ACGAN_mnist

An Auxiliary Classifier GAN(ACGAN) in PyTorch to generate MNIST digits.

Literature survey

• The orginal ACGAN paper - [https://arxiv.org/abs/1610.09585]
• Code heavily inspired by dcgan pytorch repo.
• A lot of tuning and hacks borrowed from this awesome repo.

Tips

Keep in mind that these tips are those which worked in this specific project. These may or may not work if you are generating stuff other than MNIST digits.

General

• Keep the architectures of G and D symmetric.
• Use ReLU in G and BatchNorm + LReLU in D.
• Add dropout layers in D to make it 'weaker'.

Things that drastically improved performance

• Label smoothing for D
• Label flipping for D
• Adding Dropout layers in G

Results

• epoch 2 iteration 200/700

• epoch 4 iteration 500/700

• epoch 6 iteration 500/700

• epoch 8 iteration 400/700

• epoch 10 iteration 200/700