convert_checkpoint.py

# KAGGLE SPECIFIC: This script is used to make training compatible with Kaggle's notebook environment.

import torch
import os

def load_and_save_checkpoint(input_filename, output_filename, device):
    if os.path.isfile(input_filename):
        print(f"Loading checkpoint '{input_filename}'")
        checkpoint = torch.load(input_filename, map_location=device)
        
        # Extract only the necessary state
        save_state = {
            'epoch': checkpoint['epoch'],
            'generator_state_dict': checkpoint['generator_state_dict'],
            'discriminator_state_dict': checkpoint['discriminator_state_dict'],
            'optimizerG_state_dict': checkpoint['optimizerG_state_dict'],
            'optimizerD_state_dict': checkpoint['optimizerD_state_dict'],
        }
        
        # Save the checkpoint
        torch.save(save_state, output_filename)
        print(f"Saved checkpoint to '{output_filename}'")
    else:
        print(f"No checkpoint found at '{input_filename}'")

if __name__ == "__main__":
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Using device: {device}")

    input_checkpoint = "checkpoints/latest_checkpoint.pth.tar"
    output_checkpoint = "checkpoints/converted_checkpoint.pth.tar"

    load_and_save_checkpoint(input_checkpoint, output_checkpoint, device)