad.py",
testargs = [
"run_squad.py",
"--data_dir=./examples/tests_samples/SQUAD",
"--model_name=bert-base-uncased",
"--output_dir=./examples/tests_samples/temp_dir",
......@@ -84,27 +87,24 @@ class ExamplesTests(unittest.TestCase):
"--per_gpu_train_batch_size=2",
"--per_gpu_eval_batch_size=1",
"--overwrite_output_dir",
"--seed=42"]
model_type, model_name = ("--model_type=bert",
"--model_name_or_path=bert-base-uncased")
with patch.object(sys, 'argv', testargs + [model_type, model_name]):
"--seed=42",
]
model_type, model_name = ("--model_type=bert", "--model_name_or_path=bert-base-uncased")
with patch.object(sys, "argv", testargs + [model_type, model_name]):
result = run_squad.main()
self.assertGreaterEqual(result['f1'], 30)
self.assertGreaterEqual(result['exact'], 30)
self.assertGreaterEqual(result["f1"], 30)
self.assertGreaterEqual(result["exact"], 30)
def test_generation(self):
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)
testargs = ["run_generation.py",
"--prompt=Hello",
"--length=10",
"--seed=42"]
model_type, model_name = ("--model_type=openai-gpt",
"--model_name_or_path=openai-gpt")
with patch.object(sys, 'argv', testargs + [model_type, model_name]):
testargs = ["run_generation.py", "--prompt=Hello", "--length=10", "--seed=42"]
model_type, model_name = ("--model_type=openai-gpt", "--model_name_or_path=openai-gpt")
with patch.object(sys, "argv", testargs + [model_type, model_name]):
result = run_generation.main()
self.assertGreaterEqual(len(result), 10)
if __name__ == "__main__":
unittest.main()
examples/utils_multiple_choice.py
View file @ fa84ae26
......@@ -55,19 +55,10 @@ class InputExample(object):
class InputFeatures(object):
def __init__(self,
example_id,
choices_features,
label
):
def __init__(self, example_id, choices_features, label):
self.example_id = example_id
self.choices_features = [
{
'input_ids': input_ids,
'input_mask': input_mask,
'segment_ids': segment_ids
}
{"input_ids": input_ids, "input_mask": input_mask, "segment_ids": segment_ids}
for input_ids, input_mask, segment_ids in choices_features
]
self.label = label
......@@ -99,29 +90,29 @@ class RaceProcessor(DataProcessor):
def get_train_examples(self, data_dir):
"""See base class."""
logger.info("LOOKING AT {} train".format(data_dir))
high = os.path.join(data_dir, 'train/high')
middle = os.path.join(data_dir, 'train/middle')
high = os.path.join(data_dir, "train/high")
middle = os.path.join(data_dir, "train/middle")
high = self._read_txt(high)
middle = self._read_txt(middle)
return self._create_examples(high + middle, 'train')
return self._create_examples(high + middle, "train")
def get_dev_examples(self, data_dir):
"""See base class."""
logger.info("LOOKING AT {} dev".format(data_dir))
high = os.path.join(data_dir, 'dev/high')
middle = os.path.join(data_dir, 'dev/middle')
high = os.path.join(data_dir, "dev/high")
middle = os.path.join(data_dir, "dev/middle")
high = self._read_txt(high)
middle = self._read_txt(middle)
return self._create_examples(high + middle, 'dev')
return self._create_examples(high + middle, "dev")
def get_test_examples(self, data_dir):
"""See base class."""
logger.info("LOOKING AT {} test".format(data_dir))
high = os.path.join(data_dir, 'test/high')
middle = os.path.join(data_dir, 'test/middle')
high = os.path.join(data_dir, "test/high")
middle = os.path.join(data_dir, "test/middle")
high = self._read_txt(high)
middle = self._read_txt(middle)
return self._create_examples(high + middle, 'test')
return self._create_examples(high + middle, "test")
def get_labels(self):
"""See base class."""
......@@ -131,13 +122,12 @@ class RaceProcessor(DataProcessor):
lines = []
files = glob.glob(input_dir + "/*txt")
for file in tqdm.tqdm(files, desc="read files"):
with open(file, 'r', encoding='utf-8') as fin:
with open(file, "r", encoding="utf-8") as fin:
data_raw = json.load(fin)
data_raw["race_id"] = file
lines.append(data_raw)
return lines
def _create_examples(self, lines, set_type):
"""Creates examples for the training and dev sets."""
examples = []
......@@ -145,9 +135,9 @@ class RaceProcessor(DataProcessor):
race_id = "%s-%s" % (set_type, data_raw["race_id"])
article = data_raw["article"]
for i in range(len(data_raw["answers"])):
truth = str(ord(data_raw['answers'][i]) - ord('A'))
question = data_raw['questions'][i]
options = data_raw['options'][i]
truth = str(ord(data_raw["answers"][i]) - ord("A"))
question = data_raw["questions"][i]
options = data_raw["options"][i]
examples.append(
InputExample(
......@@ -155,9 +145,12 @@ class RaceProcessor(DataProcessor):
question=question,
contexts=[article, article, article, article], # this is not efficient but convenient
endings=[options[0], options[1], options[2], options[3]],
label=truth))
label=truth,
)
)
return examples
class SwagProcessor(DataProcessor):
"""Processor for the SWAG data set."""
