def _setup_dataloader_from_config(self, cfg: DictConfig, mode: str):
tokenizer, model = self._tokenizer, self.model
start_time = perf_counter()
logging.info(f'Creating {mode} dataset')
input_file = cfg.data_path
dataset = TextNormalizationDecoderDataset(
input_file,
tokenizer,
cfg.mode,
cfg.get('max_decoder_len', tokenizer.model_max_length),
cfg.get('decoder_data_augmentation', False),
)
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=constants.LABEL_PAD_TOKEN_ID,
)
dl = torch.utils.data.DataLoader(
dataset=dataset,
batch_size=cfg.batch_size,
shuffle=cfg.shuffle,
collate_fn=data_collator,
)
running_time = perf_counter() - start_time
logging.info(f'Took {running_time} seconds')
return dl
def train(self):
# Create the output directory if needed
os.makedirs(self.training_args.output_dir, exist_ok=True)
# Load pretrained model and tokenizer
print('Loading model and tokenizer for',
self.gen_args['model_name_or_path'])
tokenizer_name = self.gen_args.get('tok_name_or_path',
self.gen_args['model_name_or_path'])
self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
self.model = AutoModelForSeq2SeqLM.from_pretrained(
self.gen_args['model_name_or_path'], **self.model_args)
self.model.config.task_specific_params = {'parse_amr': self.gen_args}
assert self.model.config.no_repeat_ngram_size == 0 # required for good results
# `use_cache=True` is incompatible with gradient checkpointing
if self.training_args.gradient_checkpointing:
self.model.config.use_cache = False
# Save the tokenizer
if self.gen_args.get('save_tokenizer', False):
self.tokenizer.save_pretrained(self.training_args.output_dir)
# Load the datasets
print('Building datasets')
train_file_path = os.path.join(self.gen_args['corpus_dir'],
self.gen_args['train_fn'])
train_dataset = self.build_dataset(train_file_path)
print('Training data is {:,} after removing {:,} long entries'.format( \
len(train_dataset), len(train_dataset.bad_indexes)))
# Load the evaluation dataset
eval_fn = self.gen_args.get('eval_fn')
if eval_fn:
eval_file_path = os.path.join(self.gen_args['corpus_dir'], eval_fn)
eval_samples = load_and_serialize(eval_file_path)
print('Evaluation data is {:,} samples'.format(
len(eval_samples['graphs'])))
else:
eval_samples = None
# Train the model
print('Training')
collator = DataCollatorForSeq2Seq(
self.tokenizer,
self.model,
padding=True,
max_length=self.gen_args['max_train_graph_len'],
pad_to_multiple_of=8 if self.training_args.fp16 else None)
trainer = AMRTrainer(model=self.model,
args=self.training_args,
train_dataset=train_dataset,
data_collator=collator,
eval_tokenizer=self.tokenizer,
eval_samples=eval_samples,
callbacks=[CudaMemoryCB()])
# If resume_from_checkpoint is True it will look for the last checkpoint in the value of output_dir
# passed via TrainingArguments. If it's a path to a specific checkpoint it will use that saved
# checkpoint folder to resume the training from.
trainer.train(
resume_from_checkpoint=self.training_args.resume_from_checkpoint)
# Save the results
if self.gen_args.get('save_at_end', False):
print('Saving model')
trainer.save_model(self.training_args.output_dir)
def loaders(self):
if self._loaders is None:
ps, t = self.params, self.tokenizer
c = DataCollatorForSeq2Seq(
t,
model=self.model,
label_pad_token_id=-100 if ps.ignore_pad_token_for_loss else t.PAD,
pad_to_multiple_of=8 if ps.fp16 else None,
)
t = DataLoader(
self.train_ds, shuffle=True, collate_fn=c, batch_size=ps.train_batch_size
)
e = DataLoader(self.eval_ds, collate_fn=c, batch_size=ps.eval_batch_size)
self._loaders = {TRAIN: t, EVAL: e}
return self._loaders
def get_dataloader(self,
type_path: str,
batch_size: int,
shuffle: bool = False) -> DataLoader:
dataset = self.get_dataset(type_path)
data_collator = DataCollatorForSeq2Seq(
self.tokenizer,
label_pad_token_id=-100, #self.tokenizer.pad_token_id,
pad_to_multiple_of=8 if self.hparams.fp16 else None,
)
#sampling method
#used to disable the distributed setting. Indeed
#local_rank: int = field(default=-1, metadata={"help": "For distributed training: local_rank"})
def _get_train_sampler(dataset):
#########################################
if isinstance(dataset, torch.utils.data.IterableDataset):
return None
# elif is_torch_tpu_available():
# return get_tpu_sampler(dataset)
else:
if (self.hparams.sortish_sampler):
dataset.make_sortish_sampler(
self.hparams.train_batch_size,
distributed=(self.hparams.local_rank != -1)
#distributed=(self.args.parallel_mode == ParallelMode.DISTRIBUTED),
) #check def parallel_mode(self): in training_args.py for more details.
return (RandomSampler(dataset) if self.hparams.local_rank == -1
else DistributedSampler(self.train_dataset))
##########################################
train_sampler = _get_train_sampler(dataset)
return DataLoader(
dataset,
batch_size=batch_size,
sampler=train_sampler,
collate_fn=data_collator,
num_workers=self.num_workers,
)
def main():
# Parse the arguments
args = parse_args()
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# 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()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.model_name_or_path)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if args.model_name_or_path:
model = AutoModelForSeq2SeqLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForSeq2SeqLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (args.target_lang is not None and args.source_lang
is not None), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = args.source_prefix if args.source_prefix is not None else ""
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if args.source_lang is not None:
tokenizer.src_lang = args.source_lang
if args.target_lang is not None:
tokenizer.tgt_lang = args.target_lang
# Get the language codes for input/target.
source_lang = args.source_lang.split("_")[0]
target_lang = args.target_lang.split("_")[0]
padding = "max_length" if args.pad_to_max_length else False
# Temporarily set max_target_length for training.
max_target_length = args.max_target_length
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
label_pad_token_id = -100 if args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
)
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_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 = 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)
# 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,
)
metric = load_metric("sacrebleu")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {args.per_device_train_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)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(
train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
if args.val_max_target_length is None:
args.val_max_target_length = args.max_target_length
gen_kwargs = {
"max_length":
args.val_max_target_length
if args is not None else config.max_length,
"num_beams":
args.num_beams,
}
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
generated_tokens = accelerator.unwrap_model(model).generate(
batch["input_ids"],
attention_mask=batch["attention_mask"],
**gen_kwargs,
)
generated_tokens = accelerator.pad_across_processes(
generated_tokens, dim=1, pad_index=tokenizer.pad_token_id)
labels = batch["labels"]
if not args.pad_to_max_length:
# If we did not pad to max length, we need to pad the labels too
labels = accelerator.pad_across_processes(
batch["labels"],
dim=1,
pad_index=tokenizer.pad_token_id)
generated_tokens = accelerator.gather(
generated_tokens).cpu().numpy()
labels = accelerator.gather(labels).cpu().numpy()
if args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels,
tokenizer.pad_token_id)
decoded_preds = tokenizer.batch_decode(
generated_tokens, skip_special_tokens=True)
decoded_labels = tokenizer.batch_decode(
labels, skip_special_tokens=True)
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
metric.add_batch(predictions=decoded_preds,
references=decoded_labels)
eval_metric = metric.compute()
logger.info({"bleu": eval_metric["score"]})
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir,
save_function=accelerator.save)
def main():
parser = HfArgumentParser((ModelArguments, DataTrainingArguments,DataValidationArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
model_args, data_args, test_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, test_args, training_args = parser.parse_args_into_dataclasses()
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.target_lang]
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
# Get the default prefix if None is passed.
if data_args.source_prefix is None:
task_specific_params = model.config.task_specific_params
if task_specific_params is not None:
prefix = task_specific_params.get("prefix", "")
else:
prefix = ""
else:
