import numpy as np
import evaluate
from datasets import load_dataset
from transformers import (
    MT5ForConditionalGeneration, 
    MT5Tokenizer, 
    Seq2SeqTrainingArguments, 
    Seq2SeqTrainer, 
    DataCollatorForSeq2Seq
)

# Load metrics
cer_metric = evaluate.load("cer")
wer_metric = evaluate.load("wer")

model_nm = "google/mt5-small"
tokenizer = MT5Tokenizer.from_pretrained(model_nm)
model = MT5ForConditionalGeneration.from_pretrained(model_nm)

def compute_metrics(eval_preds):
    preds, labels = eval_preds
    if isinstance(preds, tuple):
        preds = preds[0]
    
    decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
    # Replace -100 in labels as we can't decode them
    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

    cer = cer_metric.compute(predictions=decoded_preds, references=decoded_labels)
    wer = wer_metric.compute(predictions=decoded_preds, references=decoded_labels)
    
    return {"cer": cer, "wer": wer}

def tokenize_fn(batch):
    inputs = tokenizer(batch['source'], padding="max_length", truncation=True, max_length=64)
    labels = tokenizer(batch['target'], padding="max_length", truncation=True, max_length=64)
    inputs["labels"] = labels["input_ids"]
    return inputs

# Load and process data
dataset = load_dataset('csv', data_files={'train': 'train.csv', 'test': 'val.csv'})
tokenized_dataset = dataset.map(tokenize_fn, batched=True)

args = Seq2SeqTrainingArguments(
    output_dir="./translit-results",
    evaluation_strategy="epoch",
    learning_rate=2e-4,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=16,
    weight_decay=0.01,
    save_total_limit=2,
    num_train_epochs=3,
    predict_with_generate=True,
    fp16=True, # Set to False if not using GPU
    logging_steps=100,
)

trainer = Seq2SeqTrainer(
    model=model,
    args=args,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["test"],
    tokenizer=tokenizer,
    data_collator=DataCollatorForSeq2Seq(tokenizer, model=model),
    compute_metrics=compute_metrics
)

trainer.train()
model.save_pretrained("./final_model")
tokenizer.save_pretrained("./final_model")