rmachado23/mnist-conditional-gan

MNIST Conditional GAN (cGAN) Generator

Conditional GAN trained on MNIST to generate handwritten digits 0-9.

Architecture

• Type: Conditional DCGAN Generator
• Input: noise vector (100,) + integer label (0-9)
• Output: grayscale image (1, 28, 28), Tanh activation [-1, 1]
• Parameters: ~1.9M
• Conditioning: label embedding (10-dim) concatenated with noise

Training

• Dataset: MNIST (60,000 images)
• Epochs: 50
• Optimizer: Adam (lr=0.0002, betas=(0.5, 0.999))
• Loss: BCELoss with label smoothing (0.9)
• G steps per D step: 2

Usage

from cgan_model import Generator
import torch

G = Generator()
G.load_state_dict(torch.load("mnist_cgan_generator.pth", map_location="cpu"))
G.eval()

z = torch.randn(1, 100)
label = torch.tensor([7])
with torch.no_grad():
    img = G(z, label)  # (1, 1, 28, 28) in [-1, 1]