prefix = data_args.source_prefix
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
# To serialize preprocess_function below, each of those four variables needs to be defined (even if we won't use
# them all).
text_column, summary_column = None, None
if data_args.task.startswith("summarization"):
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None)
if data_args.text_column is None:
text_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if data_args.summary_column is None:
summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
# Tokenize Input
model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets, max_length=max_target_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(tokenizer, label_pad_token_id=label_pad_token_id)
# Metric
metric_name = "rouge"
metric = load_metric(metric_name)
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds = [pred.strip() for pred in decoded_preds]
decoded_labels = [label.strip() for label in decoded_labels]
if metric_name == "sacrebleu":
decoded_labels = [[label] for label in decoded_labels]
result = metric.compute(predictions=decoded_preds, references=decoded_labels)
# Extract a few results from ROUGE
if metric_name == "rouge":
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
else:
result = {"bleu": result["score"]}
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
result["gen_len"] = np.mean(prediction_lens)
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
train_result = trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
model = trainer.model
tokenizer = trainer.tokenizer
print("\n")
print("Running Evaluation Script")
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
output = []
inco = 0
min_length = test_args.min_summ_length
max_length = test_args.max_summ_length
df_test = pd.read_csv(data_args.validation_file)
for index,row in tqdm(df_test.iterrows(), total=df_test.shape[0]):
text = row['Text']
ref = row['Summary']
input_tokenized = tokenizer.encode(text, return_tensors='pt',max_length=data_args.max_source_length, truncation=True).to(device)
dim = list(input_tokenized.size())
summary_ids = model.generate(input_tokenized,
num_beams=test_args.num_beams,
no_repeat_ngram_size=test_args.no_repeat_ngram_size,
length_penalty=test_args.length_penalty,
min_length=min_length,
max_length=max_length,
early_stopping=True)
summ = [tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summary_ids][0]
if summ.find("nnnn")!=-1:
summ = re.sub(r'nnn*nn', '', summ)
inco = inco + 1
score = evaluate_summary(ref,summ)
output.append((score.precision,ref,summ))
print("Evaluation Completed")
precision = [round(x[0],4) for x in output]
fmeasure = [round(evaluate_summary(x[1],x[2]).fmeasure,4) for x in output]
actual = [x[1] for x in output]
generated = [re.sub(r'nnn*n', '',x[2]) for x in output]
df = pd.DataFrame({'Generated Summary':generated,'Actual Summary':actual, 'Precision': precision, 'F Score': fmeasure})
csv_output = os.path.join(training_args.output_dir, str(len(df_test)) + "-test_results.csv")
df.to_csv(csv_output)
print("Evaluation results saved in {}".format(csv_output))
output_df = pd.read_csv(csv_output)
length_df = len(output_df)
top = 10
if length_df < 20:
top = int(length_df/2) - 1
if top <=0 :
top = 1
if length_df!=0:
final_output = [(x['Precision'],x['Actual Summary'],x['Generated Summary'],x['F Score']) for ind,x in output_df.iterrows()]
output_desc = sorted(final_output, key = lambda x: -x[0])
fsc = np.mean([t[3] for t in output_desc])
pre = np.mean([t[0] for t in output_desc])
output_eval_file = os.path.join(training_args.output_dir, "evaluation_scores.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
writer.write("-----------------------------------------------------------------------------------------------------------------------------------------------------------------")
writer.write("\n")
writer.write("Mean F Measure: {:.4f}".format(fsc))
writer.write("\n")
writer.write("Mean Precision (Rouge1): {:.4f}".format(pre))
writer.write("\n")
writer.write("-----------------------------------------------------------------------------------------------------------------------------------------------------------------")
writer.write("\n")
writer.write("Best {}: ".format(top))
writer.write("\n")
for tup in output_desc[0:top]:
writer.write("F Measure: {}".format(tup[3]))
writer.write("\n")
writer.write("Precision: {}".format(tup[0]))
writer.write("\n")
writer.write("Actual Summary:")
writer.write("\n")
writer.write(tup[1])
writer.write("\n")
writer.write("Generated Summary:")
writer.write("\n")
writer.write(tup[2])
writer.write("\n")
writer.write("-----------------------------------------------------------------------------------------------------------------------------------------------------------------")
writer.write("\n")
writer.write("\n\n")
writer.write("Worst {}: ".format(top))
writer.write("\n")
n = len(final_output)
output_asc = sorted(final_output, key = lambda x: x[0])
for tup in output_asc[:top]:
writer.write("F Measure: {}".format(tup[3]))
writer.write("\n")
writer.write("Precision: {}".format(tup[0]))
writer.write("\n")
writer.write("Actual Summary: ")
writer.write("\n")
writer.write(tup[1])
writer.write("\n")
writer.write("Generated Summary:")
writer.write("\n")
writer.write(tup[2])
writer.write("\n")
writer.write("-----------------------------------------------------------------------------------------------------------------------------------------------------------------")
writer.write("\n")
print("Evaluation scores saved in {}".format(output_eval_file))
return results
def main():
args = parse_args()
if args.source_prefix is None and args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
"`--source_prefix 'summarize: ' `"
)
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# 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()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if args.model_name_or_path:
model = AutoModelForSeq2SeqLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForSeq2SeqLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
prefix = args.source_prefix if args.source_prefix is not None else ""
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(args.dataset_name, None)
if args.text_column is None:
text_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
text_column = args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if args.summary_column is None:
summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
summary_column = args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = args.max_target_length
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs, max_length=args.max_source_length, padding=padding, truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets, max_length=max_target_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
with accelerator.main_process_first():
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 1):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
label_pad_token_id = -100 if args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
)
eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_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 = 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
)
# 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,
)
# Metric
metric = load_metric("rouge")
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_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)
completed_steps = 0
for epoch in range(args.num_train_epochs):
model.train()
for step, batch in enumerate(train_dataloader):
outputs = model(**batch)
loss = outputs.loss
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if completed_steps >= args.max_train_steps:
break
model.eval()
if args.val_max_target_length is None:
args.val_max_target_length = args.max_target_length
gen_kwargs = {
"max_length": args.val_max_target_length if args is not None else config.max_length,
"num_beams": args.num_beams,
}
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
generated_tokens = accelerator.unwrap_model(model).generate(
batch["input_ids"],
attention_mask=batch["attention_mask"],
**gen_kwargs,
)
generated_tokens = accelerator.pad_across_processes(
generated_tokens, dim=1, pad_index=tokenizer.pad_token_id
)
labels = batch["labels"]
if not args.pad_to_max_length:
# If we did not pad to max length, we need to pad the labels too
labels = accelerator.pad_across_processes(batch["labels"], dim=1, pad_index=tokenizer.pad_token_id)
generated_tokens = accelerator.gather(generated_tokens).cpu().numpy()
labels = accelerator.gather(labels).cpu().numpy()
if args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
if isinstance(generated_tokens, tuple):
generated_tokens = generated_tokens[0]
decoded_preds = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
metric.add_batch(predictions=decoded_preds, references=decoded_labels)
result = metric.compute(use_stemmer=True)
# Extract a few results from ROUGE
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
result = {k: round(v, 4) for k, v in result.items()}
logger.info(result)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True
)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)
def _setup_dataloader_from_config(self, cfg: DictConfig, data_split: str):
logging.info(f"Creating {data_split} dataset")
shuffle = cfg["shuffle"]
if cfg.get("use_tarred_dataset", False):
logging.info('Tarred dataset')
metadata_file = cfg["tar_metadata_file"]
if metadata_file is None or not os.path.exists(metadata_file):
raise FileNotFoundError(
f"Trying to use tarred dataset but could not find {metadata_file}."
)
with open(metadata_file, "r") as f:
metadata = json.load(f)
num_batches = metadata["num_batches"]
tar_files = os.path.join(os.path.dirname(metadata_file),
metadata["text_tar_filepaths"])
logging.info(f"Loading {tar_files}")
dataset = TarredTextNormalizationDecoderDataset(
text_tar_filepaths=tar_files,
num_batches=num_batches,
shuffle_n=cfg.get("tar_shuffle_n", 4 *
cfg['batch_size']) if shuffle else 0,
shard_strategy=cfg.get("shard_strategy", "scatter"),
global_rank=self.global_rank,
world_size=self.world_size,
)
dl = torch.utils.data.DataLoader(
dataset=dataset,
batch_size=1,
sampler=None,
num_workers=cfg.get("num_workers", 2),
pin_memory=cfg.get("pin_memory", False),
drop_last=cfg.get("drop_last", False),
)
else:
input_file = cfg.data_path
if not os.path.exists(input_file):
raise ValueError(f"{input_file} not found.")
dataset = TextNormalizationDecoderDataset(
input_file=input_file,
tokenizer=self._tokenizer,
tokenizer_name=self.transformer_name,
mode=self.mode,
max_len=self.max_sequence_len,
decoder_data_augmentation=cfg.get('decoder_data_augmentation',
False)
if data_split == "train" else False,
lang=self.lang,
do_basic_tokenize=cfg.do_basic_tokenize,
use_cache=cfg.get('use_cache', False),
max_insts=cfg.get('max_insts', -1),
do_tokenize=True,
)
# create and save class names to class_ids mapping for validation
# (each validation set might have different classes)
if data_split in ['val', 'test']:
if not hasattr(self, "_val_class_to_id"):
self._val_class_to_id = []
self._val_id_to_class = []
self._val_class_to_id.append(dataset.label_ids_semiotic)
self._val_id_to_class.append(
{v: k
for k, v in dataset.label_ids_semiotic.items()})
data_collator = DataCollatorForSeq2Seq(
self._tokenizer,
model=self.model,
label_pad_token_id=constants.LABEL_PAD_TOKEN_ID,
padding=True)
dl = torch.utils.data.DataLoader(
dataset=dataset,
batch_size=cfg.batch_size,
shuffle=shuffle,
collate_fn=data_collator,
num_workers=cfg.get("num_workers", 3),
pin_memory=cfg.get("pin_memory", False),
drop_last=cfg.get("drop_last", False),
)
return dataset, dl
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For JSON files, this script will use the `question` column for the input question and `table` column for the corresponding table.
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# IMPORTANT: the initial BART model's decoding is penalized by no_repeat_ngram_size, and thus
# we should disable it here to avoid problematic generation
config.no_repeat_ngram_size = 0
config.max_length = 1024
config.early_stopping = False
# load tapex tokenizer
tokenizer = TapexTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
add_prefix_space=True,
)
# load Bart based Tapex model (default tapex-large)
model = BartForConditionalGeneration.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_tableqa_function(examples, is_training=False):
"""
The is_training FLAG is used to identify if we could use the supervision
to truncate the table content if it is required.
"""
# this function is specific for WikiSQL since the util function need the data structure
# to retrieve the WikiSQL answer for each question
def _convert_table_types(_table):
"""Runs the type converter over the table cells."""
ret_table = deepcopy(_table)
types = ret_table["types"]
ret_table["real_rows"] = ret_table["rows"]
typed_rows = []
for row in ret_table["rows"]:
typed_row = []
for column, cell_value in enumerate(row):
typed_row.append(
_TYPE_CONVERTER[types[column]](cell_value))
typed_rows.append(typed_row)
ret_table["rows"] = typed_rows
return ret_table
questions = [question.lower() for question in examples["question"]]
example_tables = examples["table"]
example_sqls = examples["sql"]
tables = [
pd.DataFrame.from_records(example_table["rows"],
columns=example_table["header"])
for example_table in example_tables
]
# using tapas utils to obtain wikisql answer
answers = []
for example_sql, example_table in zip(example_sqls, example_tables):
tapas_table = _convert_table_types(example_table)
answer_list: List[str] = retrieve_wikisql_query_answer_tapas(
tapas_table, example_sql)
# you can choose other delimiters to split each answer
answers.append(answer_list)
# IMPORTANT: we cannot pass by answers during evaluation, answers passed during training are used to
# truncate large tables in the train set!
if is_training:
model_inputs = tokenizer(
table=tables,
query=questions,
answer=answers,
max_length=data_args.max_source_length,
padding=padding,
truncation=True,
)
else:
model_inputs = tokenizer(table=tables,
query=questions,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
with tokenizer.as_target_tokenizer():
labels = tokenizer(
answer=[", ".join(answer) for answer in answers],
max_length=max_target_length,
padding=padding,
truncation=True,
)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
# in training, we can use the answer as extra information to truncate large tables
preprocess_tableqa_function_training = partial(preprocess_tableqa_function,
is_training=True)
if training_args.do_train:
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_tableqa_function_training,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
eval_dataset = eval_dataset.map(
preprocess_tableqa_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = datasets["test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
predict_dataset = predict_dataset.map(
preprocess_tableqa_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
delimiter = ", "
# define example evaluation
def evaluate_example(predict_str: str, ground_str: str):
predict_spans = predict_str.split(delimiter)
ground_spans = ground_str.split(delimiter)
predict_values = defaultdict(lambda: 0)
ground_values = defaultdict(lambda: 0)
for span in predict_spans:
try:
predict_values[float(span)] += 1
except ValueError:
predict_values[span.strip()] += 1
for span in ground_spans:
try:
ground_values[float(span)] += 1
except ValueError:
ground_values[span.strip()] += 1
is_correct = predict_values == ground_values
return is_correct
def get_denotation_accuracy(predictions: List[str],
references: List[str]):
assert len(predictions) == len(references)
correct_num = 0
for predict_str, ground_str in zip(predictions, references):
is_correct = evaluate_example(predict_str.lower(),
ground_str.lower())
if is_correct:
correct_num += 1
return correct_num / len(predictions)
accuracy = get_denotation_accuracy(decoded_preds, decoded_labels)
result = {"denotation_accuracy": accuracy}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
predict_results = trainer.predict(
predict_dataset,
metric_key_prefix="predict",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = predict_results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
predict_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
predictions = [pred.strip() for pred in predictions]
output_prediction_file = os.path.join(training_args.output_dir,
"tapex_predictions.txt")
with open(output_prediction_file, "w") as writer:
writer.write("\n".join(predictions))
return results
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files in the summarization task, this script will use the first column for the full texts and the