......@@ -179,27 +172,25 @@ class SwagProcessor(DataProcessor):
"setting!"
)
return self._create_examples(self._read_csv(os.path.join(data_dir, "test.csv")), "test")
def get_labels(self):
"""See base class."""
return ["0", "1", "2", "3"]
def _read_csv(self, input_file):
with open(input_file, 'r', encoding='utf-8') as f:
with open(input_file, "r", encoding="utf-8") as f:
reader = csv.reader(f)
lines = []
for line in reader:
if sys.version_info[0] == 2:
line = list(unicode(cell, 'utf-8') for cell in line)
line = list(unicode(cell, "utf-8") for cell in line)
lines.append(line)
return lines
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(
"For training, the input file must contain a label column."
)
if type == "train" and lines[0][-1] != "label":
raise ValueError("For training, the input file must contain a label column.")
examples = [
InputExample(
......@@ -207,10 +198,11 @@ class SwagProcessor(DataProcessor):
question=line[5], # in the swag dataset, the
# common beginning of each
# choice is stored in "sent2".
contexts = [line[4], line[4], line[4], line[4]],
endings = [line[7], line[8], line[9], line[10]],
label=line[11]
) for line in lines[1:] # we skip the line with the column names
contexts=[line[4], line[4], line[4], line[4]],
endings=[line[7], line[8], line[9], line[10]],
label=line[11],
)
for line in lines[1:] # we skip the line with the column names
]
return examples
......@@ -238,15 +230,14 @@ class ArcProcessor(DataProcessor):
return ["0", "1", "2", "3"]
def _read_json(self, input_file):
with open(input_file, 'r', encoding='utf-8') as fin:
with open(input_file, "r", encoding="utf-8") as fin:
lines = fin.readlines()
return lines
def _create_examples(self, lines, type):
"""Creates examples for the training and dev sets."""
#There are two types of labels. They should be normalized
# There are two types of labels. They should be normalized
def normalize(truth):
if truth in "ABCD":
return ord(truth) - ord("A")
......@@ -283,12 +274,18 @@ class ArcProcessor(DataProcessor):
if len(options) == 4:
examples.append(
InputExample(
example_id = id,
example_id=id,
question=question,
contexts=[options[0]["para"].replace("_", ""), options[1]["para"].replace("_", ""),
options[2]["para"].replace("_", ""), options[3]["para"].replace("_", "")],
contexts=[
options[0]["para"].replace("_", ""),
options[1]["para"].replace("_", ""),
options[2]["para"].replace("_", ""),
options[3]["para"].replace("_", ""),
],
endings=[options[0]["text"], options[1]["text"], options[2]["text"], options[3]["text"]],
label=truth))
label=truth,
)
)
if type == "train":
assert len(examples) > 1
......@@ -316,7 +313,7 @@ def convert_examples_to_features(
Loads a data file into a list of `InputFeatures`
"""
label_map = {label : i for i, label in enumerate(label_list)}
label_map = {label: i for i, label in enumerate(label_list)}
features = []
for (ex_index, example) in tqdm.tqdm(enumerate(examples), desc="convert examples to features"):
......@@ -331,16 +328,13 @@ def convert_examples_to_features(
else:
text_b = example.question + " " + ending
inputs = tokenizer.encode_plus(
text_a,
text_b,
add_special_tokens=True,
max_length=max_length,
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!"
