def main():
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
if args.test_file is not None:
data_files["test"] = args.test_file
extension = args.train_file.split(".")[-1]
raw_datasets = load_dataset(extension,
data_files=data_files,
field="data")
# 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.
config = XLNetConfig.from_pretrained(args.model_name_or_path)
tokenizer = XLNetTokenizerFast.from_pretrained(args.model_name_or_path)
model = XLNetForQuestionAnswering.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config)
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
column_names = raw_datasets["train"].column_names
question_column_name = "question" if "question" in column_names else column_names[
0]
context_column_name = "context" if "context" in column_names else column_names[
1]
answer_column_name = "answers" if "answers" in column_names else column_names[
2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
if args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(args.max_seq_length, tokenizer.model_max_length)
# Training preprocessing
def prepare_train_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
tokenized_examples["is_impossible"] = []
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
# The cls token gets 1.0 too (for predictions of empty answers).
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != context_idx:
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != context_idx:
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[
token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(
token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(
token_end_index + 1)
tokenized_examples["is_impossible"].append(0.0)
return tokenized_examples
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets["train"]
if args.max_train_samples is not None:
# We will select sample from whole data if agument is specified
train_dataset = train_dataset.select(range(args.max_train_samples))
# Create train feature from dataset
with accelerator.main_process_first():
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if args.max_train_samples is not None:
# Number of samples might increase during Feature Creation, We select only specified max samples
train_dataset = train_dataset.select(range(args.max_train_samples))
# Validation preprocessing
def prepare_validation_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
# We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, input_ids in enumerate(tokenized_examples["input_ids"]):
# Find the CLS token in the input ids.
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(
examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_idx else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if args.max_eval_samples is not None:
# We will select sample from whole data
eval_examples = eval_examples.select(range(args.max_eval_samples))
# Validation Feature Creation
with accelerator.main_process_first():
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if args.max_eval_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
eval_dataset = eval_dataset.select(range(args.max_eval_samples))
if args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(
range(args.max_predict_samples))
# Predict Feature Creation
with accelerator.main_process_first():
predict_dataset = predict_examples.map(
prepare_validation_features,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
if args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
predict_dataset = predict_dataset.select(
range(args.max_predict_samples))
# 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:
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 = DataCollatorWithPadding(
tokenizer,
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_dataset_for_model = eval_dataset.remove_columns(
["example_id", "offset_mapping"])
eval_dataloader = DataLoader(eval_dataset_for_model,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
if args.do_predict:
predict_dataset_for_model = predict_dataset.remove_columns(
["example_id", "offset_mapping"])
predict_dataloader = DataLoader(
predict_dataset_for_model,
collate_fn=data_collator,
batch_size=args.per_device_eval_batch_size)
# Post-processing:
def post_processing_function(examples,
features,
predictions,
stage="eval"):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=args.version_2_with_negative,
n_best_size=args.n_best_size,
max_answer_length=args.max_answer_length,
start_n_top=model.config.start_n_top,
end_n_top=model.config.end_n_top,
output_dir=args.output_dir,
prefix=stage,
)
# Format the result to the format the metric expects.
if args.version_2_with_negative:
formatted_predictions = [{
"id":
k,
"prediction_text":
v,
"no_answer_probability":
scores_diff_json[k]
} for k, v in predictions.items()]
else:
formatted_predictions = [{
"id": k,
"prediction_text": v
} for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex[answer_column_name]
} for ex in examples]
return EvalPrediction(predictions=formatted_predictions,
label_ids=references)
metric = load_metric(
"squad_v2" if args.version_2_with_negative else "squad")
def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
"""
Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
Args:
start_or_end_logits(:obj:`tensor`):
This is the output predictions of the model. We can only enter either start or end logits.