# second column for the summaries (unless you specify column names for this with the `text_column` and
# `summary_column` arguments).
# For translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = MeanSumT5.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (
data_args.target_lang is not None and data_args.source_lang is not None
), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
return
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if data_args.task.startswith("translation") or isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if data_args.source_lang is not None:
tokenizer.src_lang = data_args.source_lang
if data_args.target_lang is not None:
tokenizer.tgt_lang = data_args.target_lang
# To serialize preprocess_function below, each of those four variables needs to be defined (even if we won't use
# them all).
source_lang, target_lang, text_column, summary_column = None, None, None, None
if data_args.task.startswith("summarization"):
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None)
if data_args.text_column is None:
text_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if data_args.summary_column is None:
summary_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
else:
# Get the language codes for input/target.
lang_search = re.match("translation_([a-z]+)_to_([a-z]+)", data_args.task)
if data_args.source_lang is not None:
source_lang = data_args.source_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a source language via --source_lang or rename your task 'translation_xx_to_yy'."
source_lang = lang_search.groups()[0]
if data_args.target_lang is not None:
target_lang = data_args.target_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a target language via --target_lang or rename your task 'translation_xx_to_yy'."
target_lang = lang_search.groups()[1]
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(model, "prepare_decoder_input_ids_from_labels"):
logger.warn(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
articles = examples[text_column]
article_sentences = [nltk.sent_tokenize(inp) for inp in articles]
sentence_input_ids = []
sentence_attention_masks = []
for sents in article_sentences:
cur_sents = []
split = len(sents) // 5
for i in range(5):
if i < 4:
cur_sents.append(' '.join(sents[i * split: (i + 1) * split]))
else:
cur_sents.append(' '.join(sents[i * split:]))
assert(len(cur_sents) == 5)
features = tokenizer(cur_sents, max_length=data_args.max_source_length, padding="max_length", truncation=True)
sentence_input_ids.append(features['input_ids'])
sentence_attention_masks.append(features['attention_mask'])
targets = examples['highlights']
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets, max_length=200, padding="max_length", truncation=True) #
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_input = dict(input_ids=sentence_input_ids,
attention_mask=sentence_attention_masks,
labels=labels['input_ids'])
return model_input
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
# model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
metric_name = "rouge" if data_args.task.startswith("summarization") else "sacrebleu"
metric = load_metric(metric_name)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
if metric_name == "rouge":
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
else: # sacrebleu
labels = [[label] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
print(decoded_preds[:5])
if metric_name == "rouge":
result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
# Extract a few results from ROUGE
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
else:
result = metric.compute(predictions=decoded_preds, references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
)
all_metrics = {}
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
if trainer.is_world_process_zero():
logger.info("***** train metrics *****")
for key in sorted(metrics.keys()):
logger.info(f" {key} = {metrics[key]}")
save_json(metrics, os.path.join(training_args.output_dir, "train_results.json"))
all_metrics.update(metrics)
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(
max_length=data_args.val_max_target_length, num_beams=data_args.num_beams, metric_key_prefix="val"
)
metrics = {k: round(v, 4) for k, v in metrics.items()}
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(eval_dataset)
metrics["val_samples"] = min(max_val_samples, len(eval_dataset))
if trainer.is_world_process_zero():
logger.info("***** val metrics *****")
for key in sorted(metrics.keys()):
logger.info(f" {key} = {metrics[key]}")
save_json(metrics, os.path.join(training_args.output_dir, "val_results.json"))
all_metrics.update(metrics)
if training_args.do_predict:
logger.info("*** Test ***")
test_results = trainer.predict(
test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = test_results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
metrics = {k: round(v, 4) for k, v in metrics.items()}
if trainer.is_world_process_zero():
logger.info("***** test metrics *****")
for key in sorted(metrics.keys()):
logger.info(f" {key} = {metrics[key]}")
save_json(metrics, os.path.join(training_args.output_dir, "test_results.json"))
all_metrics.update(metrics)
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
test_preds = [pred.strip() for pred in test_preds]
output_test_preds_file = os.path.join(training_args.output_dir, "test_preds_seq2seq.txt")
with open(output_test_preds_file, "w") as writer:
writer.write("\n".join(test_preds))
if trainer.is_world_process_zero():
save_json(all_metrics, os.path.join(training_args.output_dir, "all_results.json"))
return results
def create_trainer(self):
training_args = self.training_args
data_args = self.data_args
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else self.tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
self.tokenizer,
label_pad_token_id=label_pad_token_id,
)
# Metric
metric = load_metric("sacrebleu")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
def compute_metrics(p: EvalPrediction):
preds = p.predictions[0] if isinstance(p.predictions,
tuple) else p.predictions
labels = p.label_ids
decoded_preds = self.tokenizer.batch_decode(
preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels,
self.tokenizer.pad_token_id)
decoded_labels = self.tokenizer.batch_decode(
labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != self.tokenizer.pad_token_id)
for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
all_args = self.get_all_args(exclude_base=True)
# Initialize our Trainer
self.trainer = self.SEQ2SEQ_TRAINER_CLASS(
model=None,
args=training_args,
train_dataset=self.train_dataset
if training_args.do_train else None,
eval_dataset=self.validation_dataset
if training_args.do_eval else None,
tokenizer=self.tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
model_init=self.model_init,
**all_args,
)
self.trainer._max_length = data_args.max_target_length
self.trainer._num_beams = data_args.num_beams
def main():
parser = argparse.ArgumentParser()
# Input and output configs
parser.add_argument("--data_folder", default=None, type=str, required=True,
help="the folder to save the processed data")
parser.add_argument("--last_utterance_only", default=False, required=False, action="store_true",
help="Train with the whole context or the last utterance only")
args = parser.parse_args()
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
handlers=[
logging.StreamHandler()
]
)
tasks = ['mantis', 'msdialog', 'ubuntu_dstc8']
#Downloading Conversation Response Ranking
for task in tasks:
if not os.path.isdir(args.data_folder+task):
logging.info("Starting downloader for task {}".format(task))
dataDownloader = downloader.DataDownloader(task, args.data_folder)
dataDownloader.download_and_preprocess()
all_df = []
for task in tasks:
train = pd.read_csv(args.data_folder+task+"/train.tsv", sep="\t")
train['task'] = task
replace = train.shape[0]<80000
train = train.sample(80000, replace=replace)
all_df.append(train)
all_df = pd.concat(all_df)
def preprocess_response(r):
# some tokens that only appear in MSDialogue are removed here
r = r.replace("<<<AGENT>>>:", "")
r = r.replace("PERSON_PLACEHOLDER", "")
r = r.replace("AGENT", "")
return r
def preprocess_context(r):
#removes beginning of context and keeps only last utterance.
if 'msdialog' in r['task']:
context = r['context'].split("[TURN_SEP]")[-1].split("[UTTERANCE_SEP]")[0].strip()
else:
context = r['context'].split("[TURN_SEP]")[-1].split("[UTTERANCE_SEP]")[-2].strip()
return context
all_df["response"] = all_df.apply(lambda r,f=preprocess_response: f(r['response']), axis=1)
if args.last_utterance_only:
all_df["context"] = all_df.apply(lambda r,f=preprocess_context: f(r), axis=1)
dataset = Dataset.from_pandas(all_df)
# all_df["len_context"] = all_df.apply(lambda r: len(r['context'].split(" ")), axis=1)
# all_df["len_response"] = all_df.apply(lambda r: len(r['response'].split(" ")), axis=1)
model_checkpoint = "t5-base" #["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b"]
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
max_input_length = 100
if args.last_utterance_only:
max_target_length = 100
else:
max_target_length = 400
col_from = "response"
col_to = "context"
def preprocess_function(examples):
inputs = [preprocess_response(doc) for doc in examples[col_from]]
model_inputs = tokenizer(inputs, max_length=max_input_length, truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer([t for t in examples[col_to]], max_length=max_target_length, truncation=True)
model_inputs["labels"] = labels["input_ids"]
return model_inputs
tokenized_datasets = dataset.map(preprocess_function, batched=True)
model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint)
batch_size = 5
train_args = Seq2SeqTrainingArguments(
"response2context_lu_{}".format(args.last_utterance_only),
learning_rate=2e-5,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
weight_decay=0.01,
save_total_limit=3,
num_train_epochs=2,
predict_with_generate=True,
seed=42
)
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
trainer = Seq2SeqTrainer(
model,
train_args,
train_dataset=tokenized_datasets,
data_collator=data_collator,
tokenizer=tokenizer
)
print("Fine-tuning T5.")
