xlsr-train.py

import os

from datasets import load_dataset, Audio, concatenate_datasets

# Set processors (optional)

num_proc = os.cpu_count()//2
num_dataloaders = os.cpu_count()//2

print(f"Cpu count: {os.cpu_count()}\nNum proc: {num_proc}\nNum dataloaders: {num_dataloaders}")

# Load datasets

train = load_dataset()
dev = load_dataset()
test = load_dataset()

import unicodedata
import re

def preprocess_text(batch):
  # Convert to lowercase
  batch['sentence'] = batch['sentence'].lower()

  # Normalize text
  batch['sentence'] = unicodedata.normalize('NFKC', batch['sentence'])
  batch['sentence'] = re.sub(r'[\’\ʻ\ʼ\ʽ\‘]', "'", batch['sentence'])

  # Remove punctuation and special characters
  batch['sentence'] = re.sub(r'[^\w\s\']', '', batch['sentence'])
  batch['sentence'] = re.sub(r'_', ' ', batch['sentence'])

  # Remove excessive whitespace
  batch['sentence'] = ' '.join(batch['sentence'].split())

  return batch

import librosa
import numpy as np

def get_lens(batch):
  try:
    audio_len = librosa.get_duration(y=batch['audio']['array'], sr=batch['audio']['sampling_rate'])
  except:
    del batch['audio']
    batch['audio'] = None
    audio_len = 0.0
    transcript_len = len(batch['sentence'])
    batch['audio_len'] = audio_len
    batch['transcript_len'] = transcript_len
    batch['len_ratio'] = float(audio_len)/float(transcript_len)
    batch['num_feature_vecs'] = int(np.round(audio_len * 1000 / 20))
    return batch
  
  transcript_len = len(batch['sentence'])

  batch['audio_len'] = audio_len
  batch['transcript_len'] = transcript_len
  batch['len_ratio'] = float(audio_len)/float(transcript_len)
  batch['num_feature_vecs'] = int(np.round(audio_len * 1000 / 20)) # seconds -> milliseconds, divide by 20 millisecond feature_win_step, round up to nearest int

  return batch

def data_checks(batch):
  audio_check = (batch['audio_len']>1.0 and batch['audio_len']<30.0)
  transcript_check = (batch['transcript_len']>10)

  input_output_ratio = float(batch['num_feature_vecs']) / float(batch['transcript_len'])
  input_output_ratio_check = input_output_ratio>1.0 # CTC algorithm assumes the input is not shorter than the ouput

  return (audio_check and transcript_check and input_output_ratio_check)

train = train.map(preprocess_text, num_proc=num_proc)
dev = dev.map(preprocess_text, num_proc=num_proc)

try:
  train = train.map(get_lens, num_proc=num_proc)
except:
  train = train.map(get_lens, num_proc=4)

try:
  dev = dev.map(get_lens, num_proc=num_proc)
except:
  dev = dev.map(get_lens, num_proc=4)

train = train.filter(data_checks, num_proc=num_proc)
dev = dev.filter(data_checks, num_proc=num_proc)

train_mean = np.mean(train['len_ratio'])
train_std = np.std(train['len_ratio'])

dev_mean = np.mean(dev['len_ratio'])
dev_std = np.std(dev['len_ratio'])

num_std_devs = 2
train = train.filter(lambda batch: (abs(batch['len_ratio'] - train_mean) - (num_std_devs * train_std)) <= 0, num_proc=num_proc)
dev = dev.filter(lambda batch: (abs(batch['len_ratio'] - dev_mean) - (num_std_devs * dev_std)) <= 0, num_proc=num_proc)

print(f"Train hours: {sum(train['audio_len'])/3600}\nDev hours: {sum(dev['audio_len'])/3600}")

train = train.remove_columns(['audio_len', 'transcript_len', 'len_ratio', 'num_feature_vecs'])
dev = dev.remove_columns(['audio_len', 'transcript_len', 'len_ratio', 'num_feature_vecs'])

alphabet = None # define the language's alphabet here e.g. " 'abcdefghijklmnorstuwyzƙƴɓɗ" for Hausa

alphabet = sorted(list(set(alphabet)))

vocab_dict = {v: k for k, v in enumerate(alphabet)}

vocab_dict["|"] = vocab_dict[" "]
del vocab_dict[" "]

vocab_dict["[UNK]"] = len(vocab_dict)
vocab_dict["[PAD]"] = len(vocab_dict)

import json

with open('vocab.json', 'w') as vocab_file:
  json.dump(vocab_dict, vocab_file)

from transformers import Wav2Vec2CTCTokenizer

tokenizer = Wav2Vec2CTCTokenizer.from_pretrained("./", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")

from transformers import Wav2Vec2FeatureExtractor

feature_extractor = Wav2Vec2FeatureExtractor(feature_size=1, sampling_rate=16000, padding_value=0.0, do_normalize=True, return_attention_mask=True)

from transformers import Wav2Vec2Processor

processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)

def prepare_dataset(batch):
  audio = batch["audio"]
  batch["input_values"] = processor(audio=audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
  batch["input_length"] = len(batch["input_values"])
  batch["labels"] = processor(text=batch["sentence"]).input_ids
  return batch

try:
  train = train.map(prepare_dataset, remove_columns=train.column_names, num_proc=num_proc)
except:
  train = train.map(prepare_dataset, remove_columns=train.column_names, num_proc=4)

