def generate_summaries(
examples: list,
out_file: str,
model_name: str,
batch_size: int = 8,
device: str = DEFAULT_DEVICE,
fp16=True,
task="summarization",
decoder_start_token_id=None,
finetune_flag: int = 0,
checkpoint_path: str = "",
**gen_kwargs,
) -> None:
fout = Path(out_file).open("w", encoding="utf-8")
# initialize tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
# if our goal is to evaluate the original checkpoint
if finetune_flag < 1:
# initialize the model checkpoints
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
# if our goal is to evaluate our fine-tuned checkpoint
else:
# load the finetuned checkpoints
model = AutoModelForSeq2SeqLM.from_pretrained(
f"{checkpoint_path}/best_tfmr").to(device)
if fp16:
model = model.half()
if decoder_start_token_id is None:
decoder_start_token_id = gen_kwargs.pop("decoder_start_token_id", None)
# update config with summarization specific params
use_task_specific_params(model, task)
for batch in tqdm(list(chunks(examples, batch_size))):
batch = tokenizer(batch,
return_tensors="pt",
truncation=True,
padding="max_length").to(device)
input_ids, attention_mask = trim_batch(
**batch, pad_token_id=tokenizer.pad_token_id)
# -----------------------------------------
# Topic Modeling - GSM
# -----------------------------------------
docs = []
# load dict
dictionary = Dictionary.load(datapath('dict-www-cnndm-unigram'))
# remove [SEP]
sep_list = [
'[SEP_0]', '[SEP_1]', '[SEP_2]', '[SEP_3]', '[SEP_4]', '[SEP_5]',
'[SEP_6]', '[SEP_7]', '[SEP_8]', '[SEP_9]'
]
# vocab size for topic modeling
vocab_size = len(dictionary)
# load config for GSM
config = yaml_load(f"data/config/gsm.yaml")
# model
config['hidden']['features'][0] = vocab_size
# trainer batch
config['trainer_batch']['test_sample'] = 1
config = extend_config_reference(config)
gsm_trainer = config['GSMtrainer']
gsm_trainer['base_dir'] = f"log/bart-large-cnn-finetune"
gsm_trainer = GSMTrainer.from_config(gsm_trainer)
total_sample = len(batch['input_ids'])
for batch_num in range(total_sample):
# extract the batch_sentence
batch_sentence = tokenizer.decode(
batch['input_ids'][batch_num].tolist(),
skip_special_tokens=True)
# change to lowercase and split to list
batch_sentence_list = batch_sentence.split(" ")
# remove [SEP]
batch_sentence_list_nosep = [
item for item in batch_sentence_list if item not in sep_list
]
text = ' '.join([x for x in batch_sentence_list_nosep])
fine_text = text.replace(' ##', '').lower()
batch_sentence = re.sub(r'[^\w\s]', '', fine_text)
# batch_sentence: change to the cleaned news for topic modeling
# change to training data format in topic modeling
gsm_data_bow = dictionary.doc2bow(batch_sentence.split(" "))
docs.append(gsm_data_bow)
# gsm_data: data for topic modeling
gsm_data = DataLoader(DocDataset(docs, len(dictionary), device='cuda'),
batch_size=config['dataset']['batch_size'],
drop_last=False,
num_workers=0)
gsm_trainer.__dict__['train_iterator'] = gsm_data
gsm_loss, gsm_p = gsm_trainer.co_train(vocab_size=vocab_size,
training=False)
del gsm_data
topic_p = gsm_p.cuda()
summaries = model.generate(
input_ids=input_ids,
attention_mask=attention_mask,
decoder_start_token_id=decoder_start_token_id,
topic_p=topic_p,
**gen_kwargs,
)
dec = tokenizer.batch_decode(summaries,
skip_special_tokens=True,
clean_up_tokenization_spaces=False)
for hypothesis in dec:
fout.write(hypothesis + "\n")
fout.flush()
def test_integration_torch_conversation_blenderbot_400M_input_ids(self):
tokenizer = AutoTokenizer.from_pretrained(
"facebook/blenderbot-400M-distill")
model = AutoModelForSeq2SeqLM.from_pretrained(
"facebook/blenderbot-400M-distill")
conversation_agent = ConversationalPipeline(model=model,
tokenizer=tokenizer)
# test1
conversation_1 = Conversation("hello")
inputs = conversation_agent._parse_and_tokenize([conversation_1])
self.assertEqual(inputs["input_ids"].tolist(), [[1710, 86, 2]])
# test2
conversation_1 = Conversation(
"I like lasagne.",
past_user_inputs=["hello"],
generated_responses=[
" Do you like lasagne? It is a traditional Italian dish consisting of a shepherd's pie."
],
)
inputs = conversation_agent._parse_and_tokenize([conversation_1])
self.assertEqual(
inputs["input_ids"].tolist(),
[
# This should be compared with the same conversation on ParlAI `safe_interactive` demo.
[
1710, # hello
86,
228, # Double space
228,
946,
304,
398,
6881,
558,
964,
38,
452,
315,
265,
6252,
452,
322,
968,
6884,
3146,
278,
306,
265,
617,
87,
388,
75,
341,
286,
521,
21,
228, # Double space
228,
281, # I like lasagne.