trainer.train()
if args.last_utterance_only:
model.save_pretrained("{}/{}_response2context_last_utt_only".format(args.data_folder, model_checkpoint))
else:
model.save_pretrained("{}/{}_response2context".format(args.data_folder, model_checkpoint))
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files, field="data", cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=True,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
# Preprocessing the datasets.
# We need to generate and tokenize inputs and targets.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
elif training_args.do_predict:
column_names = raw_datasets["test"].column_names
else:
logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
return
# Get the column names for input/target.
dataset_columns = question_answering_column_name_mapping.get(data_args.dataset_name, None)
if data_args.question_column is None:
question_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
else:
question_column = data_args.question_column
if question_column not in column_names:
raise ValueError(
f"--question_column' value '{data_args.question_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.context_column is None:
context_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
else:
context_column = data_args.context_column
if context_column not in column_names:
raise ValueError(
f"--context_column' value '{data_args.context_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.answer_column is None:
answer_column = dataset_columns[2] if dataset_columns is not None else column_names[2]
else:
answer_column = data_args.answer_column
if answer_column not in column_names:
raise ValueError(
f"--answer_column' value '{data_args.answer_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_answer_length for training.
max_answer_length = data_args.max_answer_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
def preprocess_sqaud_batch(
examples,
question_column: str,
context_column: str,
answer_column: str,
) -> Tuple[List[str], List[str]]:
questions = examples[question_column]
contexts = examples[context_column]
answers = examples[answer_column]
def generate_input(_question, _context):
return " ".join(["question:", _question, "context:", _context])
inputs = [generate_input(question, context) for question, context in zip(questions, contexts)]
targets = [answer["text"][0] if len(answer["text"]) > 0 else "" for answer in answers]
return inputs, targets
def preprocess_function(examples):
inputs, targets = preprocess_sqaud_batch(examples, question_column, context_column, answer_column)
model_inputs = tokenizer(inputs, max_length=max_seq_length, padding=padding, truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets, max_length=max_answer_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if data_args.max_train_samples is not None:
# We will select sample from whole data if agument is specified
train_dataset = train_dataset.select(range(data_args.max_train_samples))
# Create train feature from dataset
with training_args.main_process_first(desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if data_args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
train_dataset = train_dataset.select(range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
# We will select sample from whole data
eval_examples = eval_examples.select(range(data_args.max_eval_samples))
# Validation Feature Creation
with training_args.main_process_first(desc="validation dataset map pre-processing"):
eval_dataset = eval_examples.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if data_args.max_eval_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
if training_args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if data_args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(range(data_args.max_predict_samples))
# Predict Feature Creation
with training_args.main_process_first(desc="prediction dataset map pre-processing"):
predict_dataset = predict_examples.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
if data_args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
predict_dataset = predict_dataset.select(range(data_args.max_predict_samples))
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Post-processing:
def postprocess_text(preds, labels):
preds = [" ".join(pred) for pred in preds]
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
metric = load_metric("rouge")
def compute_metrics(eval_preds: EvalPrediction):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = [tokenizer.batch_decode(pred, skip_special_tokens=True) for pred in preds]
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
# Extract a few results from ROUGE
result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
max_length = (
training_args.generation_max_length
if training_args.generation_max_length is not None
else data_args.val_max_answer_length
)
num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=max_length, num_beams=num_beams, metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Prediction
if training_args.do_predict:
logger.info("*** Predict ***")
results = trainer.predict(predict_dataset, predict_examples)
metrics = results.metrics
max_predict_samples = (
data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)
)
metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if training_args.push_to_hub:
kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "question-answering"}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
trainer.push_to_hub(**kwargs)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S")
logger.setLevel(logging.INFO)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files in the summarization task, this script will use the first column for the full texts and the
# second column for the summaries (unless you specify column names for this with the `text_column` and
# `summary_column` arguments).
# For translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (data_args.target_lang is not None and data_args.source_lang
is not None), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
dataset_columns = summarization_name_mapping.get(data_args.dataset_name,
None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
label_pad_token_id=label_pad_token_id,
)
# Metric
metric_name = "rouge" if data_args.task.startswith(
"summarization") else "sacrebleu"
metric = load_metric(metric_name)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
if metric_name == "rouge":
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
else: # sacrebleu
labels = [[label] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
if metric_name == "rouge":
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
else:
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
all_metrics = {}
# Training
if training_args.do_train:
train_result = trainer.train()
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
if trainer.is_world_process_zero():
logger.info("***** train metrics *****")
for key in sorted(metrics.keys()):
logger.info(f" {key} = {metrics[key]}")
save_json(
metrics,
os.path.join(training_args.output_dir, "train_results.json"))
all_metrics.update(metrics)
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="val")
metrics = {k: round(v, 4) for k, v in metrics.items()}
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["val_samples"] = min(max_val_samples, len(eval_dataset))
if trainer.is_world_process_zero():
logger.info("***** val metrics *****")
for key in sorted(metrics.keys()):
logger.info(f" {key} = {metrics[key]}")
save_json(
metrics,
os.path.join(training_args.output_dir, "val_results.json"))
all_metrics.update(metrics)
if training_args.do_predict:
logger.info("*** Test ***")
test_results = trainer.predict(
test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = test_results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(
test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
metrics = {k: round(v, 4) for k, v in metrics.items()}
if trainer.is_world_process_zero():
logger.info("***** test metrics *****")
for key in sorted(metrics.keys()):
logger.info(f" {key} = {metrics[key]}")
save_json(
metrics,
os.path.join(training_args.output_dir, "test_results.json"))
all_metrics.update(metrics)
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
test_preds = [pred.strip() for pred in test_preds]
output_test_preds_file = os.path.join(
training_args.output_dir, "test_preds_seq2seq.txt")
with open(output_test_preds_file, "w") as writer:
writer.write("\n".join(test_preds))
if trainer.is_world_process_zero():
save_json(all_metrics,
os.path.join(training_args.output_dir, "all_results.json"))
return results
def main():
# 查看SRC/Transformers/Training_args.py中的所有可能参数
# or by passing the --help flag to this script.
# 我们现在保留不同的args集，以便更干净地分离关注的参数。
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# 如果我们只向脚本传递一个参数，而且是一个json文件的路径，那么我们就来解析它以获得我们的参数。
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"你正在运行一个T5模型，但没有提供一个源前缀，这是必须的, e.g. with `--source_prefix 'summarize: ' `"
)
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# 获取数据集：你可以提供你自己的JSON训练和评估文件（见下文）。
# 或只是提供中心上的一个公共数据集的名称 at https://huggingface.co/datasets/
# （数据集将自动从数据集hub下载）。
#
# 对于翻译，只支持JSON文件，其中有一个名为 "translation"的字段，包含源语言和目标语言的两个key（除非你调整下面的内容）。
# 在分布式训练中，load_dataset函数保证只有一个本地进程可以同时下载数据集。
if data_args.dataset_name == 'csl':
datasets = load_dataset(path='data_csl/csl.py',
name='csl',
data_files={
'train': 'data_csl/train.json',
'validation': 'data_csl/dev.json',
'test': 'data_csl/test.json'
})
elif data_args.dataset_name is not None:
# 从hub下载并加载数据集。
if os.path.exists('data_script/' + data_args.dataset_name):
dataset_path = 'data_script/' + data_args.dataset_name + '.py'
else:
dataset_path = data_args.dataset_name
#如果手动把数据下载到本地，那么请设置忽略校验md5
datasets = load_dataset(dataset_path,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
ignore_verifications=True)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# 查看更多关于加载任何类型的标准或自定义数据集（从文件、python dict、pandas DataFrame等）的信息，请访问
# https://huggingface.co/docs/datasets/loading_datasets.html.
# 2. 加载预训练的模型和tokenizer
#
# 分布式训练：
# .from_pretrained方法保证只有一个本地进程可以同时下载模型和单词表。
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# 对数据集进行预处理。
# 我们需要对输入和目标进行tokenize。, column_names: ['article', 'highlights', 'id']
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info("没有什么可以做的。 请传入 `do_train`, `do_eval` and/or `do_predict`.")
return
# 获取输入/目标的列名称。, 我们只需要 dataset_columns: ('article', 'highlights')
dataset_columns = summarization_name_mapping.get(data_args.dataset_name,
None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
"""
数据预处理的函数
Args:
examples ():
Returns:
"""
# inputs 是一个批次的数据，是全文， targets是摘要数据，是一个批次的, prefix代表每个输入前都加前缀，代表要执行的任务: 'summarize: '
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
if os.path.exists('data_script/rouge.py'):
metric = load_metric("data_script/rouge.py")
else:
metric = load_metric("rouge")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
logger.info("*** 训练 ***")
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** 评估 ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="eval")
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** 测试 ***")
test_results = trainer.predict(
test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = test_results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(
test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
test_preds = [pred.strip() for pred in test_preds]
output_test_preds_file = os.path.join(training_args.output_dir,
"test_generations.txt")
with open(output_test_preds_file, "w") as writer:
writer.write("\n".join(test_preds))
return results
def main():
args = parse_args()
# this requires having preprocessed framenet
# using `frame.cli:preprocess-framenet
paths = data_paths(args.data)
dataset = load_dataset('json', data_files=paths)
metric = load_metric("rouge")
train_test = dataset["train"].train_test_split(test_size=0.1)
test_valid = train_test["test"].train_test_split(test_size=0.5)
datasets = DatasetDict({
"train": train_test["train"],
"test": test_valid["test"],
"valid": test_valid["train"]
})
tokenizer = AutoTokenizer.from_pretrained(args.model)
# the family of t5 models expect input sentences to be prefixed with `"summarize: "`
if args.model in ["t5-small", "t5-base", "t5-larg", "t5-3b", "t5-11b"]:
prefix = "summarize: "
else:
prefix = ""
# HuggingFace loves to use closures, I would prefer this
# be refactored into the library, but going this route for simplicity
def preprocess_function(examples):
"""Tokenize the data for Seq2Seq
Maps over all the examples in the dataset
to tokenize both the input framenet sentences
and the target frame definitions.
