app.py

import torch
from torch import nn, optim
from torch.utils.data import DataLoader, Dataset
from torchvision import transforms, datasets, models
from PIL import Image
import json
import os
import gradio as gr
import shutil

# Paths
image_folder = "Images/"
metadata_file = "descriptions.json"

# Define the function to load metadata
def load_metadata(metadata_file):
    with open(metadata_file, 'r') as f:
        metadata = json.load(f)
    return metadata

# Custom Dataset Class
class ImageDescriptionDataset(Dataset):
    def __init__(self, image_folder, metadata):
        self.image_folder = image_folder
        self.metadata = metadata
        self.image_names = list(metadata.keys())  # List of image filenames
        self.transform = transforms.Compose([
            transforms.Resize((512, 512)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
        ])

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

    def __getitem__(self, idx):
        image_name = self.image_names[idx]
        image_path = os.path.join(self.image_folder, image_name)
        image = Image.open(image_path).convert("RGB")
        description = self.metadata[image_name]
        image = self.transform(image)
        return image, description

# LoRA Layer Implementation
class LoRALayer(nn.Module):
    def __init__(self, original_layer, rank=4):
        super(LoRALayer, self).__init__()
        self.original_layer = original_layer
        self.rank = rank
        self.lora_up = nn.Linear(original_layer.in_features, rank, bias=False)
        self.lora_down = nn.Linear(rank, original_layer.out_features, bias=False)

    def forward(self, x):
        return self.original_layer(x) + self.lora_down(self.lora_up(x))

# LoRA Model Class
class LoRAModel(nn.Module):
    def __init__(self):
        super(LoRAModel, self).__init__()
        self.backbone = models.resnet18(pretrained=True)  # Base model
        self.backbone.fc = LoRALayer(self.backbone.fc)  # Replace the final layer with LoRA

    def forward(self, x):
        return self.backbone(x)

# Training Function
def train_lora(image_folder, metadata):
    print("Starting LoRA training process...")

    # Create dataset and dataloader
    dataset = ImageDescriptionDataset(image_folder, metadata)
    dataloader = DataLoader(dataset, batch_size=8, shuffle=True)

    # Initialize model, loss function, and optimizer
    model = LoRAModel()
    criterion = nn.CrossEntropyLoss()  # Update this if your task changes
    optimizer = optim.Adam(model.parameters(), lr=0.001)

    # Training loop
    num_epochs = 5  # Adjust the number of epochs based on your needs
    for epoch in range(num_epochs):
        print(f"Epoch {epoch + 1}/{num_epochs}")
        for batch_idx, (images, descriptions) in enumerate(dataloader):
            # Convert descriptions to a numerical format (if applicable)
            labels = torch.randint(0, 100, (images.size(0),))  # Placeholder labels

            # Forward pass
            outputs = model(images)
            loss = criterion(outputs, labels)
            
            # Backward pass
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            
            if batch_idx % 10 == 0:  # Log every 10 batches
                print(f"Batch {batch_idx}, Loss: {loss.item()}")

    # Save the trained model to /mnt/data/ for Hugging Face Space to access
    save_path = '/mnt/data/lora_model.pth'
    torch.save(model.state_dict(), save_path)
    print(f"Model saved at {save_path}")

    # Move the file to a location where we can access it for download
    # Here, /mnt/data is directly accessible from the Hugging Face Space interface
    print(f"Training completed. The model is saved and ready for download at {save_path}.")

    return f"Training completed. Download the model from: [Download Model](sandbox:/mnt/data/lora_model.pth)"

# Gradio App
def start_training_gradio():
    print("Loading metadata and preparing dataset...")
    metadata = load_metadata(metadata_file)
    return train_lora(image_folder, metadata)

demo = gr.Interface(
    fn=start_training_gradio,
    inputs=None,
    outputs="text",
    title="Train LoRA Model",
    description="Fine-tune a model using LoRA for consistent image generation."
)

demo.launch()