398,
6881,
558,
964,
21,
2, # EOS
]
],
)
def main():
args = parse_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 pick up all supported trackers in the environment
accelerator = Accelerator(log_with="all", logging_dir=args.output_dir) if args.with_tracking else Accelerator()
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: ' `"
)
# 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))
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)
# 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,
)
# 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)
args.max_train_steps = args.num_train_epochs * 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
if args.with_tracking:
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("summarization_no_trainer", experiment_config)
# 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
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:
resume_step = int(training_difference.replace("step_", ""))
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
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:
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, labels = accelerator.gather((generated_tokens, labels))
generated_tokens = generated_tokens.cpu().numpy()
labels = 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)
# 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,
)
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.with_tracking:
result["train_loss"] = total_loss
result["epoch"] = epoch
result["step"] = completed_steps
accelerator.log(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, 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_rouge1": result["rouge1"],
"eval_rouge2": result["rouge2"],
"eval_rougeL": result["rougeL"],
"eval_rougeLsum": result["rougeLsum"],
},
f,
)
def __init__(self,
config=None,
*inputs,
args,
tasks,
vocab_sets,
save_directory=None,
**kwargs):
config = AutoConfig.from_pretrained(args.pretrained_model,
cache_dir=args.embeddings)
super().__init__(config)
self.args = args
args.dimension = config.d_model
self._is_bart_large = self.args.pretrained_model == 'facebook/bart-large'
self._is_mbart = 'mbart' in self.args.pretrained_model
self._is_mbart50 = self._is_mbart and '-50-' in self.args.pretrained_model
self.src_lang, self.tgt_lang = adjust_language_code(
config, args.pretrained_model, kwargs.get('src_lang', 'en'),
kwargs.get('tgt_lang', 'en'))
if save_directory is not None:
self.model = AutoModelForSeq2SeqLM.from_config(config)
else:
self.model = AutoModelForSeq2SeqLM.from_pretrained(
self.args.pretrained_model, cache_dir=self.args.embeddings)
self.numericalizer = TransformerNumericalizer(
self.args.pretrained_model, args, max_generative_vocab=None)
self.numericalizer.get_tokenizer(save_directory, config, self.src_lang,
self.tgt_lang)
self.init_vocab_from_data(vocab_sets, tasks, save_directory)
self.model.resize_token_embeddings(self.numericalizer.num_tokens)
# set decoder_start_token_id
# recommended by huggingface
# TODO check if it's actually useful
if self.model.config.decoder_start_token_id is None and isinstance(
self.numericalizer._tokenizer,
(MBartTokenizer, MBartTokenizerFast)):
if isinstance(self.numericalizer._tokenizer, MBartTokenizer):
self.model.config.decoder_start_token_id = self.numericalizer._tokenizer.lang_code_to_id[
self.tgt_lang]
else:
self.model.config.decoder_start_token_id = self.numericalizer._tokenizer.convert_tokens_to_ids(
self.tgt_lang)
if self.model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that decoder_start_token_id for the model is defined"
)
if args.dropper_ratio > 0:
# lazy import since dropper is an optional dependency
from loss_dropper import LossDropper
self.dropper = LossDropper(dropc=args.dropper_ratio,
min_count=args.dropper_min_count)
else:
self.dropper = None
self.criterion = LabelSmoothingCrossEntropy(args.label_smoothing)
def main(args_file=None):
# 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, TrainingArguments))
if (len(sys.argv) == 2
and sys.argv[1].endswith(".json")) or args_file is not None:
# 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.
args_file_path = os.path.abspath(
sys.argv[1]) if args_file is None else args_file
model_args, data_args, training_args = parser.parse_json_file(
json_file=args_file_path)
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
assert model_args.model_type in list(
MODEL_TYPE_TO_TOKENIZER.keys()), "model type should be 't5' or 'bart'"
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 training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# Set project name
os.environ["WANDB_PROJECT"] = "question-generation"
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
train_dataset, valid_dataset = processed_data(data_args, model_args,
training_args)
tokenizer_cls = MODEL_TYPE_TO_TOKENIZER[model_args.model_type]
tokenizer = tokenizer_cls.from_pretrained(
model_args.tokenizer_name_or_path if model_args.tokenizer_name_or_path
else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
model.resize_token_embeddings(len(tokenizer))
if model_args.freeze_embeds:
logger.info("freezing embeddings of the model")
freeze_embeds(model)
assert_not_all_frozen(model)
'''
# Get datasets
logger.info('loading dataset')
train_dataset = torch.load(data_args.train_file_path) if training_args.do_train else None
valid_dataset = torch.load(data_args.valid_file_path) if training_args.do_eval else None
logger.info('finished loading dataset')
'''
# Initialize data_collator
data_collator = T2TDataCollator(tokenizer=tokenizer,
model_type=model_args.model_type,
mode="training",
using_tpu=training_args.tpu_num_cores
is not None)
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=valid_dataset,
data_collator=data_collator,
#prediction_loss_only=True,
label_smoothing=model_args.label_smoothing)
# disable wandb console logs
logging.getLogger('wandb.run_manager').setLevel(logging.WARNING)
# Training
if training_args.do_train:
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()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
