def _load_model(self):
self.model = AutoModelForCTC.from_pretrained(self.model_path)
self.model.to(self.device)
try:
self.processor = Wav2Vec2Processor.from_pretrained(self.model_path)
self.token_set = TokenSet.from_processor(self.processor)
except Exception:
logger.warning("Not fine-tuned model! You'll need to fine-tune it before use this model for audio transcription")
self.processor = None
self.token_set = 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)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
task_name = data_args.task
lang_id = data_args.language
if task_name is None:
raise ValueError("Set --task should be set to '<xtreme_s_task>' "
"(e.g. 'fleurs-asr', 'mls', 'covost2', 'minds14') ")
if lang_id is None:
raise ValueError(
"Set --language should be set to the language id of the sub dataset "
"config to be used (e.g. 'pl', 'en.tr', 'fr-FR') or 'all'"
" for multi-lingual fine-tuning.")
if data_args.target_column_name is None:
target_column_name = TASK_TO_TARGET_COLUMN_NAME[task_name]
else:
target_column_name = data_args.target_column_name
# here we differentiate between tasks with text as the target and classification tasks
is_text_target = target_column_name in ("transcription", "translation")
config_name = ".".join([task_name.split("-")[0], lang_id])
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.audio_column_name not in raw_datasets[
"train"].column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
if target_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--target_column_name {target_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--target_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(
range(data_args.max_train_samples))
if not is_text_target:
label_list = raw_datasets["train"].features[
target_column_name].names
num_labels = len(label_list)
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(
range(data_args.max_eval_samples))
if training_args.do_predict:
raw_datasets["predict"] = load_dataset(
data_args.dataset_name,
config_name,
split=data_args.predict_split_name,
use_auth_token=data_args.use_auth_token,
cache_dir=model_args.cache_dir,
)
if data_args.max_predict_samples is not None:
raw_datasets["predict"] = raw_datasets["predict"].select(
range(data_args.max_predict_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (f'[{"".join(data_args.chars_to_ignore)}]' if
data_args.chars_to_ignore is not None else None)
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[target_column_name]).lower() + " "
else:
batch["target_text"] = batch[target_column_name].lower() + " "
return batch
if is_text_target:
with training_args.main_process_first(
desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[target_column_name],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk_token = data_args.unk_token
pad_token = data_args.pad_token
# 3. Next, let's load the config as we might need it to create
# the tokenizer
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
if is_text_target:
# 4. (Optional, for ASR and translation) If no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kwargs = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.output_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_output_dir and os.path.isfile(
vocab_file):
os.remove(vocab_file)
with training_args.main_process_first(
desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk_token=unk_token,
pad_token=pad_token,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
if not config.is_encoder_decoder:
tokenizer_kwargs = {
"config":
config if config.tokenizer_class is not None else None,
"tokenizer_type":
config.model_type
if config.tokenizer_class is None else None,
"unk_token":
unk_token,
"pad_token":
pad_token,
"word_delimiter_token":
word_delimiter_token,
}
else:
tokenizer_kwargs = {}
# 5. Now we can instantiate the feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load feature_extractor and tokenizer
if is_text_target:
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_auth_token=data_args.use_auth_token,
**tokenizer_kwargs,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
# adapt config
config.update({
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"activation_dropout": model_args.activation_dropout,
})
if training_args.do_train:
if is_text_target:
config.pad_token_id = tokenizer.pad_token_id
config.vocab_size = len(tokenizer)
else:
label_to_id = {v: i for i, v in enumerate(label_list)}
config.label2id = label_to_id
config.id2label = {id: label for label, id in label_to_id.items()}
config.num_labels = num_labels
# create model
if target_column_name == "transcription":
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
elif config.is_encoder_decoder:
model = AutoModelForSpeechSeq2Seq.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
if model.config.decoder_start_token_id is None:
raise ValueError(
"Make sure that `config.decoder_start_token_id` is correctly defined"
)
else:
model = AutoModelForAudioClassification.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(raw_datasets.values())).features[
data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name,
datasets.features.Audio(
sampling_rate=feature_extractor.sampling_rate))
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
num_workers = data_args.preprocessing_num_workers
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"],
sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
if is_text_target:
batch["labels"] = tokenizer(batch["target_text"],
**additional_kwargs).input_ids
else:
batch["labels"] = batch[target_column_name]
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
if training_args.do_train:
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets["train"] = vectorized_datasets["train"].filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["length"],
)