)
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, token_type_ids = inputs["input_ids"], inputs["token_type_ids"]
......@@ -364,7 +358,6 @@ def convert_examples_to_features(
assert len(token_type_ids) == max_length
choices_features.append((input_ids, attention_mask, token_type_ids))
label = label_map[example.label]
if ex_index < 2:
......@@ -372,33 +365,17 @@ def convert_examples_to_features(
logger.info("race_id: {}".format(example.example_id))
for choice_idx, (input_ids, attention_mask, token_type_ids) in enumerate(choices_features):
logger.info("choice: {}".format(choice_idx))
logger.info("input_ids: {}".format(' '.join(map(str, input_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("input_ids: {}".format(" ".join(map(str, input_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,
)
)
features.append(InputFeatures(example_id=example.example_id, choices_features=choices_features, label=label,))
return features
processors = {"race": RaceProcessor, "swag": SwagProcessor, "arc": ArcProcessor}
processors = {
"race": RaceProcessor,
"swag": SwagProcessor,
"arc": ArcProcessor
}
MULTIPLE_CHOICE_TASKS_NUM_LABELS = {
"race", 4,
"swag", 4,
"arc", 4
}
MULTIPLE_CHOICE_TASKS_NUM_LABELS = {"race", 4, "swag", 4, "arc", 4}
examples/utils_ner.py
View file @ fa84ae26
......@@ -61,9 +61,7 @@ def read_examples_from_file(data_dir, mode):
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))
examples.append(InputExample(guid="{}-{}".format(mode, guid_index), words=words, labels=labels))
guid_index += 1
words = []
labels = []
......@@ -76,13 +74,12 @@ def read_examples_from_file(data_dir, mode):
# 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))
examples.append(InputExample(guid="%s-%d".format(mode, guid_index), words=words, labels=labels))
return examples
def convert_examples_to_features(examples,
def convert_examples_to_features(
examples,
label_list,
max_seq_length,
tokenizer,
......@@ -96,7 +93,8 @@ def convert_examples_to_features(examples,
pad_token_segment_id=0,
pad_token_label_id=-100,
sequence_a_segment_id=0,
mask_padding_with_zero=True):
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]
......@@ -122,8 +120,8 @@ def convert_examples_to_features(examples,
# 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)]
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:
......@@ -174,10 +172,10 @@ def convert_examples_to_features(examples,
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)
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
......@@ -194,10 +192,8 @@ def convert_examples_to_features(examples,
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))
InputFeatures(input_ids=input_ids, input_mask=input_mask, segment_ids=segment_ids, label_ids=label_ids)
)
return features
......
hubconf.py
View file @ fa84ae26
from transformers import (
AutoTokenizer, AutoConfig, AutoModel, AutoModelWithLMHead, AutoModelForSequenceClassification, AutoModelForQuestionAnswering
AutoTokenizer,
AutoConfig,
AutoModel,
AutoModelWithLMHead,
AutoModelForSequenceClassification,
AutoModelForQuestionAnswering,
)
from transformers.file_utils import add_start_docstrings
dependencies = ['torch', 'tqdm', 'boto3', 'requests', 'regex', 'sentencepiece', 'sacremoses']
dependencies = ["torch", "tqdm", "boto3", "requests", "regex", "sentencepiece", "sacremoses"]
@add_start_docstrings(AutoConfig.__doc__)
def config(*args, **kwargs):
......@@ -57,6 +63,7 @@ def model(*args, **kwargs):
return AutoModel.from_pretrained(*args, **kwargs)
@add_start_docstrings(AutoModelWithLMHead.__doc__)
def modelWithLMHead(*args, **kwargs):
r"""
......
setup.py
View file @ fa84ae26
......@@ -38,11 +38,11 @@ from setuptools import find_packages, setup
extras = {
'serving': ['pydantic', 'uvicorn', 'fastapi'],
'serving-tf': ['pydantic', 'uvicorn', 'fastapi', 'tensorflow'],
'serving-torch': ['pydantic', 'uvicorn', 'fastapi', 'torch']
"serving": ["pydantic", "uvicorn", "fastapi"],
"serving-tf": ["pydantic", "uvicorn", "fastapi", "tensorflow"],
"serving-torch": ["pydantic", "uvicorn", "fastapi", "torch"],
}
extras['all'] = [package for package in extras.values()]
extras["all"] = [package for package in extras.values()]
setup(
name="transformers",
......@@ -50,30 +50,29 @@ setup(
author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
author_email="thomas@huggingface.co",
description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",
long_description=open("README.md", "r", encoding='utf-8').read(),
long_description=open("README.md", "r", encoding="utf-8").read(),
long_description_content_type="text/markdown",
keywords='NLP deep learning transformer pytorch tensorflow BERT GPT GPT-2 google openai CMU',
license='Apache',
keywords="NLP deep learning transformer pytorch tensorflow BERT GPT GPT-2 google openai CMU",
license="Apache",
url="https://github.com/huggingface/transformers",
packages=find_packages(exclude=["*.tests", "*.tests.*",
"tests.*", "tests"]),
install_requires=['numpy',
'boto3',
'filelock',
'requests',
'tqdm',
'regex != 2019.12.17',
'sentencepiece',
'sacremoses'],
extras_require=extras,
scripts=[
'transformers-cli'
packages=find_packages(exclude=["*.tests", "*.tests.*", "tests.*", "tests"]),
install_requires=[
"numpy",
"boto3",
"filelock",
"requests",
"tqdm",
"regex != 2019.12.17",
"sentencepiece",
"sacremoses",
],
extras_require=extras,
scripts=["transformers-cli"],
# python_requires='>=3.5.0',
classifiers=[
'Intended Audience :: Science/Research',
'License :: OSI Approved :: Apache Software License',
'Programming Language :: Python :: 3',
'Topic :: Scientific/Engineering :: Artificial Intelligence',
"Intended Audience :: Science/Research",
"License :: OSI Approved :: Apache Software License",
"Programming Language :: Python :: 3",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
],
)
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