nreimers's picture
upload
761beea
import argparse
import logging
import math
import os
from datetime import datetime
import datasets
import torch
from torch.utils.data import DataLoader
from tqdm.auto import tqdm
import sys
import transformers
from accelerate import Accelerator, DistributedType
from shutil import copyfile
import wandb
import numpy as np
from transformers import (
MODEL_MAPPING,
AutoModelForMaskedLM,
AutoTokenizer,
DataCollatorForLanguageModeling,
SchedulerType,
get_scheduler
)
from transformers.utils.versions import require_version
class TrainDataset(torch.utils.data.IterableDataset):
def __init__(self, filepath, tokenizer, max_length, batch_size, train_samples):
self.tokenizer = tokenizer
self.fIn = open(filepath)
self.max_length = max_length
self.batch_size = batch_size
self.train_samples = train_samples
def __iter__(self):
batch = []
for sent in self.fIn:
batch.append(sent.strip()[0:1000])
if len(batch) >= self.batch_size:
#Use multi process tokenization
encoded = self.tokenizer(batch, add_special_tokens=True, truncation=True, max_length=self.max_length, return_special_tokens_mask=True, padding=True)
#print(len(encoded['input_ids'][0]))
for idx in range(len(batch)):
single_sample = {key: encoded[key][idx] for key in encoded}
yield single_sample
batch = []
def __len__(self):
return self.train_samples
## Dev dataset
class DevDataset(torch.utils.data.Dataset):
def __init__(self, filepath, tokenizer, max_length):
self.tokenizer = tokenizer
self.max_length = max_length
with open(filepath) as fIn:
sentences = [sent.strip() for sent in fIn]
self.num_sentences = len(sentences)
self.tokenized = self.tokenizer(sentences, add_special_tokens=True, truncation=True, max_length=self.max_length, return_special_tokens_mask=True)
def __getitem__(self, idx):
return {key: self.tokenized[key][idx] for key in self.tokenized}
def __len__(self):
return self.num_sentences
logger = logging.getLogger(__name__)
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")
MODEL_CONFIG_CLASSES = list(MODEL_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
def parse_args():
parser = argparse.ArgumentParser(description="Finetune a transformers model on a Masked Language Modeling task")
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The configuration name of the dataset to use (via the datasets library).",
)
parser.add_argument(
"--train_file", type=str, default=None, help="A text file data (1 text per line).."
)
parser.add_argument(
"--dev_file", type=str, default=None, help="A text file data (1 text per line)."
)
parser.add_argument(
"--model_name",
default="nicoladecao/msmarco-word2vec256000-distilbert-base-uncased",
type=str,
help="Path to pretrained model or model identifier from huggingface.co/models."
)
parser.add_argument(
"--per_device_batch_size",
type=int,
default=16,
help="Batch size (per device) for the training dataloader.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=5e-5,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument("--weight_decay", type=float, default=0.01, help="Weight decay to use.")
parser.add_argument("--num_train_epochs", type=int, default=1, help="Total number of training epochs to perform.")
parser.add_argument(
"--max_train_steps",
type=int,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
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(
"--lr_scheduler_type",
type=SchedulerType,
default="linear",
help="The scheduler type to use.",
choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
)
parser.add_argument(
"--num_warmup_steps", type=int, default=1000, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument(
"--model_type",
type=str,
default=None,
help="Model type to use if training from scratch.",
choices=MODEL_TYPES,
)
parser.add_argument(
"--max_seq_length",
type=int,
default=256,
help="The maximum total input sequence length after tokenization. Sequences longer than this will be truncated.",
)
parser.add_argument(
"--line_by_line",
type=bool,
default=True,
help="Whether distinct lines of text in the dataset are to be handled as distinct sequences.",
)
parser.add_argument(
"--overwrite_cache", type=bool, default=False, help="Overwrite the cached training and evaluation sets"
)
parser.add_argument(
"--mlm_probability", type=float, default=0.15, help="Ratio of tokens to mask for masked language modeling loss"
)
parser.add_argument("--mixed_precision", default="fp16")
parser.add_argument("--train_samples", required=True, type=int)
parser.add_argument("--eval_steps", default=10000, type=int)
parser.add_argument("--max_grad_norm", default=1.0, type=float)
parser.add_argument("--project", default="bert-word2vec")
parser.add_argument("--freeze_emb_layer", default=False, action='store_true')
parser.add_argument("--log_interval", default=1000, type=int)
parser.add_argument("--ckp_steps", default=50000, type=int)
args = parser.parse_args()
return args
def main():
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator(mixed_precision=args.mixed_precision)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state)