# 7. Next, we can prepare for the training step.
# Let's use the appropriate XTREME-S evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metric = load_metric("xtreme_s", task_name)
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(
f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
)
return
def compute_asr_metric(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metric = eval_metric.compute(predictions=pred_str,
references=label_str)
return metric
def compute_classification_metric(pred):
pred_ids = np.argmax(pred.predictions, axis=1)
metric = eval_metric.compute(predictions=pred_ids,
references=pred.label_ids)
return metric
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
if is_text_target:
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
# wait until configs are saved in the main process before loading the processor
torch.distributed.barrier()
if is_text_target:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
else:
processor = AutoFeatureExtractor.from_pretrained(
training_args.output_dir)
# Instantiate custom data collator
data_collator = SpeechDataCollatorWithPadding(processor=processor,
pad_labels=is_text_target)
# Initialize Trainer
if target_column_name == "translation":
trainer = Seq2SeqTrainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_asr_metric
if training_args.predict_with_generate else None,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=feature_extractor,
)
else:
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_asr_metric
if is_text_target else compute_classification_metric,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=feature_extractor,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
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()
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(vectorized_datasets["train"]))
metrics["train_samples"] = min(max_train_samples,
len(vectorized_datasets["train"]))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation on the test set
results = {}
if training_args.do_predict:
logger.info(
f"*** Evaluating on the `{data_args.predict_split_name}` set ***")
metrics = trainer.evaluate(vectorized_datasets["predict"])
max_predict_samples = (data_args.max_predict_samples
if data_args.max_predict_samples is not None
else len(vectorized_datasets["predict"]))
metrics["predict_samples"] = min(max_predict_samples,
len(vectorized_datasets["predict"]))
trainer.log_metrics("predict", metrics)
trainer.save_metrics("predict", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from": model_args.model_name_or_path,
"tasks": task_name,
"tags": [task_name, data_args.dataset_name],
"dataset_args":
f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}, Predict split: {data_args.predict_split_name}",
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
"language": data_args.language,
}
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, 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(
)
# 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)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
raw_datasets["train"] = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name)
raw_datasets["eval"] = load_dataset(data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name)
if data_args.audio_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--audio_column_name {data_args.audio_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column_name` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
if data_args.text_column_name not in raw_datasets["train"].column_names:
raise ValueError(
f"--text_column_name {data_args.audio_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
# prepare dataset
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(
range(data_args.max_train_samples))
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(
range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (f'[{"".join(data_args.chars_to_ignore)}]' if
data_args.chars_to_ignore is not None else None)
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[data_args.text_column_name]).lower() + " "
else:
batch["target_text"] = batch[
data_args.text_column_name].lower() + " "
return batch
with training_args.main_process_first(
desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[data_args.text_column_name],
desc="remove special characters from datasets",
)
# 3. Next, we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
with training_args.main_process_first(
desc="dataset map vocabulary creation"):
vocab_dict = create_vocabulary_from_data(raw_datasets)
vocab_file = os.path.join(training_args.output_dir, "vocab.json")
# save vocab dict to be loaded into tokenizer
os.makedirs(training_args.output_dir, exist_ok=True)
if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
os.remove(vocab_file)