#if trainer.is_world_master():
# tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
results = {}
if training_args.do_eval and training_args.local_rank in [-1, 0]:
logger.info("*** Evaluate ***")
eval_output = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir,
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(eval_output.keys()):
logger.info(" %s = %s", key, str(eval_output[key]))
writer.write("%s = %s\n" % (key, str(eval_output[key])))
results.update(eval_output)
return results
def model(self):
model = AutoModelForSeq2SeqLM.from_pretrained(
self.ckpt).to(torch_device)
if torch_device == "cuda":
model = model.half()
return model
def model(self):
return AutoModelForSeq2SeqLM.from_pretrained(
self.checkpoint_name).to(torch_device)
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 training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# 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,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=".ckpt" in model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
)
tokenizer.add_special_tokens({'additional_special_tokens': [
'<|HOME|>', '<|AWAY|>',
'<|PLAYER-START_POSITION|>', '<|PLAYER-MIN|>', '<|PLAYER-PTS|>', '<|PLAYER-FGM|>', '<|PLAYER-FGA|>', '<|PLAYER-FG_PCT|>', '<|PLAYER-FG3M|>', '<|PLAYER-FG3A|>', '<|PLAYER-FG3_PCT|>', '<|PLAYER-FTM|>', '<|PLAYER-FTA|>', '<|PLAYER-FT_PCT|>', '<|PLAYER-OREB|>', '<|PLAYER-DREB|>', '<|PLAYER-REB|>', '<|PLAYER-AST|>', '<|PLAYER-TO|>', '<|PLAYER-STL|>', '<|PLAYER-BLK|>', '<|PLAYER-PF|>',
'<|TEAM-PTS_QTR1|>', '<|TEAM-PTS_QTR2|>', '<|TEAM-PTS_QTR3|>', '<|TEAM-PTS_QTR4|>', '<|TEAM-PTS|>', '<|TEAM-FG_PCT|>', '<|TEAM-FG3_PCT|>', '<|TEAM-FT_PCT|>', '<|TEAM-REB|>', '<|TEAM-AST|>', '<|TEAM-TOV|>', '<|TEAM-WINS|>', '<|TEAM-LOSSES|>', '<|TEAM-CITY|>', '<|TEAM-NAME|>',
]})
# tokenizer.model_max_length = 1300
# tokenizer.max_length = 1300
print(tokenizer.model_max_length)
model.resize_token_embeddings(len(tokenizer))
freq_seqs = get_freq_sequences(data_args.data_dir, tokenizer) if model_args.trigram_loss else None
# use task specific params
use_task_specific_params(model, data_args.task)
# set num_beams for evaluation
if data_args.eval_beams is not None:
model.config.num_beams = data_args.eval_beams
assert model.config.num_beams >= 1, f"got eval_beams={model.config.num_beams}. Need an integer >= 1"
# set max length for generation
model.config.max_generate_length = data_args.val_max_target_length
# set decoder_start_token_id for MBart
if model.config.decoder_start_token_id is None and isinstance(tokenizer, MBartTokenizer):
decoder_start_token_id = tokenizer.lang_code_to_id[data_args.tgt_lang]
model.config.decoder_start_token_id = decoder_start_token_id
def build_compute_metrics_fn(task_name: str) -> Callable[[EvalPrediction], Dict]:
def non_pad_len(tokens: np.ndarray) -> int:
return np.count_nonzero(tokens != tokenizer.pad_token_id)
def decode_pred(pred: EvalPrediction) -> Tuple[List[str], List[str]]:
pred_str = tokenizer.batch_decode(pred.predictions, skip_special_tokens=True)
label_str = tokenizer.batch_decode(pred.label_ids, skip_special_tokens=True)
pred_str = lmap(str.strip, pred_str)
label_str = lmap(str.strip, label_str)
return pred_str, label_str
def summarization_metrics(pred: EvalPrediction) -> Dict:
pred_str, label_str = decode_pred(pred)
rouge: Dict = calculate_rouge(pred_str, label_str)
summ_len = np.mean(lmap(non_pad_len, pred.predictions))
rouge.update({"gen_len": summ_len})
return rouge
def translation_metrics(pred: EvalPrediction) -> Dict:
pred_str, label_str = decode_pred(pred)
bleu: Dict = calculate_bleu(pred_str, label_str)
gen_len = np.mean(lmap(non_pad_len, pred.predictions))
bleu.update({"gen_len": gen_len})
return bleu
compute_metrics_fn = summarization_metrics if "summarization" in task_name else translation_metrics
return compute_metrics_fn
def freeze_embeds(model: torch.nn.Module):
"""Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
try:
freeze_params(model.model.shared)
for d in [model.model.encoder, model.model.decoder]:
freeze_params(d.embed_positions)
freeze_params(d.embed_tokens)
except AttributeError:
freeze_params(model.shared)
for d in [model.encoder, model.decoder]:
freeze_params(d.embed_tokens)
if model_args.freeze_embeds:
freeze_embeds(model)
if model_args.freeze_encoder:
freeze_params(model.get_encoder())
assert_all_frozen(model.get_encoder())
dataset_class = Seq2SeqDataset if hasattr(tokenizer, "prepare_seq2seq_batch") else LegacySeq2SeqDataset
model.config.update(
{'early_stopping': True, 'length_penalty': 2.0, 'max_length': 600, 'min_length': 275, 'no_repeat_ngram_size': 3, 'num_beams': 4,
'max_position_embeddings': 1024
# 'n_positions': 1024,
# 'prefix': 'summarize: '
}
)
# Get datasets
train_dataset = (
dataset_class(
tokenizer,
type_path="train",
data_dir=data_args.data_dir,
n_obs=data_args.n_train,
max_target_length=data_args.max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
)
if training_args.do_train
else None
)
eval_dataset = (
dataset_class(
tokenizer,
type_path="val",
data_dir=data_args.data_dir,
n_obs=data_args.n_val,
max_target_length=data_args.val_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
)
if training_args.do_eval
else None
)
test_dataset = (
dataset_class(
tokenizer,
type_path="test",
data_dir=data_args.data_dir,
n_obs=data_args.n_test,
max_target_length=data_args.test_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
)
if training_args.do_predict
else None
)
# model.config.update({'n_positions': 1024})
# print(model.config)
# Initialize our Trainer
trainer = Seq2SeqTrainer(
freq_seqs,
data_args.data_dir,
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=Seq2SeqDataCollator(tokenizer, data_args, training_args.tpu_num_cores),
compute_metrics=build_compute_metrics_fn(data_args.task) if training_args.predict_with_generate else None,
)
# Training
if training_args.do_train:
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()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_process_zero():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
eval_results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
result = trainer.evaluate()
output_eval_file = os.path.join(training_args.output_dir, "eval_results.json")
if trainer.is_world_process_zero():
logger.info("***** Eval results *****")
for key, value in result.items():
logger.info(" %s = %s", key, value)
with open(output_eval_file, "w") as f:
json.dump(result, f)
eval_results.update(result)
if training_args.do_predict:
logging.info("*** Test ***")
test_output = trainer.predict(test_dataset=test_dataset)
test_metrics = test_output.metrics
test_metrics = {k.replace("eval", "test"): v for k, v in test_metrics.items()}
output_test_file = os.path.join(training_args.output_dir, "test_results.json")
if trainer.is_world_process_zero():
logger.info("***** Test results *****")
for key, value in test_metrics.items():
logger.info(" %s = %s", key, value)
with open(output_test_file, "w") as f:
json.dump(test_metrics, f)
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(test_output.predictions, skip_special_tokens=True)
test_preds = lmap(str.strip, test_preds)
output_test_pred_file = os.path.join(training_args.output_dir, "test_generations.txt")
with open(output_test_pred_file, "w") as f:
f.write("\n".join(test_preds))
return eval_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))
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():
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")
import numpy as np
from tqdm.auto import tqdm
from sari.SARI import SARIsent
from rouge_score import rouge_scorer, scoring
from Data_Preprocessing import preprocess_QG
from TextSimp_QG_model import run_model
from transformers import (
AutoConfig,
AutoModelForSeq2SeqLM,
AutoTokenizer,
)
path = "/content/drive/My Drive/model/best_tfmr"
model = AutoModelForSeq2SeqLM.from_pretrained(path)
tokenizer = AutoTokenizer.from_pretrained(path)
model.eval()
def calculate_metrics(model):
_, _, dev = preprocess_QG()
sentences = pd.DataFrame(dev, columns=['Complex', 'Simple'])
sentences = sentences.groupby(
['Complex']).agg(lambda x: tuple(x)).applymap(list).reset_index()
questions = []
with open("/content/tgt-dev.txt", 'r') as f:
lines = f.readlines()
for l in lines:
questions.append(l[:-1])
def __init__(self, hparams):
super(AbstractiveSummarizer, self).__init__()
self.hparams = hparams
if "longformer-encdec" in self.hparams.model_name_or_path.lower():
self.model = (
LongformerEncoderDecoderForConditionalGeneration.from_pretrained(
self.hparams.model_name_or_path, gradient_checkpointing=True
)
)
self.tokenizer = BartTokenizerFast.from_pretrained(
self.hparams.model_name_or_path, add_prefix_space=True
)
else:
if self.hparams.decoder_model_name_or_path:
self.model = EncoderDecoderModel.from_encoder_decoder_pretrained(
self.hparams.model_name_or_path,
(
self.hparams.decoder_model_name_or_path
if self.hparams.decoder_model_name_or_path
else self.hparams.model_name_or_path
),
gradient_checkpointing=self.hparams.gradient_checkpointing,
tie_encoder_decoder=self.hparams.tie_encoder_decoder,
)
else:
self.model = AutoModelForSeq2SeqLM.from_pretrained(
self.hparams.model_name_or_path,
gradient_checkpointing=self.hparams.gradient_checkpointing,
)
self.tokenizer = AutoTokenizer.from_pretrained(
self.hparams.model_name_or_path, use_fast=True
)
self.rouge_sentence_split_token = "<q>"
self.tokenizer.add_tokens(self.rouge_sentence_split_token)
self.rouge_sentence_split_token_id = self.tokenizer.convert_tokens_to_ids(
self.rouge_sentence_split_token
)
# bo = beginning of
# eo = ending of
# seq = sequence (not using 's' because 's' stands for sentence in other places)
# Use `bos_token` for boseq if `bos_token` is set, otherwise use "[unused0]"
# Use `pad_token` for eoseq if `pad_token` is set, otherwise use "[unused1]"
do_seq_special_add = False
if self.tokenizer.bos_token:
self.target_boseq_token = self.tokenizer.bos_token
else:
self.target_boseq_token = "[unused0]"
do_seq_special_add = True
if self.tokenizer.pad_token:
self.target_eoseq_token = self.tokenizer.pad_token
else:
self.target_eoseq_token = "[unused1]"
do_seq_special_add = True
# Convert `target_boseq_token` and `target_eoseq_token` to IDs
self.target_boseq_token_id = self.tokenizer.convert_tokens_to_ids(
self.target_boseq_token
)
self.target_eoseq_token_id = self.tokenizer.convert_tokens_to_ids(
self.target_eoseq_token
)
# If the `*oseq` tokens are not already "special" then add them as special
# tokens so that they are ignored when decoding.
if do_seq_special_add:
special_tokens_dict = {
"additional_special_tokens": [
self.target_boseq_token,
self.target_eoseq_token,
]
}
self.tokenizer.add_special_tokens(special_tokens_dict)
if self.hparams.label_smoothing > 0:
self.loss_func = LabelSmoothingLoss(
self.hparams.label_smoothing,
self.tokenizer.vocab_size,
ignore_index=self.tokenizer.pad_token_id,
)
else:
self.loss_func = nn.CrossEntropyLoss(
ignore_index=self.tokenizer.pad_token_id
)
self.train_dataloader_object = None # not created yet
self.rouge_metrics = None
self.rouge_scorer = None
self.dataset = {}
self.tokenized_data_file_paths = {}
for split in ["train", "validation", "test"]:
features_cache_file = os.path.join(
self.hparams.cache_file_path, (split + "_tokenized")
)
self.tokenized_data_file_paths[split] = features_cache_file
if "longformer" in self.hparams.model_name_or_path:
self.collate_fn = partial(self.abs_collate_fn, modifier=longformer_modifier)
else:
self.collate_fn = self.abs_collate_fn
summary = summarize(input_fp, result_fp, model, max_length=3)
print(summary)
#%%
####################
########## BARThez
####################
text_sentence = "Citant les préoccupations de ses clients dénonçant des cas de censure après la suppression du compte de Trump, un fournisseur d'accès Internet de l'État de l'Idaho a décidé de bloquer Facebook et Twitter. La mesure ne concernera cependant que les clients mécontents de la politique de ces réseaux sociaux."
