In this work, introduce the Deep Convolutional Generative Adversarial Network (DCGANs), originally introduced by Radford et al. (2015), to generate images of pets, specifically cats and dogs.
Architecture guidelines for stable DCGANs:
- Replace any pooling layers with strided convolutions (discriminator) and fractional-strided convolutions (generator).
- Use batchnorm in both the generator and the discriminator (not applying batchnorm to the generator output layer and the discriminator input layer).
- Remove fully connected hidden layers for deeper architectures.
- Use ReLU activation in generator for all layers except for the output, which uses Tanh.
- Use LeakyReLU activation in the discriminator for all layers.
This project uses the AFHQ-v2 (Animal Faces-HQ v2) dataset, which contains high-quality animal face images. The original dataset consists of 3 classes: cat, dog, and wild (wild animals). Only the cat and dog classes are included, while the wild class is excluded. Each image has a resolution of 512×512. However, to enable faster training and more efficient metric evaluation, I use the 64×64 version available on Hugging Face. The training set contains 9,892 images, while the development (dev/validation) set contains 1,000 images.
| Model | FID score | Describe |
|---|---|---|
| DCGAN v.0 | 43.1442 | Implemented as in the DCGAN paper |
| DCGAN v.1 | 37.3660 | Replacing ReLU with GeLU in Generator |
| DCGAN v.2 | 31.2451 | Replacing BatchNorm with PixelNorm and use residual connection in Generator, EMA (decay=0.9996) for generate image |
All experiments are configured with the same hyper-parameters:
epoch = 300, batch_size = 128, latent_dim = 100- Input normalization: scale input images to range [-1,1] (to match the Tanh output the Generator).
- Weight initialization: initialize weights from a normal distribution ~ N(0, 0.02).
- Optimizer: Use Adam optimizer with learning_rate=0.0002, beta1 = 0.5 (instead of the default 0.9).
Install required packages:
pip install -r requirements.txt
Training:
python train.py
Generate new images:
python generate.py --num_images 64
[1] Alec Radford, Luke Metz, Soumith Chintala (2016). Unsupervised representation learning with deep convolutional generative adversarial networks. [arXiv]
[2] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio. Generative Adversarial Networks. [arXiv]
[3] Natsu6767. DCGAN-PyTorch. [Github]
[4] Yunjey Choi, Youngjung Uh, Jaejun Yoo, Jung-Woo Ha. Animal Faces-HQ v2. [Dataset]