if not os.path.isfile(vocab_file):
with open(vocab_file, "w") as vocab_file:
json.dump(vocab_dict, vocab_file)
# 4. Now we can instantiate the configuration, feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load config
config = AutoConfig.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir)
# load feature_extractor, tokenizer and create processor
tokenizer = AutoTokenizer.from_pretrained(
training_args.output_dir,
tokenizer_type=config.model_type,
unk_token="[UNK]",
pad_token="[PAD]",
word_delimiter_token="|",
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path, cache_dir=model_args.cache_dir)
processor = Wav2Vec2Processor(feature_extractor=feature_extractor,
tokenizer=tokenizer)
# adapt config
config.update({
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": processor.tokenizer.pad_token_id,
"vocab_size": len(processor.tokenizer),
"activation_dropout": model_args.activation_dropout,
})
# create model
model = AutoModelForCTC.from_pretrained(model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config)
# freeze encoder
if model_args.freeze_feature_extractor:
model.freeze_feature_extractor()
# 5. Now we preprocess the datasets which includes loading the audio, resampling and padding
# The following code should be cleaned up as soon as
# https://github.com/huggingface/datasets/pull/2324 is merged
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * processor.feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * processor.feature_extractor.sampling_rate
resampler = None
if raw_datasets["train"][data_args.audio_column_name][0].split(
".")[-1] == "mp3":
# TODO(PVP) - remove hard-coded 48_000 after audio feature is merged
resampler = torchaudio.transforms.Resample(
48_000, processor.feature_extractor.sampling_rate)
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
speech_array, sampling_rate = torchaudio.load(
batch[data_args.audio_column_name])
speech_array = speech_array.squeeze()
# if necessary resample audio
if resampler is not None:
# TODO(PVP) - remove hard-coded 48_000 after audio feature is merged
speech_array = resampler(speech_array)
sampling_rate = resampler.new_freq
speech_array = speech_array.numpy()
batch["input_values"] = processor(
speech_array,
sampling_rate=sampling_rate,
truncate=True,
max_length=max_input_length).input_values[0]
# Setup the processor for targets
with processor.as_target_processor():
batch["labels"] = processor(batch["target_text"]).input_ids
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=raw_datasets["train"].column_names,
num_proc=data_args.preprocessing_num_workers,
desc="preprocess datasets",
)
if min_input_length > 0.0:
# filter data that is shorter than min_input_length
vectorized_datasets = vectorized_datasets.filter(
lambda data: len(data["input_values"]) > min_input_length,
num_proc=data_args.preprocessing_num_workers,
)
# 6. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define Metric during training
wer_metric = load_metric("wer")
if data_args.only_data_preprocessing:
logger.info("Data preprocessing finished.")
return
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids ==
-100] = processor.tokenizer.pad_token_id
pred_str = processor.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
wer = wer_metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
# Instantiate custom data collator
data_collator = DataCollatorCTCWithPadding(processor=processor)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=processor.feature_extractor,
)
# 7. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
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
# Save the feature_extractor and the tokenizer
if is_main_process(training_args.local_rank):
processor.save_pretrained(training_args.output_dir)
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(vectorized_datasets["train"]))
metrics["train_samples"] = min(max_train_samples,
len(vectorized_datasets["train"]))
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_eval_samples = (data_args.max_eval_samples
if data_args.max_eval_samples is not None else len(
vectorized_datasets["eval"]))
metrics["eval_samples"] = min(max_eval_samples,
len(vectorized_datasets["eval"]))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
kwargs = {
"finetuned_from":
model_args.model_name_or_path,
"tasks":
"speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args":
f"Config: {data_args.dataset_config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
"dataset":
f"{data_args.dataset_name.upper()} - {data_args.dataset_config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kwargs["language"] = data_args.dataset_config_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, 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(
)