import torch
from transformers import (AutoTokenizer, AutoModelForSeq2SeqLM)
barthez_tokenizer = AutoTokenizer.from_pretrained("moussaKam/barthez")
barthez_model = AutoModelForSeq2SeqLM.from_pretrained(
"moussaKam/barthez-orangesum-abstract")
# MBartForConditionalGeneration(
# (model): MBartModel(
# (shared): Embedding(50002, 768, padding_idx=1)
# (encoder): MBartEncoder(
# (embed_tokens): Embedding(50002, 768, padding_idx=1)
# (embed_positions): MBartLearnedPositionalEmbedding(1026, 768)
input_ids = torch.tensor(
[barthez_tokenizer.encode(text_sentence, add_special_tokens=True)])
barthez_model.eval()
predict = barthez_model.generate(input_ids, max_length=100)[0]
barthez_tokenizer.decode(predict, skip_special_tokens=True)
# %%
################# BERTSUM
import json
import os
from src.helpers import create_raw_input, create_input, run_inference, device
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch
from flask import Flask, jsonify, request
app = Flask(__name__)
rewriters = {
# other rewriters go here
"t5_rewriter": {
"model": AutoModelForSeq2SeqLM.from_pretrained("castorini/t5-base-canard").to(device),
"tokenizer": AutoTokenizer.from_pretrained("castorini/t5-base-canard")
}
}
@app.route('/rewrite_raw', methods=['GET', 'POST'])
def rewrite_raw():
rewrite_request = request.get_data()
rewrite_request = json.loads(rewrite_request)
raw_query = rewrite_request["raw_query"]
context = rewrite_request["context"]
turn_separator = rewrite_request["turn_separator"]
rewriter_choice = rewrite_request["rewriter_choice"]
def test_rag_token_from_pretrained(self):
rag_config = self.get_rag_config()
rag_decoder_tokenizer = BartTokenizer.from_pretrained(
"facebook/bart-large-cnn")
rag_question_encoder_tokenizer = DPRQuestionEncoderTokenizer.from_pretrained(
"facebook/dpr-question_encoder-single-nq-base")
rag_retriever = RagRetriever(
rag_config,
question_encoder_tokenizer=rag_question_encoder_tokenizer,
generator_tokenizer=rag_decoder_tokenizer,
)
input_ids = rag_question_encoder_tokenizer(
"who sings does he love me with reba",
return_tensors="pt").input_ids
decoder_input_ids = rag_decoder_tokenizer(
"Linda Davis", return_tensors="pt").input_ids
input_ids = input_ids.to(torch_device)
decoder_input_ids = decoder_input_ids.to(torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
rag_token = RagTokenForGeneration.from_pretrained_question_encoder_generator(
"facebook/dpr-question_encoder-single-nq-base",
"facebook/bart-large-cnn",
retriever=rag_retriever,
config=rag_config,
).to(torch_device)
# check that the from pretrained methods work
rag_token.save_pretrained(tmp_dirname)
rag_token.from_pretrained(tmp_dirname, retriever=rag_retriever)
rag_token.to(torch_device)
with torch.no_grad():
output = rag_token(
input_ids,
labels=decoder_input_ids,
)
loss_pretrained = output.loss
del rag_token
question_encoder = AutoModel.from_pretrained(
"facebook/dpr-question_encoder-single-nq-base")
generator = AutoModelForSeq2SeqLM.from_pretrained(
"facebook/bart-large-cnn")
rag_token = RagTokenForGeneration(config=rag_config,
question_encoder=question_encoder,
generator=generator,
retriever=rag_retriever)
rag_token.to(torch_device)
with torch.no_grad():
output = rag_token(
input_ids,
labels=decoder_input_ids,
)
loss_init = output.loss
self.assertAlmostEqual(loss_pretrained.item(),
loss_init.item(),
places=4)
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import sentence_seperator
tokenizer = AutoTokenizer.from_pretrained('models/en_sum/distilbart-cnn-12-6')
model = AutoModelForSeq2SeqLM.from_pretrained(
'models/en_sum/distilbart-cnn-12-6')
def summary(long_text: str) -> str:
if len(long_text) > tokenizer.model_max_length:
chunks = sentence_seperator.sep_sent(long_text, tokenizer)
inputs = [tokenizer(c, return_tensors='pt') for c in chunks]
for input in inputs:
outs = model.generate(**input)
return ''.join(tokenizer.batch_decode(*outs, skip_special_tokens=True))
else:
inputs = tokenizer([long_text], return_tensors='pt')
outs = model.generate(inputs['input_ids'])
return tokenizer.batch_decode(outs, skip_special_tokens=True)
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)
# 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.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
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"]
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.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 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 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."
)
# 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.
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))
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
if (hasattr(model.config, "max_position_embeddings")
and model.config.max_position_embeddings <
data_args.max_source_length):
if model_args.resize_position_embeddings is None:
logger.warning(
f"Increasing the model's number of position embedding vectors from {model.config.max_position_embeddings} "
f"to {data_args.max_source_length}.")
model.resize_position_embeddings(data_args.max_source_length)
elif model_args.resize_position_embeddings:
model.resize_position_embeddings(data_args.max_source_length)
else:
raise ValueError(
f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has {model.config.max_position_embeddings}"
f" position encodings. Consider either reducing `--max_source_length` to {model.config.max_position_embeddings} or to automatically "
"resize the model's position encodings by passing `--resize_position_embeddings`."
)
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
# 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):
# remove pairs where at least one record is None
inputs, targets = [], []
for i in range(len(examples[text_column])):
if examples[text_column][i] is not None and examples[
summary_column][i] is not None:
inputs.append(examples[text_column][i])
targets.append(examples[summary_column][i])
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
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 = {}
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") as writer:
writer.write("\n".join(predictions))
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "summarization"
}
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
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
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(
)
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."
)
# 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 training_args.should_log 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 training_args.should_log:
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 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.