try:
  dev = dev.map(prepare_dataset, remove_columns=dev.column_names, num_proc=num_proc)
except:
  dev = dev.map(prepare_dataset, remove_columns=dev.column_names, num_proc=4)

import torch

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

@dataclass
class DataCollatorCTCWithPadding:
  """
  Data collator that will dynamically pad the inputs received.
  Args:
      processor (:class:`~transformers.Wav2Vec2Processor`)
          The processor used for proccessing the data.
      padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
          Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
          among:
          * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
            sequence if provided).
          * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
            maximum acceptable input length for the model if that argument is not provided.
          * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
            different lengths).
  """

  processor: Wav2Vec2Processor
  padding: Union[bool, str] = True

  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
      input_features = [{"input_values": feature["input_values"]} for feature in features]
      label_features = [{"input_ids": feature["labels"]} for feature in features]

      batch = self.processor.pad(
          input_features=input_features,
          padding=self.padding,
          return_tensors="pt",
      )

      labels_batch = self.processor.pad(
          labels=label_features,
          padding=self.padding,
          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)

      batch["labels"] = labels

      return batch

data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True)

"""# Model Training"""

import evaluate

wer_metric = evaluate.load("wer")
cer_metric = evaluate.load("cer")

import numpy as np

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)
  label_str = processor.batch_decode(pred.label_ids, group_tokens=False)

  wer = wer_metric.compute(predictions=pred_str, references=label_str)
  cer = cer_metric.compute(predictions=pred_str, references=label_str)

  return {"wer": wer, "cer": cer}

from transformers import Wav2Vec2ForCTC, TrainingArguments, Trainer, EarlyStoppingCallback

model_checkpoint = "facebook/wav2vec2-xls-r-300m"

model = Wav2Vec2ForCTC.from_pretrained(
    model_checkpoint,
    attention_dropout=0.0,
    hidden_dropout=0.0,
    feat_proj_dropout=0.0,
    mask_time_prob=0.05,
    layerdrop=0.0,
    ctc_loss_reduction="mean",
    pad_token_id=processor.tokenizer.pad_token_id,
    vocab_size=len(processor.tokenizer),
)

model.freeze_feature_encoder()

import wandb

dataset = None
language = None
sample_hours = None
version = None
batch_size = None
grad_acc = 1
eval_batch_size = batch_size//2
epochs = None
output_dir = f"{model_checkpoint.split('/')[-1]}-{dataset}-{language}-{sample_hours}hrs-{version}"

wandb.init(
    project="ASR Africa",
    entity="asr-africa-research-team",
    name=output_dir,
)

training_args = TrainingArguments(
  output_dir=output_dir,
  group_by_length=True,
  per_device_train_batch_size=batch_size,
  per_device_eval_batch_size=eval_batch_size,
  gradient_accumulation_steps=grad_acc,
  eval_strategy="epoch",
  logging_strategy="epoch",
  save_strategy="epoch",
  num_train_epochs=epochs,
  gradient_checkpointing=True,
  fp16=True,
  learning_rate=None,
  lr_scheduler_type='linear',
  warmup_ratio=None,
  save_total_limit=2,
  load_best_model_at_end=True,
  metric_for_best_model="wer",
  greater_is_better=False,
  optim='adamw_torch',
  push_to_hub=True,
  hub_model_id=f"asr-africa/{output_dir}",
  hub_private_repo=True,
  dataloader_num_workers=num_dataloaders,
)

trainer = Trainer(
  model=model,
  data_collator=data_collator,
  args=training_args,
  compute_metrics=compute_metrics,
  train_dataset=train,
  eval_dataset=dev,
  tokenizer=processor.feature_extractor,
  callbacks=[
    EarlyStoppingCallback(
      early_stopping_patience=10,
      early_stopping_threshold=1e-3
    )
  ],
)

trainer.train()

kwargs = {
    "dataset_tags": "",
    "dataset": "",
    "language": "",
    "model_name": "",
    "finetuned_from": model_checkpoint,
    "tasks": "automatic-speech-recognition",
}

trainer.push_to_hub(**kwargs)

other_test_dataset_1 = load_dataset()
other_test_dataset_2 = load_dataset()

test = concatenate_datasets([test, other_test_dataset_1, other_test_dataset_2]).shuffle(42)

test = test.map(preprocess_text, num_proc=num_proc)

try:
  test = test.map(get_lens, num_proc=num_proc)
except:
  test = test.map(get_lens, num_proc=4)

test = test.filter(data_checks, num_proc=num_proc)

test_mean = np.mean(test['len_ratio'])
test_std = np.std(test['len_ratio'])
num_std_devs = 2
test = test.filter(lambda batch: (abs(batch['len_ratio'] - test_mean) - (num_std_devs * test_std)) <= 0, num_proc=num_proc)

print(f"Test hours: {sum(test['audio_len'])/3600}")

test = test.remove_columns(['audio_len', 'transcript_len', 'len_ratio', 'num_feature_vecs'])

try:
  test = test.map(prepare_dataset, remove_columns=test.column_names, num_proc=num_proc)
except:
  test = test.map(prepare_dataset, remove_columns=test.column_names, num_proc=4)

results = trainer.evaluate(eval_dataset=test, metric_key_prefix="test")
print(results)

wandb.log(results)

train.cleanup_cache_files()
dev.cleanup_cache_files()
test.cleanup_cache_files()

torch.cuda.empty_cache()