eval_dataset: Evaluation dataset
max_len(:obj:`int`):
The maximum length of the output tensor. ( See the model.eval() part for more details )
"""
step = 0
# create a numpy array and fill it with -100.
logits_concat = np.full((len(dataset), max_len),
-100,
dtype=np.float32)
# Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather
for i, output_logit in enumerate(
start_or_end_logits): # populate columns
# We have to fill it such that we have to take the whole tensor and replace it on the newly created array
# And after every iteration we have to change the step
batch_size = output_logit.shape[0]
cols = output_logit.shape[1]
if step + batch_size < len(dataset):
logits_concat[step:step + batch_size, :cols] = output_logit
else:
logits_concat[step:, :cols] = output_logit[:len(dataset) -
step]
step += batch_size
return logits_concat
# 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,
)
# 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
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)
# intialize all lists to collect the batches
all_start_top_log_probs = []
all_start_top_index = []
all_end_top_log_probs = []
all_end_top_index = []
all_cls_logits = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch)
start_top_log_probs = outputs.start_top_log_probs
start_top_index = outputs.start_top_index
end_top_log_probs = outputs.end_top_log_probs
end_top_index = outputs.end_top_index
cls_logits = outputs.cls_logits
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
start_top_log_probs = accelerator.pad_across_processes(
start_top_log_probs, dim=1, pad_index=-100)
start_top_index = accelerator.pad_across_processes(
start_top_index, dim=1, pad_index=-100)
end_top_log_probs = accelerator.pad_across_processes(
end_top_log_probs, dim=1, pad_index=-100)
end_top_index = accelerator.pad_across_processes(
end_top_index, dim=1, pad_index=-100)
cls_logits = accelerator.pad_across_processes(cls_logits,
dim=1,
pad_index=-100)
all_start_top_log_probs.append(
accelerator.gather(start_top_log_probs).cpu().numpy())
all_start_top_index.append(
accelerator.gather(start_top_index).cpu().numpy())
all_end_top_log_probs.append(
accelerator.gather(end_top_log_probs).cpu().numpy())
all_end_top_index.append(
accelerator.gather(end_top_index).cpu().numpy())
all_cls_logits.append(accelerator.gather(cls_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_end_top_log_probs
]) # Get the max_length of the tensor
# concatenate all numpy arrays collected above
start_top_log_probs_concat = create_and_fill_np_array(
all_start_top_log_probs, eval_dataset, max_len)
start_top_index_concat = create_and_fill_np_array(all_start_top_index,
eval_dataset, max_len)
end_top_log_probs_concat = create_and_fill_np_array(
all_end_top_log_probs, eval_dataset, max_len)
end_top_index_concat = create_and_fill_np_array(all_end_top_index,
eval_dataset, max_len)
cls_logits_concat = np.concatenate(all_cls_logits, axis=0)
# delete the list of numpy arrays
del start_top_log_probs
del start_top_index
del end_top_log_probs
del end_top_index
del cls_logits
outputs_numpy = (
start_top_log_probs_concat,
start_top_index_concat,
end_top_log_probs_concat,
end_top_index_concat,
cls_logits_concat,
)
prediction = post_processing_function(eval_examples, eval_dataset,
outputs_numpy)
eval_metric = metric.compute(predictions=prediction.predictions,
references=prediction.label_ids)
logger.info(f"Evaluation metrics: {eval_metric}")
if args.do_predict:
# intialize all lists to collect the batches
all_start_top_log_probs = []
all_start_top_index = []
all_end_top_log_probs = []
all_end_top_index = []
all_cls_logits = []
for step, batch in enumerate(predict_dataloader):
with torch.no_grad():
outputs = model(**batch)
start_top_log_probs = outputs.start_top_log_probs
start_top_index = outputs.start_top_index
end_top_log_probs = outputs.end_top_log_probs
end_top_index = outputs.end_top_index
cls_logits = outputs.cls_logits
if not args.pad_to_max_length: # necessary to pad predictions and labels for being gathered
start_top_log_probs = accelerator.pad_across_processes(
start_top_log_probs, dim=1, pad_index=-100)
start_top_index = accelerator.pad_across_processes(
start_top_index, dim=1, pad_index=-100)
end_top_log_probs = accelerator.pad_across_processes(
end_top_log_probs, dim=1, pad_index=-100)
end_top_index = accelerator.pad_across_processes(