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))
# 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 = 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 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 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_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
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 = {}
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",
encoding="utf-8") as writer:
writer.write("\n".join(predictions))
if training_args.push_to_hub:
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tags": "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
trainer.push_to_hub(**kwargs)
return results
dataset = datasets.load_dataset("wmt14", "de-en", split="test")
dataset = list(dataset)
dataset = [xx['translation'] for xx in dataset]
dataset = [
dataset[i] for i in np.random.RandomState(123).permutation(len(dataset))
]
# --
# Load model
tokenizer = AutoTokenizer.from_pretrained("google/bert2bert_L-24_wmt_de_en",
pad_token="<pad>",
eos_token="</s>",
bos_token="<s>")
model = AutoModelForSeq2SeqLM.from_pretrained(
"google/bert2bert_L-24_wmt_de_en").cuda()
# --
# Dataloader
def collate_fn(batch):
inputs = [xx['de'] for xx in batch]
inputs = tokenizer(inputs,
padding="max_length",
truncation=True,
max_length=128,
return_tensors="pt")
targets = [xx['en'] for xx in batch]
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, TrainingArguments, ConfigArguments))
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, config_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1])
)
else:
model_args, data_args, training_args, config_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()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
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/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.
datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, keep_in_memory=False, cache_dir=model_args.cache_dir)
if "validation" not in datasets.keys():
datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
keep_in_memory=False,
cache_dir=model_args.cache_dir
)
datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
keep_in_memory=False,
cache_dir=model_args.cache_dir
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = (
data_args.train_file.split(".")[-1]
if data_args.train_file is not None
else data_args.validation_file.split(".")[-1]
)
if extension == "txt":
extension = "text"
datasets = load_dataset(extension, data_files=data_files, keep_in_memory=False, cache_dir=model_args.cache_dir)
if data_args.sanity:
datasets["train"] = datasets["train"].shard(100, index=0, contiguous=True)
# 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_kwargs = {
"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_args.config_name:
config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
elif model_args.model_name_or_path:
config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
else:
config = CONFIG_MAPPING[model_args.model_type](**vars(config_args), **config_kwargs)
logger.warning("You are instantiating a new config instance from scratch.")
tokenizer_kwargs = {
"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,
}
if model_args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
elif model_args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
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 model_args.model_name_or_path:
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,
)
else:
logger.info("Training new model from scratch")
model = AutoModelForSeq2SeqLM.from_config(config)
model.resize_token_embeddings(len(tokenizer))
# Preprocessing the datasets.
# First we tokenize all the texts.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
# since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
datasets = datasets.shuffle()
def tokenize_function(examples):
with CaptureLogger(tok_logger) as cl:
output = tokenizer(examples[text_column_name])
# clm input could be much much longer than block_size
if "Token indices sequence length is longer than the" in cl.out:
tok_logger.warning(
"^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model."
)
return output
# Ensures only the main process does dataset pre-processing; the other ones will load the `map` cache
if not is_main_process(training_args.local_rank):
print("waiting for main process to execute mapping")
torch.distributed.barrier()
logger.info("Mapping dataset to tokenized dataset.",)
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
keep_in_memory=False
)
if data_args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warning(
f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
"Picking 1024 instead. You can change that default value by passing --block_size xxx."
)
block_size = 1024
else:
if data_args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(data_args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
total_length = (total_length // (2 * block_size)) * 2 * block_size
# Split by chunks of max_len.
result = {
k: [t[i : i + block_size] for i in range(0, total_length, 2*block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = [
concatenated_examples['input_ids'][i : i + block_size]
for i in range(block_size, total_length, 2*block_size)
]
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
logger.info("Chunking tokenized dataset.")
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
keep_in_memory=False
)
# Now the other ones can catch up.
if training_args.local_rank != -1 and is_main_process(training_args.local_rank):
print("loading results from main process")
torch.distributed.barrier()
if training_args.do_train:
if "train" not in tokenized_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = lm_datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(range(data_args.max_train_samples))
if training_args.do_eval:
if "validation" not in tokenized_datasets:
cutoff = data_args.validation_split_percentage * len(lm_datasets["train"]) // 100
train_dataset = lm_datasets["train"].select(range(cutoff, len(lm_datasets["train"])))
eval_dataset = lm_datasets["train"].select(range(cutoff))
else:
eval_dataset = lm_datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(data_args.max_val_samples))
# Initialize our Trainer
trainer = Trainer(
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 will default to DataCollatorWithPadding, so we change it.
data_collator=default_data_collator,
callbacks=[LogFlosCallback, TensorBoardFloIndexedCallback]
)
# 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
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
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))
perplexity = math.exp(metrics["eval_loss"])
metrics["perplexity"] = perplexity
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(
)
check_output_dir(training_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.parallel_mode == ParallelMode.DISTRIBUTED),
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()
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
logger.info("Training/evaluation parameters %s", training_args)
# Set seed
set_seed(training_args.seed)
# 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,
)
extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout",
"attention_dropout")
for p in extra_model_params:
if getattr(training_args, p, None):
assert hasattr(
config, p
), f"({config.__class__.__name__}) doesn't have a `{p}` attribute"
setattr(config, p, getattr(training_args, p))