# 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)
# 1. First, let's load the dataset
raw_datasets = IterableDatasetDict()
raw_column_names = {}
def load_streaming_dataset(split, sampling_rate, **kwargs):
if "+" in split:
dataset_splits = [
load_dataset(split=split_name, **kwargs)
for split_name in split.split("+")
]
# `features` and `cast_column` won't be available after interleaving, so we'll use them here
features = dataset_splits[0].features
# make sure that the dataset decodes audio with a correct sampling rate
dataset_splits = [
dataset.cast_column(
data_args.audio_column_name,
datasets.features.Audio(sampling_rate=sampling_rate))
for dataset in dataset_splits
]
interleaved_dataset = interleave_datasets(dataset_splits)
return interleaved_dataset, features
else:
dataset = load_dataset(split=split, **kwargs)
features = dataset.features
# make sure that the dataset decodes audio with a correct sampling rate
dataset = dataset.cast_column(
data_args.audio_column_name,
datasets.features.Audio(sampling_rate=sampling_rate))
return dataset, features
# `datasets` takes care of automatically loading and resampling the audio,
# so we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
if training_args.do_train:
raw_datasets["train"], train_features = load_streaming_dataset(
path=data_args.dataset_name,
name=data_args.dataset_config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
streaming=True,
sampling_rate=feature_extractor.sampling_rate,
)
raw_column_names["train"] = list(train_features.keys())
if data_args.audio_column_name not in raw_column_names["train"]:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'."
" Make sure to set `--audio_column_name` to the correct audio column - one of"
f" {', '.join(raw_column_names['train'])}.")
if data_args.text_column_name not in raw_column_names["train"]:
raise ValueError(
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(raw_column_names['train'])}.")
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].take(
range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"], eval_features = load_streaming_dataset(
path=data_args.dataset_name,
name=data_args.dataset_config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
streaming=True,
sampling_rate=feature_extractor.sampling_rate,
)
raw_column_names["eval"] = list(eval_features.keys())
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].take(
range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (f'[{"".join(data_args.chars_to_ignore)}]' if
data_args.chars_to_ignore is not None else None)
text_column_name = data_args.text_column_name
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[text_column_name]).lower() + " "
else:
batch["target_text"] = batch[text_column_name].lower() + " "
return batch
with training_args.main_process_first(
desc="dataset map special characters removal"):
for split, dataset in raw_datasets.items():
raw_datasets[split] = dataset.map(
remove_special_characters, ).remove_columns([text_column_name])
# 3. Next, let's load the config as we might need it to create
# the tokenizer
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
# 4. Now we can instantiate the tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
tokenizer_name_or_path = model_args.tokenizer_name_or_path
if tokenizer_name_or_path is None:
raise ValueError(
"Tokenizer has to be created before training in streaming mode. Please specify --tokenizer_name_or_path"
)
# load feature_extractor and tokenizer
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
config=config,
use_auth_token=data_args.use_auth_token,
)
# adapt config
config.update({
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": tokenizer.pad_token_id,
"vocab_size": len(tokenizer),
"activation_dropout": model_args.activation_dropout,
})
# create model
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 5. Now we preprocess the datasets including loading the audio, resampling and normalization
audio_column_name = data_args.audio_column_name
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"],
sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
batch["labels"] = tokenizer(batch["target_text"],
**additional_kwargs).input_ids
return batch
vectorized_datasets = IterableDatasetDict()
with training_args.main_process_first(desc="dataset map preprocessing"):
for split, dataset in raw_datasets.items():
vectorized_datasets[split] = (
dataset.map(prepare_dataset).remove_columns(
raw_column_names[split] +
["target_text"]).with_format("torch"))
if split == "train":
vectorized_datasets[split] = vectorized_datasets[
split].shuffle(
buffer_size=data_args.shuffle_buffer_size,
seed=training_args.seed,
)
# 6. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metrics = {
metric: load_metric(metric)
for metric in data_args.eval_metrics
}
