run_whisper_finetuning.py

#!/usr/bin/env python
# coding=utf-8
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and

""" Fine-tuning a 🤗 Transformers Whisper model for automatic speech recognition"""

import functools
import json
import logging
import os
import re
import sys
import warnings
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Union
import evaluate

import numpy as np
import torch
from pprint import pprint
import evaluate
from datasets import DatasetDict, load_dataset
from datasets import Audio

from transformers import (
    HfArgumentParser,
    TrainingArguments,
    set_seed,
    WhisperFeatureExtractor,
    WhisperTokenizer,
    WhisperForConditionalGeneration,
    WhisperProcessor,
    Seq2SeqTrainer,
    Seq2SeqTrainingArguments,
)
from transformers.trainer_utils import get_last_checkpoint, is_main_process
from transformers.utils import check_min_version
from transformers.utils.versions import require_version


def list_field(default=None, metadata=None):
    return field(default_factory=lambda: default, metadata=metadata)


@dataclass
class Seq2SeqTrainingArguments(TrainingArguments):
    """
    Args:
        sortish_sampler (`bool`, *optional*, defaults to `False`):
            Whether to use a *sortish sampler* or not. Only possible if the underlying datasets are *Seq2SeqDataset*
            for now but will become generally available in the near future.
            It sorts the inputs according to lengths in order to minimize the padding size, with a bit of randomness
            for the training set.
        predict_with_generate (`bool`, *optional*, defaults to `False`):
            Whether to use generate to calculate generative metrics (ROUGE, BLEU).
        generation_max_length (`int`, *optional*):
            The `max_length` to use on each evaluation loop when `predict_with_generate=True`. Will default to the
            `max_length` value of the model configuration.
        generation_num_beams (`int`, *optional*):
            The `num_beams` to use on each evaluation loop when `predict_with_generate=True`. Will default to the
            `num_beams` value of the model configuration.
    """

    sortish_sampler: bool = field(default=False, metadata={
                                  "help": "Whether to use SortishSampler or not."})
    predict_with_generate: bool = field(
        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
    )
    generation_max_length: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "The `max_length` to use on each evaluation loop when `predict_with_generate=True`. Will default "
                "to the `max_length` value of the model configuration."
            )
        },
    )
    generation_num_beams: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "The `num_beams` to use on each evaluation loop when `predict_with_generate=True`. Will default "
                "to the `num_beams` value of the model configuration."
            )
        },
    )
    xla: bool = field(default=False, metadata={
        "help": "Whether to activate the XLA compilation or not"})


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """

    model_name_or_path: str = field(
        metadata={
            "help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )
    language: str = field(
        metadata={"help": "Whisper specific language"}
    )
    task: str = field(
        metadata={
            "help": "Whisper specific task, i.e., 'transcribe' or 'translate'"}
    )
    tokenizer_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={
            "help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
    )
    freeze_feature_encoder: bool = field(
        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
    )
    attention_dropout: float = field(
        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
    )
    activation_dropout: float = field(
        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
    )
    feat_proj_dropout: float = field(default=0.0, metadata={
                                     "help": "The dropout ratio for the projected features."})
    hidden_dropout: float = field(
        default=0.0,
        metadata={
            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
        },
    )
    final_dropout: float = field(
        default=0.0,
        metadata={
            "help": "The dropout probability for the final projection layer."},
    )
    mask_time_prob: float = field(
        default=0.05,
        metadata={
            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
                    "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
                    "vectors will be masked along the time axis."
        },
    )
    mask_time_length: int = field(
        default=10,
        metadata={"help": "Length of vector span to mask along the time axis."},
    )
    mask_feature_prob: float = field(
        default=0.0,
        metadata={
            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
                    "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
        },
    )
    mask_feature_length: int = field(
        default=10,
        metadata={"help": "Length of vector span to mask along the feature axis."},
    )
    layerdrop: float = field(default=0.0, metadata={
                             "help": "The LayerDrop probability."})
    ctc_loss_reduction: Optional[str] = field(
        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
    )
    ctc_zero_infinity: Optional[bool] = field(
        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
    )


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.

