import os
import zipfile
import numpy as np
import torch
from transformers import ViTForImageClassification, AdamW
import nibabel as nib
from PIL import Image
from torch.utils.data import Dataset, DataLoader
import streamlit as st

# 1. Function to extract zip files
def extract_zip(zip_file, extract_to):
    with zipfile.ZipFile(zip_file, 'r') as zip_ref:
        zip_ref.extractall(extract_to)

# 2. Preprocess images
def preprocess_image(image_path):
    ext = os.path.splitext(image_path)[-1].lower()

    if ext == '.nii' or ext == '.nii.gz':
        nii_image = nib.load(image_path)
        image_data = nii_image.get_fdata()
        image_tensor = torch.tensor(image_data).float()
        if len(image_tensor.shape) == 3:
            image_tensor = image_tensor.unsqueeze(0)

    elif ext in ['.jpg', '.jpeg']:
        img = Image.open(image_path).convert('RGB').resize((224, 224))
        img_np = np.array(img)
        image_tensor = torch.tensor(img_np).permute(2, 0, 1).float()

    else:
        raise ValueError(f"Unsupported format: {ext}")

    image_tensor /= 255.0  # Normalize to [0, 1]
    return image_tensor

# 3. Label images
def prepare_dataset(extracted_folder):
    image_paths = []
    labels = []
    for disease_folder in ['alzheimers', 'parkinsons', 'ms']:
        folder_path = os.path.join(extracted_folder, disease_folder)
        label = {'alzheimers': 0, 'parkinsons': 1, 'ms': 2}[disease_folder]
        for img_file in os.listdir(folder_path):
            if img_file.endswith(('.nii', '.jpg', '.jpeg')):
                image_paths.append(os.path.join(folder_path, img_file))
                labels.append(label)
    return image_paths, labels

# 4. Custom Dataset
class CustomImageDataset(Dataset):
    def __init__(self, image_paths, labels):
        self.image_paths = image_paths
        self.labels = labels

    def __len__(self):
        return len(self.image_paths)

    def __getitem__(self, idx):
        image = preprocess_image(self.image_paths[idx])
        label = self.labels[idx]
        return image, label

# 5. Training function
def fine_tune_model(train_loader):
    model = ViTForImageClassification.from_pretrained('google/vit-base-patch16-224-in21k', num_labels=3)
    model.train()
    optimizer = AdamW(model.parameters(), lr=1e-4)
    criterion = torch.nn.CrossEntropyLoss()
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model.to(device)
    for epoch in range(10):
        running_loss = 0.0
        for images, labels in train_loader:
            images, labels = images.to(device), labels.to(device)
            optimizer.zero_grad()
            outputs = model(pixel_values=images).logits
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
            running_loss += loss.item()
    return running_loss / len(train_loader)

# Streamlit UI
st.title("Fine-tune ViT on MRI Scans")

if st.button("Start Training"):
    extract_zip('your_zip_file.zip', 'extracted_folder/')
    image_paths, labels = prepare_dataset('extracted_folder/')
    dataset = CustomImageDataset(image_paths, labels)
    train_loader = DataLoader(dataset, batch_size=32, shuffle=True)
    final_loss = fine_tune_model(train_loader)
    st.write(f"Training Complete with Final Loss: {final_loss}")