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metrics = {
k: v.compute(predictions=pred_str, references=label_str)
for k, v in eval_metrics.items()
}
return metrics
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
try:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
except (OSError, KeyError):
warnings.warn(
"Loading a processor from a feature extractor config that does not"
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
" attribute to your `preprocessor_config.json` file to suppress this warning: "
" `'processor_class': 'Wav2Vec2Processor'`",
FutureWarning,
)
processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
# Instantiate custom data collator
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
data_collator = DataCollatorCTCWithPadding(processor=processor,
max_length=max_input_length)
# trainer callback to reinitialize and reshuffle the streamable datasets at the beginning of each epoch
class ShuffleCallback(TrainerCallback):
def on_epoch_begin(self, args, state, control, train_dataloader,
**kwargs):
if isinstance(train_dataloader.dataset, IterableDatasetShard):
pass # set_epoch() is handled by the Trainer
elif isinstance(train_dataloader.dataset, IterableDataset):
train_dataloader.dataset.set_epoch(
train_dataloader.dataset._epoch + 1)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=processor,
callbacks=[ShuffleCallback()],
)
# 7. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
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()
metrics = train_result.metrics
if data_args.max_train_samples:
metrics["train_samples"] = data_args.max_train_samples
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()
if data_args.max_eval_samples:
metrics["eval_samples"] = data_args.max_eval_samples
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
kwargs = {
"finetuned_from":
model_args.model_name_or_path,
"tasks":
"speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args":
(f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split:"
f" {data_args.eval_split_name}"),
"dataset":
f"{data_args.dataset_name.upper()} - {config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kwargs["language"] = config_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, 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_speech_recognition_ctc", model_args, data_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:
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)
# 1. First, let's load the dataset
raw_datasets = DatasetDict()
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.audio_column_name not in raw_datasets[
"train"].column_names:
raise ValueError(
f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'."
" Make sure to set `--audio_column_name` to the correct audio column - one of"
f" {', '.join(raw_datasets['train'].column_names)}.")
if data_args.text_column_name not in raw_datasets[
"train"].column_names:
raise ValueError(
f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column_name` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}.")
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(
range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(
range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (f'[{"".join(data_args.chars_to_ignore)}]' if
data_args.chars_to_ignore is not None else None)
text_column_name = data_args.text_column_name
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = re.sub(
chars_to_ignore_regex, "",
batch[text_column_name]).lower() + " "
else:
batch["target_text"] = batch[text_column_name].lower() + " "
return batch
with training_args.main_process_first(
desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[text_column_name],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk_token = data_args.unk_token
pad_token = data_args.pad_token
# 3. Next, let's load the config as we might need it to create
# the tokenizer
# load config
config = AutoConfig.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
# 4. Next, if no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kwargs = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.output_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_output_dir and os.path.isfile(
vocab_file):
try:
os.remove(vocab_file)
except OSError:
# in shared file-systems it might be the case that
# two processes try to delete the vocab file at the some time
pass
with training_args.main_process_first(
desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk_token=unk_token,
pad_token=pad_token,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
tokenizer_kwargs = {
"config":
config if config.tokenizer_class is not None else None,
"tokenizer_type":
config.model_type if config.tokenizer_class is None else None,
"unk_token":
unk_token,
"pad_token":
pad_token,
"word_delimiter_token":
