TinyDiffusion

Feb 17, 2025#diffusion56

AI Translation

This post is translated from Chinese into English through AI.View Original

AI-generated summary

The document outlines the process of implementing TinyDiffusion on Kaggle using the CIFAR-10 dataset. It includes steps for data preparation, such as cloning the TinyDiffusion repository and uploading the CIFAR-10 dataset. The instructions detail how to set up the environment, copy datasets to the output directory, and create directories for saving images generated every 10 epochs. It also describes running scripts for data processing, noise addition, and training the model for 100 epochs. Finally, it provides code to check for saved images and display them.

Reference：Datawhale tinydiffusion

Kaggle Implementation:#

Data Preparation#

git clone to download tinydiffusion locally
Download the cifar-10-python dataset to the datasets/ folder, maintaining the directory structure as datasets/cifar-10-batches-py.
Package and upload to Kaggle dataset, named cifar-10-python.

Run#

Create a new notebook, add the dataset cifar-10-python as input.
Run the default code
Configure the environment

!pip install -r /kaggle/input/tinydiffusion/TinyDiffusion/ddpm/requirements.txt

Copy the dataset to the output path

import shutil
import os

# Define the input and output directories for the dataset
input_dir = '/kaggle/input/tinydiffusion'
output_dir = '/kaggle/output/tinydiffusion'

# Check if the output directory exists, if not, create it
if not os.path.exists(output_dir):
    os.makedirs(output_dir)

# Traverse all files and folders in the input directory
for root, dirs, files in os.walk(input_dir):
    # Build the corresponding output directory
    relative_path = os.path.relpath(root, input_dir)
    output_sub_dir = os.path.join(output_dir, relative_path)
    
    # If the output subdirectory does not exist, create it
    if not os.path.exists(output_sub_dir):
        os.makedirs(output_sub_dir)
    
    # Copy files to the output directory
    for file in files:
        input_file_path = os.path.join(root, file)
        output_file_path = os.path.join(output_sub_dir, file)
        shutil.copy2(input_file_path, output_file_path)

print("Dataset copy completed!")

Create a path to store images generated every 10 epochs

import os

# Define the target directory for saving images
save_dir = '/kaggle/output/tinydiffusion/TinyDiffusion/samples'

# Check if the directory exists, if not, create it
if not os.path.exists(save_dir):
    os.makedirs(save_dir)
    print(f"Successfully created directory: {save_dir}")
else:
    print(f"Directory {save_dir} already exists.")

Enter the output path

cd /kaggle/output/tinydiffusion/TinyDiffusion

Process the data

!python /kaggle/input/tinydiffusion/TinyDiffusion/ddpm/dataloader.py

Print the structure

!python /kaggle/input/tinydiffusion/TinyDiffusion/ddpm/unet.py

Add noise to the images

!python /kaggle/input/tinydiffusion/TinyDiffusion/ddpm/diffusion.py

Train, here only training for 100 times, if you want to modify, you can change it in the command line

!python /kaggle/output/tinydiffusion/TinyDiffusion/ddpm/train.py --epochs 100 --batch_size 128 --lr 1e-4 --img_size 32

View the images

import os
from PIL import Image
import matplotlib.pyplot as plt

# Define the image directory
save_dir = '/kaggle/output/tinydiffusion/TinyDiffusion/samples'

# Check if the directory exists
if not os.path.exists(save_dir):
    print(f"Directory {save_dir} does not exist.")
else:
    # Get all files in the directory
    all_files = os.listdir(save_dir)
    
    # Filter out image files (assuming image file extensions are .png, .jpg, .jpeg)
    image_files = [f for f in all_files if f.lower().endswith(('.png', '.jpg', '.jpeg'))]
    
    if not image_files:
        print(f"There are no image files in the directory {save_dir}.")
    else:
        print(f"Found the following image files:")
        for idx, image_file in enumerate(image_files):
            print(f"{idx + 1}. {image_file}")
            # Display the image (optional)
            image_path = os.path.join(save_dir, image_file)
            img = Image.open(image_path)
            plt.imshow(img)
            plt.title(image_file)
            plt.axis('off')
            plt.show()

Hanah

TinyDiffusion

Kaggle Implementation:#

Data Preparation#

Run#

Results#