CGAN_Jittor

A Jittor implementation of Conditional GAN (CGAN) for handwritten digit image generation.

Project Overview

This project implements a Conditional GAN using the Jittor deep learning framework to generate handwritten digit images trained on the MNIST dataset.

Key features

• Conditional generation of specific digits through label-guided adversarial training.
• Lightweight implementation optimized for Jittor’s dynamic computation graph.
• Real-time visualization of generated samples during training.

Installation

Prerequisites

• Supported Systems: Linux or Windows (including WSL). macOS users require a virtual machine.
• Dependencies: Python (≥3.7) and a C++ compiler (g++ ≥5.4.0 for linux or clang ≥8.0 for mac).

Installing Jittor

Choose one of these methods (full guide here):

# Pip (recommended)
python -m pip install jittor

# Docker (CPU-only)
docker run -it --network host jittor/jittor

# For CUDA support, see official installation docs.

Usage

Clone this repository:

git clone https://gitlink.org.cn/chen_03/CGAN_Jittor.git

Run the CGAN script with default parameters:

# Default: Generate "0123456789"
python CGAN.py

Customize training (examples):

# Generate "02468", train for 100 epochs with learning rate 0.0001
python CGAN.py --number="02468" --n_epochs=100 --lr=0.0001

# Generate "13579", batch size 128, latent dimension 100
python CGAN.py --number="13579" --batch_size=128 --latent_dim=100

Key Arguments

Parameter	Default	Description
--number	"0123456789"	Labels of digits to generate
--n_epochs	50	Total training epochs
--lr	0.0002	Learning rate for Adam optimizer
--batch_size	64	Training batch size
--latent_dim	100	Dimension of noise vector

Output

• Training phase: Generated samples saved every 1000 steps in ./samples/ for debugging.
• Final result: result.png showing grid of generated digits after training completion.

Contributing

Contributions are welcome! Open an issue or submit a PR for:

• Performance optimizations
• Additional visualization features
• Extended documentation