end_top_index, dim=1, pad_index=-100)
cls_logits = accelerator.pad_across_processes(
cls_logits, dim=1, pad_index=-100)
all_start_top_log_probs.append(
accelerator.gather(start_top_log_probs).cpu().numpy())
all_start_top_index.append(
accelerator.gather(start_top_index).cpu().numpy())
all_end_top_log_probs.append(
accelerator.gather(end_top_log_probs).cpu().numpy())
all_end_top_index.append(
accelerator.gather(end_top_index).cpu().numpy())
all_cls_logits.append(
accelerator.gather(cls_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_end_top_log_probs
]) # Get the max_length of the tensor
# concatenate all numpy arrays collected above
start_top_log_probs_concat = create_and_fill_np_array(
all_start_top_log_probs, predict_dataset, max_len)
start_top_index_concat = create_and_fill_np_array(
all_start_top_index, predict_dataset, max_len)
end_top_log_probs_concat = create_and_fill_np_array(
all_end_top_log_probs, predict_dataset, max_len)
end_top_index_concat = create_and_fill_np_array(
all_end_top_index, predict_dataset, max_len)
cls_logits_concat = np.concatenate(all_cls_logits, axis=0)
# delete the list of numpy arrays
del start_top_log_probs
del start_top_index
del end_top_log_probs
del end_top_index
del cls_logits
outputs_numpy = (
start_top_log_probs_concat,
start_top_index_concat,
end_top_log_probs_concat,
end_top_index_concat,
cls_logits_concat,
)
prediction = post_processing_function(predict_examples,
predict_dataset, outputs_numpy)
predict_metric = metric.compute(predictions=prediction.predictions,
references=prediction.label_ids)
logger.info(f"Predict metrics: {predict_metric}")
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, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(
json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses(
)
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_qa_beam_search", model_args, data_args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = training_args.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
logger.info(f"Training/evaluation parameters {training_args}")
# Detecting last checkpoint.
last_checkpoint = None
if os.path.isdir(
training_args.output_dir
) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(
training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome.")
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
extension = data_args.train_file.split(".")[-1]
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
extension = data_args.validation_file.split(".")[-1]
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.test_file.split(".")[-1]
raw_datasets = load_dataset(
extension,
data_files=data_files,
field="data",
cache_dir=model_args.cache_dir,
use_auth_token=True if model_args.use_auth_token else None,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = XLNetConfig.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 = XLNetTokenizerFast.from_pretrained(
model_args.tokenizer_name
if model_args.tokenizer_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,
)
model = XLNetForQuestionAnswering.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,
)
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
if training_args.do_train:
column_names = raw_datasets["train"].column_names
elif training_args.do_eval:
column_names = raw_datasets["validation"].column_names
else:
column_names = raw_datasets["test"].column_names
question_column_name = "question" if "question" in column_names else column_names[
0]
context_column_name = "context" if "context" in column_names else column_names[
1]
answer_column_name = "answers" if "answers" in column_names else column_names[
2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
if data_args.max_seq_length > tokenizer.model_max_length:
logger.warning(
f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
)
max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
# Training preprocessing
def prepare_train_features(examples):
# Some of the questions have lots of whitespace on the left, which is not useful and will make the
# truncation of the context fail (the tokenized question will take a lots of space). So we remove that
# left whitespace
examples[question_column_name] = [
q.lstrip() for q in examples[question_column_name]
]
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
tokenized_examples["is_impossible"] = []
