app.py

import os
import torch
from PIL import Image
import gradio as gr
from datasets import load_dataset
from transformers import AutoProcessor, MllamaForConditionalGeneration, Trainer, TrainingArguments

# === CONFIG ===
CKPT = "unsloth/Llama-3.2-11B-Vision-Instruct"
MODEL_SAVE_PATH = "./llama3-handwriting-ocr-simple"
DATASET_NAME = "om440/partial-rimes-handwritten-small"
MAX_TOKENS = 128
TRAIN_EXAMPLES = 20

# === PROMPT ===
TRAINING_PROMPT = """Extract all text from this image exactly as written.
Include both handwritten and printed text.
Preserve original formatting and line breaks."""

# === LOAD MODEL ET PROCESSOR ===
def load_model():
    if os.path.exists(MODEL_SAVE_PATH):
        model = MllamaForConditionalGeneration.from_pretrained(MODEL_SAVE_PATH)
        processor = AutoProcessor.from_pretrained(MODEL_SAVE_PATH)
    else:
        model = MllamaForConditionalGeneration.from_pretrained(CKPT)
        processor = AutoProcessor.from_pretrained(CKPT)
    return model, processor

# === PREPROCESSING SIMPLE ===
def preprocess(examples, processor):
    images = []
    for img in examples["image"]:
        if hasattr(img, "convert"):
            image = img.convert("RGB")
        else:
            image = Image.fromarray(img).convert("RGB")
        image.thumbnail((336, 336), Image.Resampling.LANCZOS)
        images.append(image)

    inputs = processor(
        text=[TRAINING_PROMPT] * len(images),
        images=images,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=64,
    )
    labels = processor.tokenizer(
        examples["text"],
        padding="max_length",
        truncation=True,
        max_length=MAX_TOKENS,
        return_tensors="pt"
    )
    return {
        "input_ids": inputs["input_ids"],
        "attention_mask": inputs["attention_mask"],
        "pixel_values": inputs["pixel_values"],
        "labels": labels["input_ids"]
    }

# === TRAINING ===
def train(num_examples=TRAIN_EXAMPLES):
    dataset = load_dataset(DATASET_NAME)["train"].select(range(num_examples))
    model, processor = load_model()

    tokenized = dataset.map(
        lambda ex: preprocess(ex, processor),
        batched=True,
        batch_size=1,
        remove_columns=dataset.column_names
    )

    training_args = TrainingArguments(
        output_dir=MODEL_SAVE_PATH,
        per_device_train_batch_size=1,
        num_train_epochs=1,
        logging_steps=1,
        save_steps=10,
        save_total_limit=1,
        remove_unused_columns=False,
        report_to="none",
        gradient_accumulation_steps=1,
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized,
        tokenizer=processor.tokenizer,
        data_collator=lambda data: {
            "input_ids": torch.stack([f["input_ids"] for f in data]),
            "attention_mask": torch.stack([f["attention_mask"] for f in data]),
            "pixel_values": torch.stack([f["pixel_values"] for f in data]),
            "labels": torch.stack([f["labels"] for f in data]),
        }
    )

    trainer.train()
    model.save_pretrained(MODEL_SAVE_PATH)
    processor.save_pretrained(MODEL_SAVE_PATH)
    return f"✅ Entraînement terminé sur {num_examples} exemples."

# === INFERENCE ===
def extract_text(image_path):
    model, processor = load_model()
    image = Image.open(image_path).convert("RGB")
    image.thumbnail((336, 336), Image.Resampling.LANCZOS)

    prompt = TRAINING_PROMPT

    inputs = processor(
        text=prompt,
        images=[image],
        return_tensors="pt"
    )

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=150,
            pad_token_id=processor.tokenizer.pad_token_id,
            eos_token_id=processor.tokenizer.eos_token_id,
            do_sample=False,
        )

    result = processor.decode(outputs[0], skip_special_tokens=True)
    if prompt in result:
        result = result.replace(prompt, "").strip()
    return result or "Aucun texte détecté."

# === GRADIO UI ===
def launch_gradio():
    with gr.Blocks() as demo:
        gr.Markdown("# LLaMA 3.2 Vision OCR simple")

        with gr.Tab("Entraînement"):
            train_btn = gr.Button("🚀 Entraîner sur 20 exemples")
            train_output = gr.Textbox(lines=4, interactive=False)
            train_btn.click(fn=train, inputs=[], outputs=train_output)

        with gr.Tab("Extraction"):
            image_input = gr.Image(type="filepath", label="Uploader une image")
            extract_btn = gr.Button("🔍 Extraire le texte")
            text_output = gr.Textbox(lines=10, interactive=False)
            extract_btn.click(fn=extract_text, inputs=image_input, outputs=text_output)

    return demo

if __name__ == "__main__":
    demo = launch_gradio()
    demo.launch(share=True)