DCGANs

Implementing a DCGAN on MNIST digits data and CIFAR-10 data using Tensorflow.

Generally DCGANs contains two networks, a Generator and a Discriminator. So as name says generator keeps on generating images by taking in some noise and discriminator distinguishes between original images and generated images.

Generator

Generator takes in some noise as input and we need to follow the following rules which are experimental observations as mentioned in paper for good results:

• We need to use Fractional strided convolution(or Transposed convolution) to upsample so that it can learn its own upsampling.
• We need to use ReLU activation in all layers except in last layer where we use tanh activation.
• We need to use Batch Normalization in all layers except output and we need not use dense layers on top of conv layers as we need to return an image.

Discriminator

Discriminator takes in images as input and we need to follow following rules which are experimental observations as mentioned in paper for good results:

• We need to use strided convolution instead of max-pooling layers to downsample so that it can learn its own downsampling.
• We need to use LeakyReLU activation with a leak of 0.2-0.3 in all layers except in last layer where we use sigmoid activation as this is used for binary classification task.
• We need to Batch Normalization in all layers except output and we need not use dense layers but in the output layer we just use one unit as output (real or fake)

Results

GIFs created using generated images by generator.

The one on the left was trained for 100 epochs which took around an hour and the one on the right was trained for 40 epochs which took around one and half hour.

References

• Ian Goodfellow et al., Generative Adversarial Networks

• Alec Radford et al., Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

© V I S H W A S