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
# The cls token gets 1.0 too (for predictions of empty answers).
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != context_idx:
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != context_idx:
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char
and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[
token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(
token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(
token_end_index + 1)
tokenized_examples["is_impossible"].append(0.0)
return tokenized_examples
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:
# Select samples from Dataset, This will help to decrease processing time
max_train_samples = min(len(train_dataset),
data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
# Create Training Features
with training_args.main_process_first(
desc="train dataset map pre-processing"):
train_dataset = train_dataset.map(
prepare_train_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on train dataset",
)
if data_args.max_train_samples is not None:
# Select samples from dataset again since Feature Creation might increase number of features
max_train_samples = min(len(train_dataset),
data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
# Validation preprocessing
def prepare_validation_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[
question_column_name if pad_on_right else context_column_name],
examples[
context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
# We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, input_ids in enumerate(tokenized_examples["input_ids"]):
# Find the CLS token in the input ids.
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
tokenized_examples["p_mask"].append([
0.0 if (not special_tokens[i][k] and s == context_idx)
or k == cls_index else 1.0 for k, s in enumerate(sequence_ids)
])
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(
examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_idx else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if training_args.do_eval:
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_examples = raw_datasets["validation"]
if data_args.max_eval_samples is not None:
# Selecting Eval Samples from Dataset
max_eval_samples = min(len(eval_examples),
data_args.max_eval_samples)
eval_examples = eval_examples.select(range(max_eval_samples))
# Create Features from Eval Dataset
with training_args.main_process_first(
desc="validation dataset map pre-processing"):
eval_dataset = eval_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on validation dataset",
)
if data_args.max_eval_samples is not None:
# Selecting Samples from Dataset again since Feature Creation might increase samples size
max_eval_samples = min(len(eval_dataset),
data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
if training_args.do_predict:
if "test" not in raw_datasets:
raise ValueError("--do_predict requires a test dataset")
predict_examples = raw_datasets["test"]
if data_args.max_predict_samples is not None:
# We will select sample from whole data
predict_examples = predict_examples.select(
range(data_args.max_predict_samples))
# Test Feature Creation
with training_args.main_process_first(
desc="prediction dataset map pre-processing"):
predict_dataset = predict_examples.map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on prediction dataset",
)
if data_args.max_predict_samples is not None:
# During Feature creation dataset samples might increase, we will select required samples again
max_predict_samples = min(len(predict_dataset),
data_args.max_predict_samples)
predict_dataset = predict_dataset.select(
range(max_predict_samples))
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
data_collator = (default_data_collator if data_args.pad_to_max_length else
DataCollatorWithPadding(
tokenizer,
pad_to_multiple_of=8 if training_args.fp16 else None))
# Post-processing:
def post_processing_function(examples,
features,
predictions,
stage="eval"):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=data_args.version_2_with_negative,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
start_n_top=model.config.start_n_top,
end_n_top=model.config.end_n_top,
output_dir=training_args.output_dir,
log_level=log_level,
prefix=stage,
)
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [{
"id":
k,
"prediction_text":
v,
"no_answer_probability":
scores_diff_json[k]
} for k, v in predictions.items()]
else:
formatted_predictions = [{
"id": k,
"prediction_text": v
} for k, v in predictions.items()]
references = [{
"id": ex["id"],
"answers": ex[answer_column_name]
} for ex in examples]
return EvalPrediction(predictions=formatted_predictions,
label_ids=references)
metric = evaluate.load(
"squad_v2" if data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions,
references=p.label_ids)
# Initialize our Trainer
trainer = QuestionAnsweringTrainer(
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,
eval_examples=eval_examples if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
post_process_function=post_processing_function,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
elif last_checkpoint is not None:
checkpoint = last_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model() # Saves the tokenizer too for easy upload
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
eval_dataset)
metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Prediction
if training_args.do_predict:
logger.info("*** Predict ***")
results = trainer.predict(predict_dataset, predict_examples)
metrics = results.metrics
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(predict_dataset))
metrics["predict_samples"] = min(max_predict_samples,
len(predict_dataset))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": "question-answering"
}
if data_args.dataset_name is not None:
kwargs["dataset_tags"] = data_args.dataset_name
if data_args.dataset_config_name is not None:
kwargs["dataset_args"] = data_args.dataset_config_name
kwargs[
"dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
else:
kwargs["dataset"] = data_args.dataset_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
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",
)
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/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)
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]
datasets = load_dataset(extension, data_files=data_files, field="data")
# 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 = XLNetConfig.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 = XLNetTokenizerFast.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 = XLNetForQuestionAnswering.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,
)
