GAN

Table of Contents

1. Architecture
2. Dataset
3. Usage
4. References

Architecture

Generative Adversarial Networks (GANs) are composed of two smaller fully connected networks [1]. The generator network used to created synthetic images based on the underlying distribution of the training data and the discriminator network used to train the generator network by determining if a batch of images were either synthetically generated or not. These two networks are trained in opposition to each other &mdash the generator network aims to fool the discriminator network, and the discriminator network aims to correctly identify synthetic images.

Dataset

You can download the MNIST dataset [2] using the following command:

python3 data/download_mnist.py

This command will download place the dataset at the following path: ./data/datasets.

The dataset consist of 70000 grayscale 28x28 images split into 10 classes, with approximately 7000 images per class. There are a total of 60000 training images and 10000 test images. The classes (in numerical order) in the dataset are:

• 0
• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9

Only the 60000 training images will be used to trained the generator and discriminator network.

---

You can download the FashionMNIST dataset [3] using the following command:

python3 data/download_fashionmnist.py

This command will download place the dataset at the following path: ./data/datasets.

The dataset consist of 70000 grayscale 28x28 images split into 10 classes, with approximately 7000 images per class. There are a total of 60000 training images and 10000 test images. The classes (in alphabetically order) in the dataset are:

• Ankle Boot
• Bag
• Coat
• Dress
• Pullover
• Sandal
• Shirt
• Sneaker
• T-shirt/Top
• Trouser

Only the 60000 training images will be used to trained the generator and discriminator network.

Usage

You can train a generator network on MNIST using the following command:

python3 -m utils.trainer \
        --root="./data/datasets" \
        --dataset="MNIST" \
        --batch_size=32 \
        --num_workers=4 \
        --pin_memory \
        --num_epochs=20 \
        --image_height=32 \
        --image_width=32 \
        --generator_lr=0.0001 \
        --discriminator_lr=0.0001 \
        --optimizer="Adam" \
        --beta1=0.5 \
        --beta2=0.999 \
        --latent_dim 100 \
        --generator_in_features 128 256 512 \
        --discriminator_in_features 512 256 128  \
        --use_spectral_norm \
        --ckpt_dir="./checkpoints" \
        --save_ckpt_interval=20 \
        --results_dir="./results" \
        --save_results_interval=2 \
        --nrow_for_saved_samples=9 \
        --random_seed=0

The command above will produce the following results. You can find the model checkpoint here.

Training (10 epochs):

Samples:

---

You can train a generator network on FashionMNIST using the following command:

python3 -m utils.trainer \
        --root="./data/datasets" \
        --dataset="FashionMNIST" \
        --batch_size=32 \
        --num_workers=4 \
        --pin_memory \
        --num_epochs=20 \
        --image_height=32 \
        --image_width=32 \
        --generator_lr=0.0001 \
        --discriminator_lr=0.0001 \
        --optimizer="Adam" \
        --beta1=0.5 \
        --beta2=0.999 \
        --latent_dim=100 \
        --generator_in_features 128 256 512 \
        --discriminator_in_features 512 256 128 \
        --use_spectral_norm \
        --ckpt_dir="./checkpoints" \
        --save_ckpt_interval=20 \
        --results_dir="./results" \
        --save_results_interval=2 \
        --nrow_for_saved_samples=9 \
        --random_seed=0

The command above will produce the following results. You can find the model checkpoint here.

Training (10 epochs):

Samples:

References

[1] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, pages 2672–2680, 2014.

[2] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. In Proceedings of the IEEE, pages 2278–2324, 1998.

[3] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.