# Setup logging, we only want one process per machine to log things on the screen.
# accelerator.is_local_main_process is only True for one process per machine.
logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
accelerator.wait_for_everyone()
#Load model
tokenizer = AutoTokenizer.from_pretrained(args.model_name)
model = AutoModelForMaskedLM.from_pretrained(args.model_name)
#Freeze emb layer
if args.freeze_emb_layer:
model.distilbert.embeddings.word_embeddings.requires_grad_(False)
# Logging & Co on main process
if accelerator.is_main_process:
exp_name = f'{args.model_name.replace("/", "-")}-{"freeze_emb" if args.freeze_emb_layer else "update_emb"}-{datetime.now().strftime("%Y-%m-%d_%H-%M-%S")}'
output_dir = os.path.join("output-mlm", exp_name)
wandb.init(project=args.project, name=exp_name, config=args)
os.makedirs(output_dir, exist_ok=False)
#Save tokenizer
tokenizer.save_pretrained(output_dir)
#Save train script
train_script_path = os.path.join(output_dir, 'train_script.py')
copyfile(__file__, train_script_path)
with open(train_script_path, 'a') as fOut:
fOut.write("\n\n# Script was called via:\n#python " + " ".join(sys.argv))
total_batch_size = args.per_device_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
train_dataset = TrainDataset(args.train_file, tokenizer, args.max_seq_length, batch_size=total_batch_size, train_samples=args.train_samples)
eval_dataset = DevDataset(args.dev_file, tokenizer, args.max_seq_length)
# Data collator
# This one will take care of randomly masking the tokens.
data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=args.mlm_probability)
# DataLoaders creation:
train_dataloader = DataLoader(train_dataset, collate_fn=data_collator, batch_size=args.per_device_batch_size)
eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
"weight_decay": args.weight_decay,
},
{
"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
"weight_decay": 0.0,
},
]
optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(model, optimizer, train_dataloader, eval_dataloader)
# On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
if accelerator.distributed_type == DistributedType.TPU:
model.tie_weights()
# Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
# shorter in multiprocess)
# Scheduler and math around the number of training steps.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
else:
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps,
num_training_steps=args.max_train_steps,
)
# Train!
logger.info("***** Running training *****")
logger.info(f" Num examples = {args.train_samples}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process, smoothing=0.05)
completed_steps = 0
train_loss_values = []
best_eval_loss = 999999
if accelerator.is_main_process:
best_ckp_dir = os.path.join(output_dir, "best")
tokenizer.save_pretrained(best_ckp_dir)
for epoch in range(args.num_train_epochs):
logger.info(f"Start epoch {epoch}")
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
if accelerator.is_main_process:
train_loss_values.append(loss.cpu().item())
accelerator.backward(loss)
accelerator.clip_grad_norm_(model.parameters(), args.max_grad_norm)
if step % args.gradient_accumulation_steps == 0:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
### Do logging
if accelerator.is_main_process:
if completed_steps % args.log_interval == 0:
wandb.log({"train/loss": np.mean(train_loss_values)}, step=completed_steps)
train_loss_values = []
if completed_steps % args.eval_steps == 0:
model.eval()
losses = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
loss = outputs.loss
losses.append(accelerator.gather(loss.repeat(args.per_device_batch_size)))
losses = torch.cat(losses)
losses = losses[: len(eval_dataset)]
try:
eval_loss = torch.mean(losses)
except OverflowError:
eval_loss = float("inf")
logger.info(f"step {completed_steps}: perplexity: {eval_loss}")
if accelerator.is_main_process:
wandb.log({"eval/loss": eval_loss}, step=completed_steps)
model.train()
#Save model
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(output_dir, save_function=accelerator.save)
with open(os.path.join(output_dir, "train_steps.log"), 'a') as fOut:
fOut.write(f"{completed_steps}: {eval_loss}\n")
#Save best model
if eval_loss < best_eval_loss:
best_eval_loss = eval_loss
unwrapped_model.save_pretrained(best_ckp_dir, save_function=accelerator.save)
with open(os.path.join(best_ckp_dir, "train_steps.log"), 'a') as fOut:
fOut.write(f"{completed_steps}: {eval_loss}\n")
if accelerator.is_main_process and completed_steps % args.ckp_steps == 0:
ckp_dir = os.path.join(output_dir, f"ckp-{int(completed_steps/1000)}k")
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(ckp_dir, save_function=accelerator.save)
tokenizer.save_pretrained(ckp_dir)
with open(os.path.join(ckp_dir, "train_steps.log"), 'a') as fOut:
fOut.write(f"{completed_steps}: {eval_loss}\n")
if completed_steps >= args.max_train_steps:
break
if args.output_dir is not None:
accelerator.wait_for_everyone()
if accelerator.is_main_process:
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(output_dir, save_function=accelerator.save)
with open(os.path.join(output_dir, "train_steps.log"), 'a') as fOut:
fOut.write(f"{completed_steps}\n")
if __name__ == "__main__":
main()
# Script was called via:
#python train_mlm-iterable.py --train_file data/c4_msmarco_news_s2orc_wiki_train.txt --dev_file data/c4_msmarco_news_s2orc_wiki_dev.txt --train_samples 100000000 --model_name train-w2v-model/c4_msmarco_news_s2orc_wiki/distilbert-256k/ --freeze_emb_layer