# Preprocessing the datasets.
# Preprocessing is slighlty different for training and evaluation.
if training_args.do_train:
column_names = datasets["train"].column_names
else:
column_names = datasets["validation"].column_names
question_column_name = "question" if "question" in column_names else column_names[0]
context_column_name = "context" if "context" in column_names else column_names[1]
answer_column_name = "answers" if "answers" in column_names else column_names[2]
# Padding side determines if we do (question|context) or (context|question).
pad_on_right = tokenizer.padding_side == "right"
# Training preprocessing
def prepare_train_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[question_column_name if pad_on_right else context_column_name],
examples[context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=data_args.max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The offset mappings will give us a map from token to character position in the original context. This will
# help us compute the start_positions and end_positions.
offset_mapping = tokenized_examples.pop("offset_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# Let's label those examples!
tokenized_examples["start_positions"] = []
tokenized_examples["end_positions"] = []
tokenized_examples["is_impossible"] = []
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, offsets in enumerate(offset_mapping):
# We will label impossible answers with the index of the CLS token.
input_ids = tokenized_examples["input_ids"][i]
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others get 1.0.
# The cls token gets 1.0 too (for predictions of empty answers).
tokenized_examples["p_mask"].append(
[
0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0
for k, s in enumerate(sequence_ids)
]
)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
answers = examples[answer_column_name][sample_index]
# If no answers are given, set the cls_index as answer.
if len(answers["answer_start"]) == 0:
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Start/end character index of the answer in the text.
start_char = answers["answer_start"][0]
end_char = start_char + len(answers["text"][0])
# Start token index of the current span in the text.
token_start_index = 0
while sequence_ids[token_start_index] != context_idx:
token_start_index += 1
# End token index of the current span in the text.
token_end_index = len(input_ids) - 1
while sequence_ids[token_end_index] != context_idx:
token_end_index -= 1
# Detect if the answer is out of the span (in which case this feature is labeled with the CLS index).
if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char):
tokenized_examples["start_positions"].append(cls_index)
tokenized_examples["end_positions"].append(cls_index)
tokenized_examples["is_impossible"].append(1.0)
else:
# Otherwise move the token_start_index and token_end_index to the two ends of the answer.
# Note: we could go after the last offset if the answer is the last word (edge case).
while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char:
token_start_index += 1
tokenized_examples["start_positions"].append(token_start_index - 1)
while offsets[token_end_index][1] >= end_char:
token_end_index -= 1
tokenized_examples["end_positions"].append(token_end_index + 1)
tokenized_examples["is_impossible"].append(0.0)
return tokenized_examples
if training_args.do_train:
train_dataset = datasets["train"].map(
prepare_train_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Validation preprocessing
def prepare_validation_features(examples):
# Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results
# in one example possible giving several features when a context is long, each of those features having a
# context that overlaps a bit the context of the previous feature.
tokenized_examples = tokenizer(
examples[question_column_name if pad_on_right else context_column_name],
examples[context_column_name if pad_on_right else question_column_name],
truncation="only_second" if pad_on_right else "only_first",
max_length=data_args.max_seq_length,
stride=data_args.doc_stride,
return_overflowing_tokens=True,
return_offsets_mapping=True,
return_special_tokens_mask=True,
return_token_type_ids=True,
padding="max_length",
)
# Since one example might give us several features if it has a long context, we need a map from a feature to
# its corresponding example. This key gives us just that.
sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping")
# The special tokens will help us build the p_mask (which indicates the tokens that can't be in answers).
special_tokens = tokenized_examples.pop("special_tokens_mask")
# For evaluation, we will need to convert our predictions to substrings of the context, so we keep the
# corresponding example_id and we will store the offset mappings.
tokenized_examples["example_id"] = []
