PyTorch implementation of the DCGAN introduced in the paper "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks" by Radford et al. on the Anime Faces Dataset by McKinsey666.
DCGAN, introduced by Radford et al. in 2015, is a seminal work in the field of generative models. Its key contributions are as follows:
- Architecture Guidelines:
- DCGAN replaces fully connected layers with convolutional layers in both the generator and discriminator.
- It employs batch normalization to stabilize training.
- Pooling layers are removed, and strided convolutions are used for downsampling.
- ReLU activation functions are used in the generator (except for the output layer, which uses tanh).
- LeakyReLU activation functions are employed in the discriminator.
- Stable Training:
- DCGAN demonstrates that following the architectural guidelines leads to more stable training of GANs.
- It mitigates issues like mode collapse and vanishing gradients.
- Image Synthesis:
- DCGAN achieves impressive results in generating high-quality images.
- It can synthesize realistic images from random noise vectors.
- Latent Space:
- The generator learns a meaningful latent space where different directions correspond to distinct features.
- By interpolating between points in this space, smooth transitions between generated images can be created.
The generator architecture as shown in the paper -
The Anime Face Dataset was created by Mckinsey666. This dataset has 63,632 "high-quality" anime faces.
The data has been preprocessed using torchvision.transforms by resizing, cropping and normalizing the images.
I have trained the model using the hyperparameters used in the original paper. Feel free to play with the parameters in the colab notebook. I am planning to convert this notebook into individual scripts in the future!
After training for 25 epochs, you can see the results in the GIF below -
