whisper_trainer.py

import os
from datasets import load_from_disk, DatasetDict, concatenate_datasets

# Get all batch directories
batch_dirs = [d for d in os.listdir("processed_dataset") if d.startswith("batch_")]
batch_dirs.sort(key=lambda x: int(x.split('_')[1]))  # Sort numerically

# Load each batch and combine them
processed_batches = []
for batch_dir in batch_dirs:
    batch_path = os.path.join("processed_dataset", batch_dir)
    batch_dataset = load_from_disk(batch_path)
    processed_batches.append(batch_dataset)

# Combine all batches into one dataset
full_dataset = concatenate_datasets(processed_batches)

# Split into train and test
# First shuffle the dataset with a fixed seed for reproducibility
shuffled_dataset = full_dataset.shuffle(seed=42)

# Get the last 975 samples for test
test_size = 975
processed_test = shuffled_dataset.select(range(test_size))
processed_train = shuffled_dataset.select(range(test_size, len(shuffled_dataset)))

# Create the dataset_dict with the new split
dataset_dict = DatasetDict({
    "train": processed_train,
    "test": processed_test
})

# Verify the loading and splitting was successful
print("\nDataset split information:")
print(f"Total examples: {len(shuffled_dataset)}")
print(f"Training examples: {len(dataset_dict['train'])}")
print(f"Test examples: {len(dataset_dict['test'])}")

# Optional: Print the first example from each split to verify the structure
print("\nFirst training example structure:")
print(dataset_dict['train'][0].keys())
print("\nFirst test example structure:")
print(dataset_dict['test'][0].keys())

from transformers import WhisperForConditionalGeneration

model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

model.generation_config.language = "malayalam"

model.generation_config.task = "transcribe"

model.generation_config.forced_decoder_ids = None

from transformers import WhisperProcessor

processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="Malayalam", task="transcribe")

import torch

from dataclasses import dataclass
from typing import Any, Dict, List, Union

@dataclass
@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
    processor: Any
    decoder_start_token_id: int
    max_target_length: int = 448  # Add a parameter to control max length

    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")

        # Truncate labels to max_target_length
        labels = labels_batch["input_ids"]
        if labels.shape[1] > self.max_target_length:
            labels = labels[:, :self.max_target_length]

        # replace padding with -100 to ignore loss correctly
        labels = labels.masked_fill(labels_batch.attention_mask[:, :labels.shape[1]].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.decoder_start_token_id).all().cpu().item():
            labels = labels[:, 1:]

        batch["labels"] = labels

        return batch

# When initializing the data collator, you can now specify the max length
data_collator = DataCollatorSpeechSeq2SeqWithPadding(
    processor=processor,
    decoder_start_token_id=model.config.decoder_start_token_id,
    max_target_length=448  # Explicitly set the max length
)

import evaluate

metric = evaluate.load("wer")

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}

from transformers import Seq2SeqTrainingArguments

training_args = Seq2SeqTrainingArguments(
    output_dir="./whisper-small-mal",  # change to a repo name of your choice
    per_device_train_batch_size=16,
    gradient_accumulation_steps=1,  # increase by 2x for every 2x decrease in batch size
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=4000,
    gradient_checkpointing=True,
    fp16=True,
    fp16_full_eval=True,
    half_precision_backend='auto',
    evaluation_strategy="steps",
    per_device_eval_batch_size=8,
    predict_with_generate=True,
    generation_max_length=225,
    save_steps=1000,
    eval_steps=20,
    logging_steps=25,
    report_to=["tensorboard"],
    load_best_model_at_end=True,
    metric_for_best_model="wer",
    greater_is_better=False,
    push_to_hub=True,
)

from transformers import Seq2SeqTrainer

trainer = Seq2SeqTrainer(
    args=training_args,
    model=model,
    train_dataset=dataset_dict["train"],
    eval_dataset=dataset_dict["test"],
    data_collator=data_collator,
    compute_metrics=compute_metrics,
    tokenizer=processor.feature_extractor,
)

processor.save_pretrained(training_args.output_dir)

trainer.train()