word_delimiter_token,
}
# 5. Now we can instantiate the feature extractor, tokenizer and model
# Note for distributed training, the .from_pretrained methods guarantee that only
# one local process can concurrently download model & vocab.
# load feature_extractor and tokenizer
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_auth_token=data_args.use_auth_token,
**tokenizer_kwargs,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token)
# adapt config
config.update({
"feat_proj_dropout": model_args.feat_proj_dropout,
"attention_dropout": model_args.attention_dropout,
"hidden_dropout": model_args.hidden_dropout,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"gradient_checkpointing": training_args.gradient_checkpointing,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"pad_token_id": tokenizer.pad_token_id,
"vocab_size": len(tokenizer),
"activation_dropout": model_args.activation_dropout,
})
# create model
model = AutoModelForCTC.from_pretrained(
model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = next(iter(raw_datasets.values())).features[
data_args.audio_column_name].sampling_rate
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column_name,
datasets.features.Audio(
sampling_rate=feature_extractor.sampling_rate))
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
audio_column_name = data_args.audio_column_name
num_workers = data_args.preprocessing_num_workers
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column_name]
inputs = feature_extractor(sample["array"],
sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# encode targets
additional_kwargs = {}
if phoneme_language is not None:
additional_kwargs["phonemizer_lang"] = phoneme_language
batch["labels"] = tokenizer(batch["target_text"],
**additional_kwargs).input_ids
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets = vectorized_datasets.filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["input_length"],
)
# 7. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metrics = {
metric: evaluate.load(metric)
for metric in data_args.eval_metrics
}
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(
f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
)
return
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metrics = {
k: v.compute(predictions=pred_str, references=label_str)
for k, v in eval_metrics.items()
}
return metrics
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.output_dir)
tokenizer.save_pretrained(training_args.output_dir)
config.save_pretrained(training_args.output_dir)
try:
processor = AutoProcessor.from_pretrained(training_args.output_dir)
except (OSError, KeyError):
warnings.warn(
"Loading a processor from a feature extractor config that does not"
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
" attribute to your `preprocessor_config.json` file to suppress this warning: "
" `'processor_class': 'Wav2Vec2Processor'`",
FutureWarning,
)
processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
# Instantiate custom data collator
data_collator = DataCollatorCTCWithPadding(processor=processor)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"]
if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"]
if training_args.do_eval else None,
tokenizer=feature_extractor,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
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()
metrics = train_result.metrics
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(vectorized_datasets["train"]))
metrics["train_samples"] = min(max_train_samples,
len(vectorized_datasets["train"]))
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_eval_samples = (data_args.max_eval_samples
if data_args.max_eval_samples is not None else len(
vectorized_datasets["eval"]))
metrics["eval_samples"] = min(max_eval_samples,
len(vectorized_datasets["eval"]))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
kwargs = {
"finetuned_from":
model_args.model_name_or_path,
"tasks":
"speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args":
(f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split:"
f" {data_args.eval_split_name}"),
"dataset":
f"{data_args.dataset_name.upper()} - {config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kwargs["language"] = config_name
if training_args.push_to_hub:
trainer.push_to_hub(**kwargs)
else:
trainer.create_model_card(**kwargs)
return results
def main():
raw_datasets = DatasetDict()
if training_args.do_train:
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config,
split=data_args.train_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.audio_column not in raw_datasets["train"].column_names:
raise ValueError(
f"--audio_column '{data_args.audio_column}' not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--audio_column` to the correct audio column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.text_column not in raw_datasets["train"].column_names:
raise ValueError(
f"--text_column {data_args.text_column} not found in dataset '{data_args.dataset_name}'. "
"Make sure to set `--text_column` to the correct text column - one of "
f"{', '.join(raw_datasets['train'].column_names)}."