Args:
examples: samples in the dataset
"""
inputs = [prefix + sent for sent in examples["sentence"]]
model_inputs = tokenizer(inputs,
max_length=args.max_input_length,
truncation=True)
with tokenizer.as_target_tokenizer():
labels = tokenizer(examples["frame_definition"],
max_length=args.max_target_length,
truncation=True)
model_inputs["labels"] = labels["input_ids"]
return model_inputs
# again, with the closures - requires instance of the tokenizer
def compute_metrics(eval_pred):
predictions, labels = eval_pred
decoded_preds = tokenizer.batch_decode(predictions,
skip_special_tokens=True)
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Rouge expects a newline after each sentence
decoded_preds = [
"\n".join(nltk.sent_tokenize(pred.strip()))
for pred in decoded_preds
]
decoded_labels = [
"\n".join(nltk.sent_tokenize(label.strip()))
for label in decoded_labels
]
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
# Add mean generated length
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id)
for pred in predictions
]
result["gen_len"] = np.mean(prediction_lens)
return {k: round(v, 4) for k, v in result.items()}
tokenized_datasets = datasets.map(preprocess_function, batched=True)
model = AutoModelForSeq2SeqLM.from_pretrained(args.model)
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
training_args = Seq2SeqTrainingArguments(
"./results/summarization",
evaluation_strategy="epoch",
learning_rate=2e-5,
per_device_train_batch_size=args.batch_size,
per_device_eval_batch_size=args.batch_size,
weight_decay=0.01,
save_total_limit=3,
num_train_epochs=args.epochs,
predict_with_generate=True,
fp16=True,
)
trainer = Seq2SeqTrainer(model,
training_args,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["valid"],
data_collator=data_collator,
tokenizer=tokenizer,
compute_metrics=compute_metrics)
trainer.train()
# Train
batch_size = args.batch_size
trainer_args = Seq2SeqTrainingArguments(
args.checkpoint_path,
evaluation_strategy="steps",
learning_rate=args.learning_rate,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
weight_decay=args.weight_decay,
save_total_limit=3,
num_train_epochs=args.num_train_epochs,
predict_with_generate=True,
fp16=True,
)
data_collator = DataCollatorForSeq2Seq(m2m_tokenizer, model=fused_model)
trainer = Seq2SeqTrainer(
fused_model,
trainer_args,
train_dataset=tokenized_datasets['train'],
eval_dataset=tokenized_datasets['validation'],
data_collator=data_collator,
tokenizer=m2m_tokenizer,
)
if args.do_train:
trainer.train()
if args.do_eval:
fused_model.eval()
print(trainer.evaluate())
def main():
# region Argument parsing
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TFTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_translation",
model_args,
data_args,
framework="tensorflow")
# endregion
# region Logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO)
datasets.utils.logging.set_verbosity(logging.INFO)
transformers.utils.logging.set_verbosity(logging.INFO)
# Log on each process the small summary:
logger.info(f"Training/evaluation parameters {training_args}")
# endregion
# region Detecting last checkpoint
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# endregion
# Set seed before initializing model.
set_seed(training_args.seed)
# region Load datasets
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files this script will use the first column for the full texts and the second column for the
# summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# endregion
# region Load model config and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# endregion
# region Dataset preprocessing
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, and/or `do_eval`."
)
return
column_names = raw_datasets["train"].column_names
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
assert data_args.target_lang is not None and data_args.source_lang is not None, (
f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --source_lang and "
"--target_lang arguments.")
tokenizer.src_lang = data_args.source_lang
tokenizer.tgt_lang = data_args.target_lang
forced_bos_token_id = (
tokenizer.lang_code_to_id[data_args.forced_bos_token]
if data_args.forced_bos_token is not None else None)
# Get the language codes for input/target.
source_lang = data_args.source_lang.split("_")[0]
target_lang = data_args.target_lang.split("_")[0]
padding = "max_length" if data_args.pad_to_max_length else False
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
def preprocess_function(examples):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Tokenize targets with the `text_target` keyword argument
labels = tokenizer(text_target=targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if data_args.max_train_samples is not None:
max_train_samples = min(len(train_dataset),
data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
with training_args.main_process_first(
desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
else:
train_dataset = None
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
max_eval_samples = min(len(eval_dataset),
data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
with training_args.main_process_first(
desc="validation dataset map pre-processing"):
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
else:
eval_dataset = None
# endregion
with training_args.strategy.scope():
# region Prepare model
model = TFAutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model.resize_token_embeddings(len(tokenizer))
if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
model.config.forced_bos_token_id = forced_bos_token_id
# endregion
# region Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (data_args.target_lang is not None
and data_args.source_lang is not None
), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
# endregion
# region Prepare TF Dataset objects
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=
64, # Reduce the number of unique shapes for XLA, especially for generation
return_tensors="tf",
)
num_replicas = training_args.strategy.num_replicas_in_sync
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas
dataset_options = tf.data.Options()
dataset_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
# model.prepare_tf_dataset() wraps a Hugging Face dataset in a tf.data.Dataset which is ready to use in
# training. This is the recommended way to use a Hugging Face dataset when training with Keras. You can also
# use the lower-level dataset.to_tf_dataset() method, but you will have to specify things like column names
# yourself if you use this method, whereas they are automatically inferred from the model input names when
# using model.prepare_tf_dataset()
# For more info see the docs:
# https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.TFPreTrainedModel.prepare_tf_dataset
# https://huggingface.co/docs/datasets/main/en/package_reference/main_classes#datasets.Dataset.to_tf_dataset
tf_train_dataset = model.prepare_tf_dataset(
train_dataset,
collate_fn=data_collator,
batch_size=total_train_batch_size,
shuffle=True,
).with_options(dataset_options)
tf_eval_dataset = model.prepare_tf_dataset(
eval_dataset,
collate_fn=data_collator,
batch_size=total_eval_batch_size,
shuffle=False).with_options(dataset_options)
# endregion
# region Optimizer and LR scheduling
num_train_steps = int(
len(tf_train_dataset) * training_args.num_train_epochs)
if training_args.warmup_steps > 0:
num_warmup_steps = training_args.warmup_steps
elif training_args.warmup_ratio > 0:
num_warmup_steps = int(num_train_steps *
training_args.warmup_ratio)
else:
num_warmup_steps = 0
if training_args.do_train:
optimizer, lr_schedule = create_optimizer(
init_lr=training_args.learning_rate,
num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps,
adam_beta1=training_args.adam_beta1,
adam_beta2=training_args.adam_beta2,
adam_epsilon=training_args.adam_epsilon,
weight_decay_rate=training_args.weight_decay,
adam_global_clipnorm=training_args.max_grad_norm,
)
else:
optimizer = None
# endregion
# region Metric and postprocessing
if training_args.do_eval:
metric = evaluate.load("sacrebleu")
if data_args.val_max_target_length is None:
data_args.val_max_target_length = data_args.max_target_length
gen_kwargs = {
"max_length": data_args.val_max_target_length,
"num_beams": data_args.num_beams,
"no_repeat_ngram_size":
0, # Not supported under XLA right now, and some models set it by default
}
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
def compute_metrics(preds):
predictions, labels = preds
if isinstance(predictions, tuple):
predictions = predictions[0]
decoded_preds = tokenizer.batch_decode(
predictions, skip_special_tokens=True)
labels = np.where(labels != -100, labels,
tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(
labels, skip_special_tokens=True)
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
metrics = metric.compute(predictions=decoded_preds,
references=decoded_labels)
return {"bleu": metrics["score"]}
# The KerasMetricCallback allows metrics that are too complex to write as standard Keras metrics
# to be computed each epoch. Any Python code can be included in the metric_fn. This is especially
# useful for metrics like BLEU and ROUGE that perform string comparisons on decoded model outputs.
# For more information, see the docs at
# https://huggingface.co/docs/transformers/main_classes/keras_callbacks#transformers.KerasMetricCallback
metric_callback = KerasMetricCallback(
metric_fn=compute_metrics,
eval_dataset=tf_eval_dataset,
predict_with_generate=True,
use_xla_generation=True,
generate_kwargs=gen_kwargs,
)
callbacks = [metric_callback]
else:
callbacks = []
# endregion
# region Preparing push_to_hub and model card
push_to_hub_model_id = training_args.push_to_hub_model_id
model_name = model_args.model_name_or_path.split("/")[-1]
if not push_to_hub_model_id:
push_to_hub_model_id = f"{model_name}-finetuned-{data_args.source_lang}-{data_args.target_lang}"
model_card_kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "translation"
}
if data_args.dataset_name is not None:
model_card_kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
model_card_kwargs[
"dataset_args"] = data_args.dataset_config_name
model_card_kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
model_card_kwargs["dataset"] = data_args.dataset_name
languages = [
l for l in [data_args.source_lang, data_args.target_lang]
if l is not None
]
if len(languages) > 0:
model_card_kwargs["language"] = languages
if training_args.push_to_hub:
# Because this training can be quite long, we save once per epoch.
callbacks.append(
PushToHubCallback(
output_dir=training_args.output_dir,
model_id=push_to_hub_model_id,
organization=training_args.push_to_hub_organization,
token=training_args.push_to_hub_token,
tokenizer=tokenizer,
**model_card_kwargs,
))
# endregion
# region Training
eval_metrics = None
model.compile(optimizer=optimizer, jit_compile=training_args.xla)
if training_args.do_train:
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {training_args.num_train_epochs}")
logger.info(
f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}"
)
logger.info(f" Total train batch size = {total_train_batch_size}")
logger.info(f" Total optimization steps = {num_train_steps}")
if training_args.xla and not data_args.pad_to_max_length:
logger.warning(
"XLA training may be slow at first when --pad_to_max_length is not set "
"until all possible shapes have been compiled.")
history = model.fit(tf_train_dataset,
epochs=int(training_args.num_train_epochs),
callbacks=callbacks)
eval_metrics = {
key: val[-1]
for key, val in history.history.items()
}
# endregion
# region Validation
if training_args.do_eval and not training_args.do_train:
# Compiling generation with XLA yields enormous speedups, see https://huggingface.co/blog/tf-xla-generate
@tf.function(jit_compile=True)
def generate(**kwargs):
return model.generate(**kwargs)
if training_args.do_eval:
logger.info("Evaluation...")