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,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
model_args.model_name_or_path,
from_tf=".ckpt" in model_args.model_name_or_path,
# config=config,
min_length=data_args.min_target_length,
max_length=data_args.max_target_length,
num_beams=data_args.eval_beams,
cache_dir=model_args.cache_dir,
)
# use task specific params
use_task_specific_params(model, data_args.task)
# set num_beams for evaluation
if data_args.eval_beams is None:
data_args.eval_beams = model.config.num_beams
# set decoder_start_token_id for MBart
if model.config.decoder_start_token_id is None and isinstance(
tokenizer, MBartTokenizer):
assert (data_args.tgt_lang is not None and data_args.src_lang
is not None), "mBart requires --tgt_lang and --src_lang"
model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
data_args.tgt_lang]
if model_args.freeze_embeds:
freeze_embeds(model)
if model_args.freeze_encoder:
freeze_params(model.get_encoder())
assert_all_frozen(model.get_encoder())
dataset_class = Seq2SeqDataset
# Get datasets
train_dataset = (dataset_class(
tokenizer,
type_path="train",
data_dir=data_args.data_dir,
n_obs=data_args.n_train,
max_target_length=data_args.max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
) if training_args.do_train else None)
eval_dataset = (dataset_class(
tokenizer,
type_path="val",
data_dir=data_args.data_dir,
n_obs=data_args.n_val,
max_target_length=data_args.val_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
) if training_args.do_eval or
training_args.evaluation_strategy != EvaluationStrategy.NO
else None)
test_dataset = (dataset_class(
tokenizer,
type_path="test",
data_dir=data_args.data_dir,
n_obs=data_args.n_test,
max_target_length=data_args.test_max_target_length,
max_source_length=data_args.max_source_length,
prefix=model.config.prefix or "",
) if training_args.do_predict else None)
# Initialize our Trainer
compute_metrics_fn = (build_compute_metrics_fn(data_args.task, tokenizer)
if training_args.predict_with_generate else None)
trainer = Seq2SeqTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
data_collator=Seq2SeqDataCollator(tokenizer, data_args,
training_args.tpu_num_cores),
compute_metrics=compute_metrics_fn,
tokenizer=tokenizer,
)
all_metrics = {}
# Training
if training_args.do_train:
logger.info("*** 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)
metrics = train_result.metrics
metrics["train_n_objs"] = data_args.n_train
trainer.save_model() # this also saves the tokenizer
if trainer.is_world_process_zero():
handle_metrics("train", metrics, training_args.output_dir)
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"))
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate(metric_key_prefix="val",
max_length=data_args.val_max_target_length,
num_beams=data_args.eval_beams)
metrics["val_n_objs"] = data_args.n_val
metrics["val_loss"] = round(metrics["val_loss"], 4)
if trainer.is_world_process_zero():
handle_metrics("val", metrics, training_args.output_dir)
all_metrics.update(metrics)
if training_args.do_predict:
logger.info("*** Predict ***")
test_output = trainer.predict(
test_dataset=test_dataset,
metric_key_prefix="test",
max_length=data_args.val_max_target_length,
num_beams=data_args.eval_beams,
)
metrics = test_output.metrics
metrics["test_n_objs"] = data_args.n_test
if trainer.is_world_process_zero():
metrics["test_loss"] = round(metrics["test_loss"], 4)
handle_metrics("test", metrics, training_args.output_dir)
all_metrics.update(metrics)
if training_args.predict_with_generate:
test_preds = tokenizer.batch_decode(
test_output.predictions,
skip_special_tokens=True,
clean_up_tokenization_spaces=True)
test_preds = lmap(str.strip, test_preds)
write_txt_file(
test_preds,
os.path.join(training_args.output_dir,
"align_wnums_test_generations.txt"))
if trainer.is_world_process_zero():
save_json(all_metrics,
os.path.join(training_args.output_dir, "all_results.json"))
return all_metrics
def eval_data_dir(
data_dir,
save_dir: str,
model_name: str,
bs: int = 8,
max_source_length: int = 1024,
type_path="val",
n_obs=None,
fp16=False,
task="summarization",
local_rank=None,
**generate_kwargs,
) -> Dict:
"""Run evaluation on part of the data for one gpu and save to {save_dir}/rank_{rank}_output.json"""
model_name = str(model_name)
assert local_rank is not None
torch.distributed.init_process_group(backend="nccl", rank=local_rank)
save_dir = Path(save_dir)
save_path = save_dir.joinpath(f"rank_{local_rank}_output.json")
torch.cuda.set_device(local_rank)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name).cuda()
if fp16:
model = model.half()
tokenizer = AutoTokenizer.from_pretrained(model_name)
logger.info(f"Inferred tokenizer type: {tokenizer.__class__}"
) # if this is wrong, check config.model_type.
use_task_specific_params(model,
task) # update config with task specific params
if max_source_length is None:
max_source_length = tokenizer.model_max_length
ds = Seq2SeqDataset(
tokenizer,
data_dir,
max_source_length,
max_target_length=1024,
type_path=type_path,
n_obs=n_obs,
prefix=model.config.prefix,
)
# I set shuffle=True for a more accurate progress bar.
# If all the longest samples are first, the prog bar estimate is too high at the beginning.
sampler = ds.make_sortish_sampler(bs,
distributed=True,
add_extra_examples=False,
shuffle=True)
data_loader = DataLoader(ds,
sampler=sampler,
batch_size=bs,
collate_fn=ds.collate_fn)
results = []
for batch in tqdm(data_loader):
summaries = model.generate(
input_ids=batch["input_ids"].to(model.device),
attention_mask=batch["attention_mask"].to(model.device),
**generate_kwargs,
)
preds = tokenizer.batch_decode(summaries,
skip_special_tokens=True,
clean_up_tokenization_spaces=False)
ids = batch["ids"]
for i, pred in enumerate(preds):
results.append(dict(pred=pred, id=ids[i].item()))
save_json(results, save_path)
return results, sampler.num_replicas
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from tqdm import tqdm
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-ru-en")
model = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-ru-en").to(
device)
with open('eval-ru-100.txt') as f:
test = f.readlines()
with open('answer.txt', 'w+') as f:
for i, val in tqdm(enumerate(test), total=len(test)):
inputs = tokenizer.encode(val, return_tensors="pt").to(device)
outputs = model.generate(inputs,
max_length=75,
num_beams=16,
early_stopping=True).to(device)
seq = (tokenizer.decode(outputs[0]).replace('<pad> ', ''))
if i != len(test) - 1:
f.write(seq + '\n')
else:
f.write(seq)
def pipeline(
task: str,
model: Optional = None,
tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
qg_format: Optional[str] = "highlight",
ans_model: Optional = None,
ans_tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
use_cuda: Optional[bool] = True,
**kwargs,
):
# Retrieve the task
if task not in SUPPORTED_TASKS:
raise KeyError("Unknown task {}, available tasks are {}".format(
task, list(SUPPORTED_TASKS.keys())))
targeted_task = SUPPORTED_TASKS[task]
task_class = targeted_task["impl"]
# Use default model/config/tokenizer for the task if no model is provided
if model is None:
model = targeted_task["default"]["model"]
# Try to infer tokenizer from model or config name (if provided as str)
if tokenizer is None:
if isinstance(model, str):
tokenizer = model
else:
# Impossible to guest what is the right tokenizer here
raise Exception(
"Impossible to guess which tokenizer to use. "
"Please provided a PretrainedTokenizer class or a path/identifier to a pretrained tokenizer."