# We still provide the index of the CLS token and the p_mask to the model, but not the is_impossible label.
tokenized_examples["cls_index"] = []
tokenized_examples["p_mask"] = []
for i, input_ids in enumerate(tokenized_examples["input_ids"]):
# Find the CLS token in the input ids.
cls_index = input_ids.index(tokenizer.cls_token_id)
tokenized_examples["cls_index"].append(cls_index)
# Grab the sequence corresponding to that example (to know what is the context and what is the question).
sequence_ids = tokenized_examples["token_type_ids"][i]
for k, s in enumerate(special_tokens[i]):
if s:
sequence_ids[k] = 3
context_idx = 1 if pad_on_right else 0
# Build the p_mask: non special tokens and context gets 0.0, the others 1.0.
tokenized_examples["p_mask"].append(
[
0.0 if (not special_tokens[i][k] and s == context_idx) or k == cls_index else 1.0
for k, s in enumerate(sequence_ids)
]
)
# One example can give several spans, this is the index of the example containing this span of text.
sample_index = sample_mapping[i]
tokenized_examples["example_id"].append(examples["id"][sample_index])
# Set to None the offset_mapping that are not part of the context so it's easy to determine if a token
# position is part of the context or not.
tokenized_examples["offset_mapping"][i] = [
(o if sequence_ids[k] == context_idx else None)
for k, o in enumerate(tokenized_examples["offset_mapping"][i])
]
return tokenized_examples
if training_args.do_eval:
validation_dataset = datasets["validation"].map(
prepare_validation_features,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
# We have already padded to max length if the corresponding flag is True, otherwise we need to pad in the data
# collator.
data_collator = default_data_collator if data_args.pad_to_max_length else DataCollatorWithPadding(tokenizer)
# Post-processing:
def post_processing_function(examples, features, predictions):
# Post-processing: we match the start logits and end logits to answers in the original context.
predictions, scores_diff_json = postprocess_qa_predictions_with_beam_search(
examples=examples,
features=features,
predictions=predictions,
version_2_with_negative=data_args.version_2_with_negative,
n_best_size=data_args.n_best_size,
max_answer_length=data_args.max_answer_length,
start_n_top=model.config.start_n_top,
end_n_top=model.config.end_n_top,
output_dir=training_args.output_dir,
is_world_process_zero=trainer.is_world_process_zero(),
)
# Format the result to the format the metric expects.
if data_args.version_2_with_negative:
formatted_predictions = [
{"id": k, "prediction_text": v, "no_answer_probability": scores_diff_json[k]}
for k, v in predictions.items()
]
else:
formatted_predictions = [{"id": k, "prediction_text": v} for k, v in predictions.items()]
references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in datasets["validation"]]
return EvalPrediction(predictions=formatted_predictions, label_ids=references)
# TODO: Once the fix lands in a Datasets release, remove the _local here and the squad_v2_local folder.
current_dir = os.path.sep.join(os.path.join(__file__).split(os.path.sep)[:-1])
metric = load_metric(os.path.join(current_dir, "squad_v2_local") if data_args.version_2_with_negative else "squad")
def compute_metrics(p: EvalPrediction):
return metric.compute(predictions=p.predictions, references=p.label_ids)
# Initialize our Trainer
trainer = QuestionAnsweringTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=validation_dataset if training_args.do_eval else None,
eval_examples=datasets["validation"] if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
post_process_function=post_processing_function,
compute_metrics=compute_metrics,
)
# 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() # Saves the tokenizer too for easy upload
# 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.txt")
if trainer.is_world_process_zero():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key, value in results.items():
logger.info(f" {key} = {value}")
writer.write(f"{key} = {value}\n")
return results
def load(cls,
pretrained_model_name_or_path,
revision=None,
tokenizer_class=None,
use_fast=True,
**kwargs):
"""
Enables loading of different Tokenizer classes with a uniform interface. Either infer the class from
model config or define it manually via `tokenizer_class`.