for batch, labels in tf_eval_dataset:
batch.update(gen_kwargs)
generated_tokens = generate(**batch)
if isinstance(generated_tokens, tuple):
generated_tokens = generated_tokens[0]
decoded_preds = tokenizer.batch_decode(
generated_tokens, skip_special_tokens=True)
labels = np.where(labels != -100, labels,
tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(
labels, skip_special_tokens=True)
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
metric.add_batch(predictions=decoded_preds,
references=decoded_labels)
eval_metrics = metric.compute()
logger.info({"bleu": eval_metrics["score"]})
# endregion
if training_args.output_dir is not None and eval_metrics is not None:
output_eval_file = os.path.join(training_args.output_dir,
"all_results.json")
with open(output_eval_file, "w") as writer:
writer.write(json.dumps(eval_metrics))
if training_args.output_dir is not None and not training_args.push_to_hub:
# If we're not pushing to hub, at least save a local copy when we're done
model.save_pretrained(training_args.output_dir)
def main():
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
data_files = dict(train=data_args.train_file, validation=data_args.validation_file)
datasets = load_dataset('csv', data_files=data_files, cache_dir=model_args.cache_dir,
delimiter='\t', column_names=['img_id', 'graph', '_', 'text']) # _ is unimportant
model, tokenizer = load_model_and_tokenizer(model_args)
prefix = "translate graph to text"
column_names = datasets["train"].column_names
padding = "max_length" if data_args.pad_to_max_length else False
def preprocess_function(examples):
inputs = [ex for ex in examples['graph']]
targets = [ex for ex in examples['text']]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets, max_length=data_args.max_target_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
#train_dataset = train_dataset.select(range(1))
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
#eval_dataset = eval_dataset.select(range(1))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
metric = load_metric('sacrebleu')
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
labels = [[label] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds, references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# this is the recommended t5 finetuning setup from
# https://huggingface.co/transformers/main_classes/optimizer_schedules.html#adafactor-pytorch
optimizer = Adafactor(
model.parameters(),
lr=3e-4,
eps=(1e-30, 1e-3),
clip_threshold=1.0,
decay_rate=-0.8,
beta1=None,
weight_decay=0.0,
relative_step=False,
scale_parameter=False,
warmup_init=False)
"""
optimizer = AdamW(lr=3e-5)
"""
lr_scheduler = transformers.get_constant_schedule(optimizer)
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
optimizers=(optimizer, lr_scheduler)
)
if training_args.do_train:
train_result = trainer.train()
trainer.save_model()
metrics = train_result.metrics
max_train_samples = len(train_dataset)
metrics["train_samples"] = len(train_dataset)
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
if training_args.do_eval:
metrics = trainer.evaluate(
max_length=data_args.max_target_length, num_beams=data_args.num_beams, metric_key_prefix="eval"
)
max_val_samples = len(eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
"`--source_prefix 'summarize: ' `")
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info(f"Training/evaluation parameters {training_args}")
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files this script will use the first column for the full texts and the second column for the
# summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(data_args.dataset_name,
None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if text_column not in column_names:
raise ValueError(
f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
)
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = datasets["test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
predict_dataset = predict_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
metric = load_metric("rouge")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
predict_results = trainer.predict(
predict_dataset,
metric_key_prefix="predict",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = predict_results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
predict_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
predictions = [pred.strip() for pred in predictions]
output_prediction_file = os.path.join(
training_args.output_dir, "generated_predictions.txt")
with open(output_prediction_file, "w") as writer:
writer.write("\n".join(predictions))
if training_args.push_to_hub:
trainer.push_to_hub()
return results
def create_trainer(self):
# Data collator
training_args = self.training_args
data_args = self.data_args
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else self.tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
self.tokenizer,
label_pad_token_id=label_pad_token_id,
)
# Metric
metric = load_metric("rouge")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
def compute_metrics(p: EvalPrediction):
preds = p.predictions[0] if isinstance(p.predictions,
tuple) else p.predictions
labels = p.label_ids
decoded_preds = self.tokenizer.batch_decode(
preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels,
self.tokenizer.pad_token_id)
decoded_labels = self.tokenizer.batch_decode(
labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels,
use_stemmer=True)
# Extract a few results from ROUGE
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
prediction_lens = [
np.count_nonzero(pred != self.tokenizer.pad_token_id)
for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
all_args = self.get_all_args(exclude_base=True)
# Initialize our Trainer
self.trainer = self.SEQ2SEQ_TRAINER_CLASS(
model=None,
args=training_args,
train_dataset=self.train_dataset
if training_args.do_train else None,
eval_dataset=self.validation_dataset
if training_args.do_eval else None,
tokenizer=self.tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
model_init=self.model_init,
**all_args,
)
self.trainer._max_length = data_args.max_target_length
self.trainer._num_beams = data_args.num_beams
def main():
args = parse_args()
if args.source_prefix is None and args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
"`--source_prefix 'summarize: ' `")
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
accelerator = Accelerator()
# 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()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name,
args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.model_name_or_path)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning(
"You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(
args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(
args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if args.model_name_or_path:
model = AutoModelForSeq2SeqLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForSeq2SeqLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = args.source_prefix if args.source_prefix is not None else ""
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(args.dataset_name, None)
text_column_name = dataset_columns[
0] if dataset_columns is not None else column_names[0]
padding = "max_length" if args.pad_to_max_length else False
if args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = args.summary_column
if summary_column not in column_names:
raise ValueError(
f"--summary_column' value '{args.summary_column}' needs to be one of: {', '.join(column_names)}"
)
# Temporarily set max_target_length for training.
max_target_length = args.max_target_length
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
inputs = examples[text_column_name]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 1):
logger.info(
f"Sample {index} of the training set: {train_dataset[index]}.")
label_pad_token_id = -100 if args.ignore_pad_token_for_loss else tokenizer.pad_token_id
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
)
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [label.strip() for label in labels]
# rougeLSum expects newline after each sentence
preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]
return preds, labels
train_dataloader = DataLoader(train_dataset,
shuffle=True,
collate_fn=data_collator,
batch_size=args.per_device_train_batch_size)
eval_dataloader = DataLoader(eval_dataset,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
import deepspeed
model_engine, optimizer, train_dataloader, eval_dataloader = deepspeed.initialize(
args=cmd_args, model, model.parameters())
# 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.
# Metric
# Only show the progress bar once on each machine.
for epoch in range(args.num_train_epochs):
for step, batch in enumerate(train_dataloader):
outputs = model_engine(**batch)
loss = outputs.loss
model_engine.backward(loss)
model_engine.step()
print("OK")
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"You're running a t5 model but didn't provide a source prefix, which is expected, e.g. with "
"`--source_prefix 'translate English to German: ' `")
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
cache_dir=model_args.cache_dir)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model.resize_token_embeddings(len(tokenizer))
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
elif training_args.do_predict:
column_names = raw_datasets["test"].column_names
else:
logger.info(
"There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
)
return
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
assert data_args.target_lang is not None and data_args.source_lang is not None, (
f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --source_lang and "
"--target_lang arguments.")
tokenizer.src_lang = data_args.source_lang
tokenizer.tgt_lang = data_args.target_lang
# For multilingual translation models like mBART-50 and M2M100 we need to force the target language token
# as the first generated token. We ask the user to explicitly provide this as --forced_bos_token argument.
forced_bos_token_id = (
tokenizer.lang_code_to_id[data_args.forced_bos_token]
if data_args.forced_bos_token is not None else None)
model.config.forced_bos_token_id = forced_bos_token_id
# Get the language codes for input/target.
source_lang = data_args.source_lang.split("_")[0]
target_lang = data_args.target_lang.split("_")[0]
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
if training_args.label_smoothing_factor > 0 and not hasattr(
model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
with training_args.main_process_first(
desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
eval_dataset = eval_dataset.select(
range(data_args.max_eval_samples))
with training_args.main_process_first(
desc="validation dataset map pre-processing"):
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_dataset = raw_datasets["test"]
if data_args.max_predict_samples is not None:
predict_dataset = predict_dataset.select(
range(data_args.max_predict_samples))
with training_args.main_process_first(
desc="prediction dataset map pre-processing"):
predict_dataset = predict_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# Metric
metric = load_metric("sacrebleu")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(
decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
max_length = (training_args.generation_max_length
if training_args.generation_max_length is not None else
data_args.val_max_target_length)
num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(max_length=max_length,
num_beams=num_beams,
metric_key_prefix="eval")
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
predict_results = trainer.predict(predict_dataset,
metric_key_prefix="predict",
max_length=max_length,
num_beams=num_beams)
metrics = predict_results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
predictions = tokenizer.batch_decode(
predict_results.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
predictions = [pred.strip() for pred in predictions]
output_prediction_file = os.path.join(
training_args.output_dir, "generated_predictions.txt")
with open(output_prediction_file, "w",
encoding="utf-8") as writer:
writer.write("\n".join(predictions))
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "translation"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
languages = [
l for l in [data_args.source_lang, data_args.target_lang]
if l is not None
]
if len(languages) > 0:
kwargs["language"] = languages
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
def main(args):
df = pd.read_csv(args.input_fname,
encoding='utf-8')[[args.source_lang, args.target_lang]]
logging.info(f'Loaded {df.shape}')
#convert to dictionary
j = {'translation': []}
for i in df.itertuples():
j['translation'] += [{args.source_lang: i[1], args.target_lang: i[2]}]
train_dataset = Dataset.from_dict(j)
raw_datasets = train_dataset.train_test_split(test_size=args.valid_pct,
seed=args.seed)
logging.info(f'Datasets created {raw_datasets}')
tokenizer = MarianTokenizer.from_pretrained(args.output_dir)
logging.info(f'Tokenizer loaded from {args.output_dir}')
#tokenize datasets
tokenized_datasets = raw_datasets.map(
partial(preprocess_function,
tokenizer=tokenizer,
max_input_length=args.max_input_length,
max_target_length=args.max_target_length,
source_lang=args.source_lang,
target_lang=args.target_lang),
batched=True,
)
logging.info(f'Tokenized datasets: {tokenized_datasets}')
#filter those with too few tokens
tokenized_datasets = tokenized_datasets.filter(
lambda example: len(example['translation']['zh']) > 2)
tokenized_datasets = tokenized_datasets.filter(
lambda example: len(example['translation']['th']) > 2)
logging.info(
f'Tokenized datasets when filtered out less than 2 tokens per sequence: {tokenized_datasets}'
)
config = MarianConfig.from_pretrained(args.output_dir)
model = MarianMTModel(config)
logging.info(f'Loaded model from {args.output_dir}')
training_args = Seq2SeqTrainingArguments(
args.output_dir,
evaluation_strategy="epoch",
load_best_model_at_end=True,
learning_rate=args.learning_rate,
warmup_ratio=args.warmup_ratio,
per_device_train_batch_size=args.batch_size,
per_device_eval_batch_size=args.batch_size,
weight_decay=args.weight_decay,
save_total_limit=args.save_total_limit,
num_train_epochs=args.num_train_epochs,
predict_with_generate=True,
fp16=args.fp16,
seed=args.seed,
)
logging.info(f'Training congig {training_args}')
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
trainer = Seq2SeqTrainer(
model,
training_args,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["test"],
data_collator=data_collator,
tokenizer=tokenizer,
compute_metrics=partial(compute_metrics,
tokenizer=tokenizer,
metric=metric,
metric_tokenize=args.metric_tokenize),
)
logging.info(f'Trainer created')
trainer.train()
model.save_pretrained(f"{args.output_dir}_best")
tokenizer.save_pretrained(f"{args.output_dir}_best")
logging.info(f'Best model saved')
model.cpu()
src_text = ['我爱你', '国王有很多心事。我明白']
translated = model.generate(
**tokenizer(src_text, return_tensors="pt", padding=True))
print([tokenizer.decode(t, skip_special_tokens=True) for t in translated])
def main():
# Parse the arguments
args = parse_args()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_translation_no_trainer", args)
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
# If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers
# in the environment
accelerator = (
Accelerator(log_with=args.report_to, logging_dir=args.output_dir) if args.with_tracking else Accelerator()
)
# 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, main_process_only=False)
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()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
else:
repo_name = args.hub_model_id
repo = Repository(args.output_dir, clone_from=repo_name)
with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
if "step_*" not in gitignore:
gitignore.write("step_*\n")
if "epoch_*" not in gitignore:
gitignore.write("epoch_*\n")
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
else:
data_files = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
if args.model_name_or_path:
model = AutoModelForSeq2SeqLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForSeq2SeqLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
assert (
args.target_lang is not None and args.source_lang is not None
), "mBart requires --target_lang and --source_lang"
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
prefix = args.source_prefix if args.source_prefix is not None else ""
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if args.source_lang is not None:
tokenizer.src_lang = args.source_lang
if args.target_lang is not None:
tokenizer.tgt_lang = args.target_lang
# Get the language codes for input/target.
source_lang = args.source_lang.split("_")[0]
target_lang = args.target_lang.split("_")[0]
padding = "max_length" if args.pad_to_max_length else False
# Temporarily set max_target_length for training.
max_target_length = args.max_target_length
padding = "max_length" if args.pad_to_max_length else False
def preprocess_function(examples):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs, max_length=args.max_source_length, padding=padding, truncation=True)
# Tokenize targets with the `text_target` keyword argument
labels = tokenizer(text_target=targets, max_length=max_target_length, padding=padding, truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
with accelerator.main_process_first():
processed_datasets = raw_datasets.map(
preprocess_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
train_dataset = processed_datasets["train"]
eval_dataset = processed_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
label_pad_token_id = -100 if args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if args.pad_to_max_length:
# If padding was already done ot max length, we use the default data collator that will just convert everything
# to tensors.
data_collator = default_data_collator
else:
# Otherwise, `DataCollatorWithPadding` will apply dynamic padding for us (by padding to the maximum length of
# the samples passed). When using mixed precision, we add `pad_to_multiple_of=8` to pad all tensors to multiple
# of 8s, which will enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if accelerator.use_fp16 else None,
)
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
)
eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)
# Optimizer
# Split weights in two groups, one with weight decay and the other not.
no_decay = ["bias", "LayerNorm.weight", "layer_norm.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)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
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
overrode_max_train_steps = True
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,
)
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# Figure out how many steps we should save the Accelerator states
if hasattr(args.checkpointing_steps, "isdigit"):
checkpointing_steps = args.checkpointing_steps
if args.checkpointing_steps.isdigit():
checkpointing_steps = int(args.checkpointing_steps)
else:
checkpointing_steps = None
# We need to initialize the trackers we use, and also store our configuration.
# We initialize the trackers only on main process because `accelerator.log`
# only logs on main process and we don't want empty logs/runs on other processes.
if args.with_tracking:
if accelerator.is_main_process:
experiment_config = vars(args)
# TensorBoard cannot log Enums, need the raw value
experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
accelerator.init_trackers("translation_no_trainer", experiment_config)
metric = evaluate.load("sacrebleu")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_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)
completed_steps = 0
starting_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "":
accelerator.print(f"Resumed from checkpoint: {args.resume_from_checkpoint}")
accelerator.load_state(args.resume_from_checkpoint)
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()]
dirs.sort(key=os.path.getctime)
path = dirs[-1] # Sorts folders by date modified, most recent checkpoint is the last
# Extract `epoch_{i}` or `step_{i}`
training_difference = os.path.splitext(path)[0]
if "epoch" in training_difference:
starting_epoch = int(training_difference.replace("epoch_", "")) + 1
resume_step = None
else:
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
# update the progress_bar if load from checkpoint
progress_bar.update(starting_epoch * num_update_steps_per_epoch)
completed_steps = starting_epoch * num_update_steps_per_epoch
for epoch in range(starting_epoch, args.num_train_epochs):
model.train()
if args.with_tracking:
total_loss = 0
for step, batch in enumerate(train_dataloader):
# We need to skip steps until we reach the resumed step
if args.resume_from_checkpoint and epoch == starting_epoch:
if resume_step is not None and step < resume_step:
if step % args.gradient_accumulation_steps == 0:
progress_bar.update(1)
completed_steps += 1
continue
outputs = model(**batch)
loss = outputs.loss
# We keep track of the loss at each epoch
if args.with_tracking:
total_loss += loss.detach().float()
loss = loss / args.gradient_accumulation_steps
accelerator.backward(loss)
if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
completed_steps += 1
if isinstance(checkpointing_steps, int):
if completed_steps % checkpointing_steps == 0:
output_dir = f"step_{completed_steps }"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
accelerator.save_state(output_dir)
if completed_steps >= args.max_train_steps:
break
model.eval()
if args.val_max_target_length is None:
args.val_max_target_length = args.max_target_length
gen_kwargs = {
"max_length": args.val_max_target_length if args is not None else config.max_length,
"num_beams": args.num_beams,
}
samples_seen = 0
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
generated_tokens = accelerator.unwrap_model(model).generate(
batch["input_ids"],
attention_mask=batch["attention_mask"],
**gen_kwargs,
)
generated_tokens = accelerator.pad_across_processes(
generated_tokens, dim=1, pad_index=tokenizer.pad_token_id
)
labels = batch["labels"]
if not args.pad_to_max_length:
# If we did not pad to max length, we need to pad the labels too
labels = accelerator.pad_across_processes(batch["labels"], dim=1, pad_index=tokenizer.pad_token_id)
generated_tokens = accelerator.gather(generated_tokens).cpu().numpy()
labels = accelerator.gather(labels).cpu().numpy()
if args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_preds = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
# If we are in a multiprocess environment, the last batch has duplicates
if accelerator.num_processes > 1:
if step == len(eval_dataloader) - 1:
decoded_preds = decoded_preds[: len(eval_dataloader.dataset) - samples_seen]
decoded_labels = decoded_labels[: len(eval_dataloader.dataset) - samples_seen]
else:
samples_seen += len(decoded_labels)
metric.add_batch(predictions=decoded_preds, references=decoded_labels)
eval_metric = metric.compute()
logger.info({"bleu": eval_metric["score"]})
if args.with_tracking:
accelerator.log(
{
"bleu": eval_metric["score"],
"train_loss": total_loss.item() / len(train_dataloader),
"epoch": epoch,
"step": completed_steps,
},
step=completed_steps,
)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True
)
if args.checkpointing_steps == "epoch":
output_dir = f"epoch_{epoch}"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
accelerator.save_state(output_dir)
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if accelerator.is_main_process:
tokenizer.save_pretrained(args.output_dir)
if args.push_to_hub:
repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)
with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
json.dump({"eval_bleu": eval_metric["score"]}, f)
batch_size = 16
args = Seq2SeqTrainingArguments(
"test-summarization",
evaluation_strategy="epoch",
learning_rate=2e-5,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
weight_decay=0.01,
save_total_limit=3,
num_train_epochs=3,
predict_with_generate=True,
fp16=True,
)
data_collator = DataCollatorForSeq2Seq(tokenizer, model=model)
def compute_metrics(eval_pred):
predictions, labels = eval_pred
decoded_preds = tokenizer.batch_decode(predictions,
skip_special_tokens=True)
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Rouge expects a newline after each sentence
decoded_preds = [
"\n".join(nltk.sent_tokenize(pred.strip())) for pred in decoded_preds
]
decoded_labels = [
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
if (os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty."
"Use --overwrite_output_dir to overcome.")
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if is_main_process(training_args.local_rank) else logging.WARN,
)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files in the summarization task, this script will use the first column for the full texts and the
# second column for the summaries (unless you specify column names for this with the `text_column` and
# `summary_column` arguments).
# For translation, only JSON files are supported, with one field named "translation" containing two keys for the
# source and target languages (unless you adapt what follows).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = AutoConfig.from_pretrained(
model_args.config_name
if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# Set decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.target_lang]
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
# Get the default prefix if None is passed.
if data_args.source_prefix is None:
task_specific_params = model.config.task_specific_params
if task_specific_params is not None:
prefix = task_specific_params.get("prefix", "")
else:
prefix = ""
else:
prefix = data_args.source_prefix
# Preprocessing the datasets.
# We need to tokenize inputs and targets.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
# For translation we set the codes of our source and target languages (only useful for mBART, the others will
# ignore those attributes).
if data_args.task.startswith("translation"):
if data_args.source_lang is not None:
tokenizer.src_lang = data_args.source_lang
if data_args.target_lang is not None:
tokenizer.tgt_lang = data_args.target_lang
# To serialize preprocess_function below, each of those four variables needs to be defined (even if we won't use
# them all).
source_lang, target_lang, text_column, summary_column = None, None, None, None
if data_args.task.startswith("summarization"):
# Get the column names for input/target.
dataset_columns = summarization_name_mapping.get(
data_args.dataset_name, None)
if data_args.text_column is None:
text_column = dataset_columns[
0] if dataset_columns is not None else column_names[0]
else:
text_column = data_args.text_column
if data_args.summary_column is None:
summary_column = dataset_columns[
1] if dataset_columns is not None else column_names[1]
else:
summary_column = data_args.summary_column
else:
# Get the language codes for input/target.
lang_search = re.match("translation_([a-z]+)_to_([a-z]+)",
data_args.task)
if data_args.source_lang is not None:
source_lang = data_args.source_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a source language via --source_lang or rename your task 'translation_xx_to_yy'."