)
if data_args.max_train_samples is not None:
raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
if training_args.do_eval:
raw_datasets["eval"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config,
split=data_args.eval_split_name,
use_auth_token=data_args.use_auth_token,
)
if data_args.max_eval_samples is not None:
raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
# 2. We remove some special characters from the datasets
# that make training complicated and do not help in transcribing the speech
# E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
# that could be easily picked up by the model
chars_to_ignore_regex = (
f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
)
text_column = data_args.text_column
def remove_special_characters(batch):
if chars_to_ignore_regex is not None:
batch["target_text"] = (
re.sub(chars_to_ignore_regex, "", batch[text_column]).lower() + " "
)
else:
batch["target_text"] = batch[text_column].lower() + " "
return batch
with training_args.main_process_first(desc="dataset map special characters removal"):
raw_datasets = raw_datasets.map(
remove_special_characters,
remove_columns=[text_column],
desc="remove special characters from datasets",
)
# save special tokens for tokenizer
word_delimiter_token = data_args.word_delimiter_token
unk = data_args.unk
pad = data_args.pad
# 3. Next, let's load the config as we might need it to create
# the tokenizer
# load config
config = AutoConfig.from_pretrained(
model_args.model_name,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token,
)
# 4. Next, if no tokenizer file is defined,
# we create the vocabulary of the model by extracting all unique characters from
# the training and evaluation datasets
# We need to make sure that only first rank saves vocabulary
# make sure all processes wait until vocab is created
tokenizer_name_or_path = model_args.tokenizer_name_or_path
tokenizer_kw = {}
if tokenizer_name_or_path is None:
# save vocab in training output dir
tokenizer_name_or_path = training_args.out_dir
vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
with training_args.main_process_first():
if training_args.overwrite_out_dir and os.path.isfile(vocab_file):
os.remove(vocab_file)
with training_args.main_process_first(desc="dataset map vocabulary creation"):
if not os.path.isfile(vocab_file):
os.makedirs(tokenizer_name_or_path, exist_ok=True)
vocab_dict = create_vocabulary_from_data(
raw_datasets,
word_delimiter_token=word_delimiter_token,
unk=unk,
pad=pad,
)
# save vocab dict to be loaded into tokenizer
with open(vocab_file, "w") as file:
json.dump(vocab_dict, file)
# if tokenizer has just been created
# it is defined by `tokenizer_class` if present in config else by `model_type`
tokenizer_kw = {
"config": config if config.tokenizer_class is not None else None,
"tokenizer_type": config.model_type if config.tokenizer_class is None else None,
"unk": unk,
"pad": pad,
"word_delimiter_token": word_delimiter_token,
}
tokenizer = AutoTokenizer.from_pretrained(
tokenizer_name_or_path,
use_auth_token=data_args.use_auth_token,
**tokenizer_kw,
)
feature_extractor = AutoFeatureExtractor.from_pretrained(
model_args.model_name,
cache_dir=model_args.cache_dir,
use_auth_token=data_args.use_auth_token,
)
# adapt config
config.update(
{
"feat_proj_dropout": model_args.feat_proj_dropout,
"drop_attn": model_args.drop_attn,
"drop": model_args.drop,
"final_dropout": model_args.final_dropout,
"mask_time_prob": model_args.mask_time_prob,
"mask_time_length": model_args.mask_time_length,
"mask_feature_prob": model_args.mask_feature_prob,
"mask_feature_length": model_args.mask_feature_length,
"grad_checkpoint": training_args.grad_checkpoint,
"layerdrop": model_args.layerdrop,
"ctc_loss_reduction": model_args.ctc_loss_reduction,
"PAD": tokenizer.PAD,
"s_vocab": len(tokenizer),
"drop_act": model_args.drop_act,
}
)
# create model
model = AutoModelForCTC.from_pretrained(
model_args.model_name,
cache_dir=model_args.cache_dir,
config=config,
use_auth_token=data_args.use_auth_token,
)
# freeze encoder
if model_args.freeze_feature_encoder:
model.freeze_feature_encoder()
# 6. Now we preprocess the datasets including loading the audio, resampling and normalization
# Thankfully, `datasets` takes care of automatically loading and resampling the audio,
# so that we just need to set the correct target sampling rate and normalize the input
# via the `feature_extractor`
# make sure that dataset decodes audio with correct sampling rate
dataset_sampling_rate = (
next(iter(raw_datasets.values())).features[data_args.audio_column].sampling_rate
)
if dataset_sampling_rate != feature_extractor.sampling_rate:
raw_datasets = raw_datasets.cast_column(
data_args.audio_column,
datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate),
)