    Using `HfArgumentParser` we can turn this class
    into argparse arguments to be able to specify them on
    the command line.
    """

    dataset_name: str = field(
        metadata={
            "help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: str = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    train_split_name: str = field(
        default="train",
        metadata={
            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
        },
    )
    eval_split_name: str = field(
        default="test",
        metadata={
            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
        },
    )
    audio_column_name: str = field(
        default="audio",
        metadata={
            "help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
    )
    text_column_name: str = field(
        default="sentence",
        metadata={
            "help": "The name of the dataset column containing the text data. Defaults to 'sentence'"},
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
                    "value if set."
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
                    "value if set."
        },
    )
    chars_to_ignore: Optional[List[str]] = list_field(
        default=None,
        metadata={"help": "A list of characters to remove from the transcripts."},
    )
    eval_metrics: List[str] = list_field(
        default=["wer"],
        metadata={
            "help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
    )
    max_duration_in_seconds: float = field(
        default=20.0,
        metadata={
            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
        },
    )
    min_duration_in_seconds: float = field(
        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
    )
    preprocessing_only: bool = field(
        default=False,
        metadata={
            "help": "Whether to only do data preprocessing and skip training. "
                    "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
                    "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
                    "so that the cached datasets can consequently be loaded in distributed training"
        },
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": "If :obj:`True`, will use the token generated when running"
                    ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
        },
    )
    unk_token: str = field(
        default="[UNK]",
        metadata={"help": "The unk token for the tokenizer"},
    )
    pad_token: str = field(
        default="[PAD]",
        metadata={"help": "The padding token for the tokenizer"},
    )
    word_delimiter_token: str = field(
        default="|",
        metadata={"help": "The word delimiter token for the tokenizer"},
    )

    phoneme_language: Optional[str] = field(
        default=None,
        metadata={
            "help": "The target language that should be used be"
            " passed to the tokenizer for tokenization. Note that"
                    " this is only relevant if the model classifies the"
                    " input audio to a sequence of phoneme sequences."
        },
    )
    print_training_arguments: bool = field(
        default=True,
        metadata={
            "help": "Prints the training arguments. For debugging"
        },
    )


@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
    processor: Any

    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
        # split inputs and labels since they have to be of different lengths and need different padding methods
        # first treat the audio inputs by simply returning torch tensors
        input_features = [{"input_features": feature["input_features"]}
                          for feature in features]
        batch = self.processor.feature_extractor.pad(
            input_features, return_tensors="pt")

        # get the tokenized label sequences
        label_features = [{"input_ids": feature["labels"]}
                          for feature in features]
        # pad the labels to max length
        labels_batch = self.processor.tokenizer.pad(
            label_features, return_tensors="pt")

        # replace padding with -100 to ignore loss correctly
        labels = labels_batch["input_ids"].masked_fill(
            labels_batch.attention_mask.ne(1), -100)

        # if bos token is appended in previous tokenization step,
        # cut bos token here as it's append later anyways
        if (labels[:, 0] == self.processor.tokenizer.bos_token_id).all().cpu().item():
            labels = labels[:, 1:]

        batch["labels"] = labels
        return batch


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.
    parser = HfArgumentParser(
        (ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    # Metrics
    def compute_metrics(pred):
        pred_ids = pred.predictions
        label_ids = pred.label_ids

        # replace -100 with the pad_token_id
        label_ids[label_ids == -100] = tokenizer.pad_token_id

        # we do not want to group tokens when computing the metrics
        pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
        label_str = tokenizer.batch_decode(label_ids, skip_special_tokens=True)

        wer = 100 * metric.compute(predictions=pred_str, references=label_str)

        return {"wer": wer}

    # Prepare dataset
    def prepare_dataset(batch):
        # load and resample audio data from 48 to 16kHz
        audio = batch["audio"]

        # compute log-Mel input features from input audio array
        batch["input_features"] = feature_extractor(
            audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]

        # encode target text to label ids
        batch["labels"] = tokenizer(batch["sentence"]).input_ids
        return batch

    def print_training_arguments(model_args, data_args, training_args):
        print("Starting with the following parameters:")
        print("\n* Model arguments:")
        pprint(vars(model_args), indent=2)
        print("\n* Data arguments")
        pprint(vars(data_args), indent=2)
        print("\n* Training arguments")
        pprint(vars(training_args), indent=2)


    # Print training arguments
    if data_args.print_training_arguments:
        print_training_arguments(model_args, data_args, training_args)

    
  
