main.py

from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2Model, Wav2Vec2Processor, Wav2Vec2CTCTokenizer
import torchaudio
import torch
from huggingface_hub import notebook_login
from datasets import load_dataset
import re
from transformers import Wav2Vec2ForCTC
from torch.utils.data import DataLoader

organization_name = "ASR-Erzya-Final-Project"
#organization_name = "zmmccormick3"
dataset_name = "asr_erzya_final_data"

dataset = load_dataset(f"{organization_name}/{dataset_name}")

processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")

#://huggingface.co/datasets/ASR-Erzya-Final-Project/asr_erzya_final_data/tree/main

def preprocess_data(batch):
    inputs = processor(batch["audio"], return_tensors="pt", sampling_rate=16000)
    targets = processor(batch["text"], return_tensors="pt", padding=True)

    # Ensure that target tensor is of shape (batch_size, sequence_length)
    targets["input_ids"] = targets["input_ids"].unsqueeze(0)

    return inputs, targets

# Create DataLoader for training data
train_data_loader = DataLoader(dataset["train"], batch_size=4, collate_fn=preprocess_data)

# Training loop
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)

for epoch in range(5):  # Set the number of epochs you want to train for
    model.train()
    for batch in train_data_loader:
        inputs, targets = batch
        inputs = {key: value.to(device) for key, value in inputs.items()}
        targets = {key: value.to(device) for key, value in targets.items()}

        optimizer.zero_grad()
        outputs = model(**inputs, labels=targets["input_ids"])
        loss = outputs.loss
        loss.backward()
        optimizer.step()

# Save the trained model
model.save_pretrained("your-organization/your-wav2vec-ctc-model")
processor.save_pretrained("your-organization/your-wav2vec-ctc-model")


#common_voice_train = load_dataset(f"{organization_name}/{dataset_name}", split="train")
#common_voice_test = load_dataset(f"{organization_name}/{dataset_name}", split="test")

# notebook_login()

common_voice_train = load_dataset("common_voice", "myv", split="train")
common_voice_test = load_dataset("common_voice", "myv", split="test")

# Remove unnecessary columns
common_voice_train = common_voice_train.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "segment", "up_votes"])
common_voice_test = common_voice_test.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "segment", "up_votes"])

chars_to_remove_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\']'

def remove_special_characters(batch):
    batch["sentence"] = re.sub(chars_to_remove_regex, '', batch["sentence"]).lower()
    return batch

common_voice_train = common_voice_train.map(remove_special_characters)
common_voice_test = common_voice_test.map(remove_special_characters)

def replace_hatted_characters(batch):
    batch["sentence"] = re.sub('[â]', 'a', batch["sentence"])
    # Add similar lines for other hat characters if needed
    return batch

common_voice_train = common_voice_train.map(replace_hatted_characters)
common_voice_test = common_voice_test.map(replace_hatted_characters)

def extract_all_chars(batch):
    all_text = " ".join(batch["sentence"])
    vocab = list(set(all_text))
    return {"vocab": [vocab], "all_text": [all_text]}

vocab_train = common_voice_train.map(extract_all_chars, batched=True, batch_size=-1, keep_in_memory=True, remove_columns=common_voice_train.column_names)
vocab_test = common_voice_test.map(extract_all_chars, batched=True, batch_size=-1, keep_in_memory=True, remove_columns=common_voice_test.column_names)

vocab_list = list(set(vocab_train["vocab"][0]) | set(vocab_test["vocab"][0]))
vocab_dict = {v: k for k, v in enumerate(sorted(vocab_list))}
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)

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

repo_name = "wav2vec2-large-xls-r-300m-tr-colab" # TK repository name
tokenizer.push_to_hub(repo_name)

# tokenizer = Wav2Vec2CTCTokenizer.from_pretrained() # TK
# feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/wav2vec2-base-960h")


def downsample(batch):
    resample = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
    batch["audio"] = resample(batch["audio"])
    return batch


common_voice_train = common_voice_train.map(downsample)
common_voice_test = common_voice_test.map(downsample)


feature_extractor = Wav2Vec2FeatureExtractor(feature_size=1, sampling_rate=16000, padding_value=0.0, do_normalize=True, return_attention_mask=True)
# processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)

'''
model = Wav2Vec2Model.from_pretrained("Link/to/HuggingfaceModel")
array, fs = torchaudio.load("/Local/link/to/your/audio.wav"
input_input = processor(array.squeeze(), sampling_rate=fs, return_tensors="pt")
with torch.no_grad():
    outputs = model(**input_input)
print(f"Hidden state shape: {outputs.last_hidden_state.numpy().shape}")
'''