# derive max & min input length for sample rate & max duration
max_input_length = data_args.max_duration * feature_extractor.sampling_rate
min_input_length = data_args.min_duration * feature_extractor.sampling_rate
audio_column = data_args.audio_column
num_workers = data_args.num_workers
# `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
phoneme_language = data_args.phoneme_language
# Preprocessing the datasets.
# We need to read the audio files as arrays and tokenize the targets.
def prepare_dataset(batch):
# load audio
sample = batch[audio_column]
inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
batch["input_values"] = inputs.input_values[0]
batch["input_length"] = len(batch["input_values"])
# encode targets
additional_kw = {}
if phoneme_language is not None:
additional_kw["phonemizer_lang"] = phoneme_language
batch["labels"] = tokenizer(batch["target_text"], **additional_kw).input_ids
return batch
with training_args.main_process_first(desc="dataset map preprocessing"):
vectorized_datasets = raw_datasets.map(
prepare_dataset,
remove_columns=next(iter(raw_datasets.values())).column_names,
num_proc=num_workers,
desc="preprocess datasets",
)
def is_audio_in_length_range(length):
return length > min_input_length and length < max_input_length
# filter data that is shorter than min_input_length
vectorized_datasets = vectorized_datasets.filter(
is_audio_in_length_range,
num_proc=num_workers,
input_columns=["input_length"],
)
# 7. Next, we can prepare the training.
# Let's use word error rate (WER) as our evaluation metric,
# instantiate a data collator and the trainer
# Define evaluation metrics during training, *i.e.* word error rate, character error rate
eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
# for large datasets it is advised to run the preprocessing on a
# single machine first with ``args.preprocessing_only`` since there will mostly likely
# be a timeout when running the script in distributed mode.
# In a second step ``args.preprocessing_only`` can then be set to `False` to load the
# cached dataset
if data_args.preprocessing_only:
logger.info(
f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
)
return
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = tokenizer.PAD
pred_str = tokenizer.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
metrics = {
k: v.compute(predictions=pred_str, references=label_str)
for k, v in eval_metrics.items()
}
return metrics
# Now save everything to be able to create a single processor later
if is_main_process(training_args.local_rank):
# save feature extractor, tokenizer and config
feature_extractor.save_pretrained(training_args.out_dir)
tokenizer.save_pretrained(training_args.out_dir)
config.save_pretrained(training_args.out_dir)
try:
processor = AutoProcessor.from_pretrained(training_args.out_dir)
except (OSError, KeyError):
warnings.warn(
"Loading a processor from a feature extractor config that does not"
" include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
" attribute to your `preprocessor_config.json` file to suppress this warning: "
" `'processor_class': 'Wav2Vec2Processor'`",
FutureWarning,
)
processor = Wav2Vec2Processor.from_pretrained(training_args.out_dir)
# Instantiate custom data collator
data_collator = DataCollatorCTCWithPadding(processor=processor)
# Initialize Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
tokenizer=feature_extractor,
)
# 8. Finally, we can start training
# Training
if training_args.do_train:
# use last checkpoint if exist
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name):
checkpoint = model_args.model_name
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
trainer.save_model()
metrics = train_result.metrics
max_train_samples = (
data_args.max_train_samples
if data_args.max_train_samples is not None
else len(vectorized_datasets["train"])
)
metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
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_eval_samples = (
data_args.max_eval_samples
if data_args.max_eval_samples is not None
else len(vectorized_datasets["eval"])
)
metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Write model card and (optionally) push to hub
config_name = data_args.dataset_config if data_args.dataset_config is not None else "na"
kw = {
"finetuned_from": model_args.model_name,
"tasks": "speech-recognition",
"tags": ["automatic-speech-recognition", data_args.dataset_name],
"dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
"dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
}
if "common_voice" in data_args.dataset_name:
kw["language"] = config_name
if training_args.push_to_hub:
trainer.push_to_hub(**kw)
else:
trainer.create_model_card(**kw)
return results
def __init__(self, model_name, hotwords=[]):
self.processor = AutoProcessor.from_pretrained(model_name)
self.model = AutoModelForCTC.from_pretrained(model_name)
self.hotwords = hotwords