source_lang = lang_search.groups()[0]
if data_args.target_lang is not None:
target_lang = data_args.target_lang.split("_")[0]
else:
assert (
lang_search is not None
), "Provide a target language via --target_lang or rename your task 'translation_xx_to_yy'."
target_lang = lang_search.groups()[1]
# Temporarily set max_target_length for training.
max_target_length = data_args.max_target_length
padding = "max_length" if data_args.pad_to_max_length else False
def preprocess_function(examples):
if data_args.task.startswith("translation"):
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
else:
inputs = examples[text_column]
targets = examples[summary_column]
inputs = [prefix + inp for inp in inputs]
model_inputs = tokenizer(inputs,
max_length=data_args.max_source_length,
padding=padding,
truncation=True)
# Setup the tokenizer for targets
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets,
max_length=max_target_length,
padding=padding,
truncation=True)
# If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
# padding in the loss.
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [[
(l if l != tokenizer.pad_token_id else -100) for l in label
] for label in labels["input_ids"]]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer, label_pad_token_id=label_pad_token_id)
# Metric
metric_name = "rouge" if data_args.task.startswith(
"summarization") else "sacrebleu"
metric = load_metric(metric_name)
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels,
skip_special_tokens=True)
# Some simple post-processing
decoded_preds = [pred.strip() for pred in decoded_preds]
decoded_labels = [label.strip() for label in decoded_labels]
if metric_name == "sacrebleu":
decoded_labels = [[label] for label in decoded_labels]
result = metric.compute(predictions=decoded_preds,
references=decoded_labels)
# Extract a few results from ROUGE
if metric_name == "rouge":
result = {
key: value.mid.fmeasure * 100
for key, value in result.items()
}
else:
result = {"bleu": result["score"]}
prediction_lens = [
np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
]
result["gen_len"] = np.mean(prediction_lens)
return result
# Initialize our Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics
if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
train_result = trainer.train(
model_path=model_args.model_name_or_path if os.path.
isdir(model_args.model_name_or_path) else None)
trainer.save_model() # Saves the tokenizer too for easy upload
output_train_file = os.path.join(training_args.output_dir,
"train_results.txt")
if trainer.is_world_process_zero():
with open(output_train_file, "w") as writer:
logger.info("***** Train results *****")
for key, value in sorted(train_result.metrics.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
# Need to save the state, since Trainer.save_model saves only the tokenizer with the model
trainer.state.save_to_json(
os.path.join(training_args.output_dir, "trainer_state.json"))
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
results = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results_seq2seq.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in sorted(results.items()):
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def main():
# 查看SRC/Transformers/Training_args.py中的所有可能参数
# or by passing the --help flag to this script.
# 我们现在保留不同的args集，以便更干净地分离关注的参数。
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# 如果我们只向脚本传递一个参数，而且是一个json文件的路径, 让我们解析它以获得我们的参数。
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
if data_args.source_prefix is None and model_args.model_name_or_path in [
"t5-small",
"t5-base",
"t5-large",
"t5-3b",
"t5-11b",
]:
logger.warning(
"你正在运行一个t5模型，但没有提供一个源前缀，这是必须的的，例如用`--source_prefix 'translate English to German:'"
)
# 检测最后一个checkpoint。
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# 设置transformer logger的粗略程度为info（仅在主进程中）
if is_main_process(training_args.local_rank):
transformers.utils.logging.set_verbosity_info()
logger.info(f"训练/评估参数 {training_args}")
# 在初始化模型之前设置random种子。
set_seed(training_args.seed)
# 获取数据集：你可以提供你自己的JSON训练和评估文件（见下文）。
# 或只是提供中心上的一个公共数据集的名称 at https://huggingface.co/datasets/
# （数据集将自动从数据集hub下载）。
#
# 对于翻译，只支持JSON文件，其中有一个名为 "translation"的字段，包含源语言和目标语言的两个key（除非你调整下面的内容）。
# 在分布式训练中，load_dataset函数保证只有一个本地进程可以同时下载数据集。
if data_args.dataset_name == 'custom_zh_en':
datasets = load_dataset(path='data/custom_zh_en.py', name='custom_zh_en', data_files={'train': ['data/train.cn','data/train.en'], 'validation': ['data/dev.cn','data/dev.en'], 'test': ['data/test.cn','data/test.en']})
elif data_args.dataset_name is not None:
# 从hub下载并加载数据集。
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
# 查看更多关于加载任何类型的标准或自定义数据集（从文件、python dict、pandas DataFrame等）的信息，请访问
# https://huggingface.co/docs/datasets/loading_datasets.html.
# 2. 加载预训练的模型和tokenizer
#
# 分布式训练：
# .from_pretrained方法保证只有一个本地进程可以同时下载模型和单词表。
config = AutoConfig.from_pretrained(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
gradient_checkpointing = True,
use_cache = False
)
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
use_auth_token=True if model_args.use_auth_token else None,
)
# 设置解码的开始的token id decoder_start_token_id
if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
if isinstance(tokenizer, MBartTokenizer):
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[data_args.target_lang]
else:
model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(data_args.target_lang)
if model.config.decoder_start_token_id is None:
raise ValueError("确保正确定义了`config.decoder_start_token_id` ")
# T5模型使用
prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
# 预处理数据集。
# 我们需要tokenize输入和目标。 column_names: ['translation']
if training_args.do_train:
column_names = datasets["train"].column_names
elif training_args.do_eval:
column_names = datasets["validation"].column_names
elif training_args.do_predict:
column_names = datasets["test"].column_names
else:
logger.info("没有什么可以做的。 请传入`do_train`，`do_eval`和/或`do_predict`。 ")
return
# mBart: 对于翻译，我们设置源语言和目标语言的代码（只对mBART有用，其他的会忽略这些属性
if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
assert data_args.target_lang is not None and data_args.source_lang is not None, (
f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --source_lang and "
"--target_lang arguments."
)
# tokenizer.src_lang: 'en_XX' ; tokenizer.tgt_lang: 'ro_RO'
tokenizer.src_lang = data_args.source_lang
tokenizer.tgt_lang = data_args.target_lang
# 对于像 mBART-50 和 M2M100 这样的多语言翻译模型，我们需要强制目标语言token作为第一个生成的token。我们要求用户明确提供--forced_bos_token参数。
forced_bos_token_id = (
tokenizer.lang_code_to_id[data_args.forced_bos_token] if data_args.forced_bos_token is not None else None
)
model.config.forced_bos_token_id = forced_bos_token_id
# 获取输入/目标的语言code， source_lang：en, target_lang: ro
source_lang = data_args.source_lang.split("_")[0]
target_lang = data_args.target_lang.split("_")[0]
# 临时设置Max_target_Length进行训练。 max_target_length: 128
max_target_length = data_args.max_target_length
#是否padding到最长: False
padding = "max_length" if data_args.pad_to_max_length else False
#标签平滑因子
if training_args.label_smoothing_factor > 0 and not hasattr(model, "prepare_decoder_input_ids_from_labels"):
logger.warning(
"label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
)
def preprocess_function(examples):
"""
# 对数据进行预处理
Args:
examples (): {'translation': [{'en':xxx, 'ro':yyyy},....]} 一个translation里面包含1千条训练样本对
一条训练样本对 eg: {'en': 'Membership of Parliament: see Minutes', 'ro': 'Componenţa Parlamentului: a se vedea procesul-verbal'}
Returns:
"""
#取出所有的src语句和trg语句，形成列表
inputs = [ex[source_lang] for ex in examples["translation"]]
targets = [ex[target_lang] for ex in examples["translation"]]
# 如果是T5模型，需要加前缀
inputs = [prefix + inp for inp in inputs]
#max_source_length： 输入序列的的最大长度 , 返回tokenid和attention mask。 model_inputs.data : {'input_ids':[[xx,xx],...], 'attention_mask':[[1,1],...}
model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True)
# 为目标设置token程序， labels也是和model_inputs同样的格式
with tokenizer.as_target_tokenizer():
labels = tokenizer(targets, max_length=max_target_length, padding=padding, truncation=True)
# 如果我们在这里进行填充，当我们想忽略损失中的填充时，将标签中所有tokenizer.pad_token_id替换为-100。
if padding == "max_length" and data_args.ignore_pad_token_for_loss:
labels["input_ids"] = [
[(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
]
model_inputs["labels"] = labels["input_ids"]
return model_inputs
if training_args.do_train:
train_dataset = datasets["train"]
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
if data_args.max_train_samples is not None:
# 是否从训练样本中截取部分，一般测试用
train_dataset = train_dataset.select(range(data_args.max_train_samples))
train_dataset = train_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
max_target_length = data_args.val_max_target_length
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
max_target_length = data_args.val_max_target_length
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(range(data_args.max_test_samples))
test_dataset = test_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# 数据的 collator 函数： label_pad_token_id: -100
label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
if data_args.pad_to_max_length:
data_collator = default_data_collator
else:
data_collator = DataCollatorForSeq2Seq(
tokenizer,
model=model,
label_pad_token_id=label_pad_token_id,
pad_to_multiple_of=8 if training_args.fp16 else None,
)
# 评估指标 Metric
if os.path.exists('data/sacrebleu.py'):
metric = load_metric('data/sacrebleu.py')
else:
metric = load_metric("sacrebleu")
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
if data_args.ignore_pad_token_for_loss:
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
result = metric.compute(predictions=decoded_preds, references=decoded_labels)
result = {"bleu": result["score"]}
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result
# 初始化我们的Trainer
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
)
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(
max_length=data_args.val_max_target_length, num_beams=data_args.num_beams, metric_key_prefix="eval"
)
max_val_samples = data_args.max_val_samples if data_args.max_val_samples is not None else len(eval_dataset)
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if training_args.do_predict:
logger.info("*** Test ***")
test_results = trainer.predict(
test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.num_beams,
)
metrics = test_results.metrics
max_test_samples = data_args.max_test_samples if data_args.max_test_samples is not None else len(test_dataset)
metrics["test_samples"] = min(max_test_samples, len(test_dataset))
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
if trainer.is_world_process_zero():
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
test_preds = [pred.strip() for pred in test_preds]
output_test_preds_file = os.path.join(training_args.output_dir, "test_generations.txt")
with open(output_test_preds_file, "w") as writer:
writer.write("\n".join(test_preds))
return results