:param pretrained_model_name_or_path: The path of the saved pretrained model or its name (e.g. `bert-base-uncased`)
:type pretrained_model_name_or_path: str
:param revision: The version of model to use from the HuggingFace model hub. Can be tag name, branch name, or commit hash.
:type revision: str
:param tokenizer_class: (Optional) Name of the tokenizer class to load (e.g. `BertTokenizer`)
:type tokenizer_class: str
:param use_fast: (Optional, False by default) Indicate if FARM should try to load the fast version of the tokenizer (True) or
use the Python one (False).
Only DistilBERT, BERT and Electra fast tokenizers are supported.
:type use_fast: bool
:param kwargs:
:return: Tokenizer
"""
pretrained_model_name_or_path = str(pretrained_model_name_or_path)
kwargs["revision"] = revision
if tokenizer_class is None:
tokenizer_class = cls._infer_tokenizer_class(
pretrained_model_name_or_path)
logger.info(f"Loading tokenizer of type '{tokenizer_class}'")
# return appropriate tokenizer object
ret = None
if "AlbertTokenizer" in tokenizer_class:
if use_fast:
ret = AlbertTokenizerFast.from_pretrained(
pretrained_model_name_or_path, keep_accents=True, **kwargs)
else:
ret = AlbertTokenizer.from_pretrained(
pretrained_model_name_or_path, keep_accents=True, **kwargs)
elif "XLMRobertaTokenizer" in tokenizer_class:
if use_fast:
ret = XLMRobertaTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = XLMRobertaTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "RobertaTokenizer" in tokenizer_class:
if use_fast:
ret = RobertaTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = RobertaTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "DistilBertTokenizer" in tokenizer_class:
if use_fast:
ret = DistilBertTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = DistilBertTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "BertTokenizer" in tokenizer_class:
if use_fast:
ret = BertTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = BertTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "XLNetTokenizer" in tokenizer_class:
if use_fast:
ret = XLNetTokenizerFast.from_pretrained(
pretrained_model_name_or_path, keep_accents=True, **kwargs)
else:
ret = XLNetTokenizer.from_pretrained(
pretrained_model_name_or_path, keep_accents=True, **kwargs)
elif "ElectraTokenizer" in tokenizer_class:
if use_fast:
ret = ElectraTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = ElectraTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif tokenizer_class == "EmbeddingTokenizer":
if use_fast:
logger.error(
'EmbeddingTokenizerFast is not supported! Using EmbeddingTokenizer instead.'
)
ret = EmbeddingTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = EmbeddingTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "CamembertTokenizer" in tokenizer_class:
if use_fast:
ret = CamembertTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = CamembertTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "DPRQuestionEncoderTokenizer" in tokenizer_class:
if use_fast:
ret = DPRQuestionEncoderTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = DPRQuestionEncoderTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
elif "DPRContextEncoderTokenizer" in tokenizer_class:
if use_fast:
ret = DPRContextEncoderTokenizerFast.from_pretrained(
pretrained_model_name_or_path, **kwargs)
else:
ret = DPRContextEncoderTokenizer.from_pretrained(
pretrained_model_name_or_path, **kwargs)
if ret is None:
raise Exception("Unable to load tokenizer")
else:
return ret
'confusiton_mat': confusion_mat
}
for k in out_dict.keys():
print(k)
print(out_dict[k])
data_read = read_dataset("..\preprocess_data\with_tags.csv")
token_docs = data_read[0]
tag_docs = data_read[1]
train_texts, val_texts, train_tags, val_tags = train_test_split(
token_docs, tag_docs, test_size=.2, random_state=42)
#setting the tokenizer
tokenizer = XLNetTokenizerFast.from_pretrained('xlnet-base-cased')
train_encodings = tokenizer(train_texts,
is_split_into_words=True,
return_offsets_mapping=True,
padding=True,
truncation=True,
max_length=30)
val_encodings = tokenizer(val_texts,
is_split_into_words=True,
return_offsets_mapping=True,
padding=True,
truncation=True,
max_length=30)
train_labels = encode_tags(train_tags, train_encodings)