)
# Instantiate tokenizer if needed
if isinstance(tokenizer, (str, tuple)):
if isinstance(tokenizer, tuple):
# For tuple we have (tokenizer name, {kwargs})
tokenizer = AutoTokenizer.from_pretrained(tokenizer[0],
**tokenizer[1])
else:
tokenizer = AutoTokenizer.from_pretrained(tokenizer)
# Instantiate model if needed
if isinstance(model, str):
model = AutoModelForSeq2SeqLM.from_pretrained(model)
if task == "question-generation":
if ans_model is None:
# load default ans model
ans_model = targeted_task["default"]["ans_model"]
ans_tokenizer = AutoTokenizer.from_pretrained(ans_model)
ans_model = AutoModelForSeq2SeqLM.from_pretrained(ans_model)
else:
# Try to infer tokenizer from model or config name (if provided as str)
if ans_tokenizer is None:
if isinstance(ans_model, str):
ans_tokenizer = ans_model
else:
# Impossible to guest what is the right tokenizer here
raise Exception(
"Impossible to guess which tokenizer to use. "
"Please provided a PretrainedTokenizer class or a path/identifier to a pretrained tokenizer."
)
# Instantiate tokenizer if needed
if isinstance(ans_tokenizer, (str, tuple)):
if isinstance(ans_tokenizer, tuple):
# For tuple we have (tokenizer name, {kwargs})
ans_tokenizer = AutoTokenizer.from_pretrained(
ans_tokenizer[0], **ans_tokenizer[1])
else:
ans_tokenizer = AutoTokenizer.from_pretrained(
ans_tokenizer)
if isinstance(ans_model, str):
ans_model = AutoModelForSeq2SeqLM.from_pretrained(ans_model)
if task == "e2e-qg":
return task_class(model=model, tokenizer=tokenizer, use_cuda=use_cuda)
elif task == "question-generation":
return task_class(model=model,
tokenizer=tokenizer,
ans_model=ans_model,
ans_tokenizer=ans_tokenizer,
qg_format=qg_format,
use_cuda=use_cuda)
else:
return task_class(model=model,
tokenizer=tokenizer,
ans_model=model,
ans_tokenizer=tokenizer,
qg_format=qg_format,
use_cuda=use_cuda)
def __init__(self, tokenizer_filepath: str):
# logger.info("Initialize %s class with pretrained model %s", self.__class__.__name__, tokenizer_filepath)
self.name = f'{self.__class__.__name__}'
self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_filepath)
self.model = AutoModelForSeq2SeqLM.from_pretrained(tokenizer_filepath)
self.mode = None
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]
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,
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 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:
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
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():
parser = argparse.ArgumentParser()
parser.add_argument(
"--collection_file",
default="data/collection.tsv",
type=str,
help="The msmarco passage collection file",
)
parser.add_argument(
"--model_name_or_path",
type=str,
default=None,
help="Doc2Query predictions",
)
parser.add_argument(
"--augmented_collection_file",
type=str,
default="data/augmented_collection.jsonl",
help="The output_file for augmented doc 2 query index",
)
parser.add_argument(
"--beam_size",
type=int,
default=3,
help="number of queries to generate per passage",
)
parser.add_argument(
"--max_length",
type=int,
default=32,
help="length of document queries",
)
parser.add_argument('--no_cuda',
action="store_true",
help="Use this to not use cuda")
args = parser.parse_args()
print("Loading collection")
collection = load_qid2query(args.collection_file)
print("Collection loaded")
device = 'cuda'
if args.no_cuda:
device = 'cpu'
print("Loading model")
config = AutoConfig.from_pretrained(args.model_name_or_path, )
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, )
model = AutoModelForSeq2SeqLM.from_pretrained(args.model_name_or_path)
model.to(device)
model.resize_token_embeddings(len(tokenizer))
print("Model Loaded")
print("Augmenting passages")
augmentations = 0
#TODO Introduce batching at inference time as right now runs 1 by 1
with open(args.augmented_collection_file, 'w') as w:
for doc_id in collection:
if augmentations % 5000 == 0:
print("{} passages augmented".format(augmentations))
document_text = collection[doc_id]
input_ids = tokenizer.encode(document_text,
return_tensors='pt').to(device)
outputs = model.generate(input_ids=input_ids,
max_length=args.max_length,
do_sample=True,
top_k=10,
num_return_sequences=args.beam_size)
query_augment = ''
for i in range(args.beam_size):
query_augment += ' '
query_augment += tokenizer.decode(outputs[i],
skip_special_tokens=True)
output_dict = {
'id': doc_id,
'contents': document_text + query_augment
}
w.write(json.dumps(output_dict) + '\n')
augmentations += 1
query getArticles {
Article(order_by: {timestamp: asc}, where: {timestamp: {_gte: "2020-02-01T00:00:00", _lte: "2020-03-01T00:00:00"}}) {
timestamp
journal
id
headline
content
}
}
"""
result = asyncio.run(query_data(server_com.url, gql_query))
news = result['Article'][0]
# Put your news here
content = news['content'] # a string of news
headline = news['headline'] # a string of news' headline
# Model
# multi-news(long), newsroom(medium), wikihow (short), cnn_dailymail
tokenizer = AutoTokenizer.from_pretrained("google/pegasus-cnn_dailymail")
model = AutoModelForSeq2SeqLM.from_pretrained("google/pegasus-cnn_dailymail").to(DEVICE)
batch = tokenizer.prepare_seq2seq_batch([content, result['Article'][1]['content']], truncation=True, padding='longest', return_tensors="pt").to(DEVICE)
summary = model.generate(**batch)
summary = tokenizer.batch_decode(summary, skip_special_tokens=True)
print(headline)
print("=============================")
print(content)
print("=============================")
print(summary)
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