    # Initialise the model
    feature_extractor = WhisperFeatureExtractor.from_pretrained(
        model_args.model_name_or_path)
    tokenizer = WhisperTokenizer.from_pretrained(
        model_args.model_name_or_path, language=model_args.language, task=model_args.task)
    processor = WhisperProcessor.from_pretrained(
        model_args.model_name_or_path, language=model_args.language, task=model_args.task)
    data_collator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)

    # Saving the processor and the tokenizer
    processor.save_pretrained(training_args.output_dir)
    tokenizer.save_pretrained(training_args.output_dir)


    # Load dataset
    train_dataset = load_dataset(data_args.dataset_name, data_args.dataset_config_name,
                                 split="train", streaming=True, use_auth_token=True)
    eval_dataset = load_dataset(data_args.dataset_name, data_args.dataset_config_name,
                                split="test", streaming=True, use_auth_token=True)
    
    # Because a bug in Datasets (https://github.com/huggingface/datasets/issues/3888) we need to read the columns and keep them for later
    column_names=[x for x in train_dataset.info.features]
    
    # Make sure everything is in 16K
    train_dataset = train_dataset.cast_column(data_args.audio_column_name, Audio(sampling_rate=16000))
    eval_dataset = eval_dataset.cast_column(data_args.audio_column_name, Audio(sampling_rate=16000))
   
    # Rename columns
    if data_args.audio_column_name != "audio":
        train_dataset = train_dataset.rename_column(
            data_args.audio_column_name, "audio")
        eval_dataset = eval_dataset.rename_column(
            data_args.audio_column_name, "audio")
        column_names.remove(data_args.audio_column_name)

    if data_args.text_column_name != "sentence":
        train_dataset = train_dataset.rename_column(
            data_args.text_column_name, "sentence")
        eval_dataset = eval_dataset.rename_column(
            data_args.text_column_name, "sentence")
        column_names.remove(data_args.text_column_name)
        
    # Prepare the dataset
    train_dataset = train_dataset.map(prepare_dataset, remove_columns=column_names)
    eval_dataset = eval_dataset.map(prepare_dataset, remove_columns=column_names)
    
    # Define metrics
    metric = evaluate.load("wer")

    # 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."

            )

   # Training
    if training_args.do_train:
        # TODO I have not yet verified that this part works as expected. The checkpoint=None should also give a meaningful error.
        # The script should not allow you to train a whisper from scratch...

        # use last checkpoint if exist
        if last_checkpoint is not None:
            print("*** Found a checkpoint!")
            checkpoint = last_checkpoint
        elif os.path.isdir(model_args.model_name_or_path):
            print("*** Loading checkpoint from parameters")
            checkpoint = model_args.model_name_or_path
        else:
            checkpoint = None

         # We need to set use_cache=False here if we want to use gradient accumulation
        model = WhisperForConditionalGeneration.from_pretrained(
            "openai/whisper-small", use_cache=False)

        # Overriding generation arguments - no tokens are forced as decoder outputs (see [`forced_decoder_ids`](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.generation_utils.GenerationMixin.generate.forced_decoder_ids)), no tokens are suppressed during generation (see [`suppress_tokens`](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.generation_utils.GenerationMixin.generate.suppress_tokens)):
        model.config.forced_decoder_ids = None
        model.config.suppress_tokens = []

        # Set seed before initializing model.
        set_seed(training_args.seed)

        # TODO - I think the number of epochs needs to be set manually? Now it seems to be calculated based on the save steps. How do I do this?
        # This is currently the output from Trainer - The "Num Epochs" indicates the universe might end before training is finished
        # ***** Running training *****
        # Num examples = 480000
        # Num Epochs = 9223372036854775807
        # Instantaneous batch size per device = 48

        trainer = Seq2SeqTrainer(
            args=training_args,
            model=model,
            train_dataset=train_dataset.with_format("torch"),
            eval_dataset=eval_dataset.with_format(
                "torch").take(data_args.max_eval_samples),
            data_collator=data_collator,
            compute_metrics=compute_metrics,
            tokenizer=processor.feature_extractor,
        )

        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()

        metrics = train_result.metrics
        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)

        trainer.save_state()

        # 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": "automatic-speech-recognition",
            "tags": ["hf-asr-leaderboard", "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 training_args.push_to_hub:
            trainer.push_to_hub(**kwargs)
        else:
            trainer.create_model_card(**kwargs)

    return train_result

# XLA hook
def _mp_fn(index):
    # For xla_spawn (TPUs)
    print("The XLA is initiated")
    main()

if __name__ == "__main__":
    main()