DCGAN model

[HarshithBachimanchi]

This PR introduces DCGAN architectures into deeplay. The models are implemented as described in the original DCGAN article (link).

Example:

Creating DCGAN generator

generator = dl.DcganGenerator(
    latent_dim=100,
    features_dim=64,
    output_channels=1,
    class_conditioned_model=False,
    num_classes=10, # only used if class_conditioned_model=True
    embedding_dim=100, # only used if class_conditioned_model=True
)

Creating DCGAN discriminator

discriminator = dl.DcganDiscriminator(
    input_channels=1,
    features_dim=64,
    class_conditioned_model=False,
    num_classes=10,  # only used if class_conditioned_model=True
    embedding_dim=100,  # only used if class_conditioned_model=True
)

As DCGAN is a fixed architecture, the images need to be resized to 64 x 64 before passing through the models. The number of Conv blocks in discriminator and the ConvTranspose blocks in the generator are fixed to 4 as per the paper. However, the number of features in each block can be controlled by changing the features_dim attribute.

For generator, number of features in each ConvTranspose block goes as [features_dim*16, features_dim*8, features_dim*4, features_dim*2]

For discriminator, the number of features in each Conv block goes by [features_dim, features_dim*2, features_dim*4, features_dim*8].

By changing the features_dim parameter, the relative strength of generator and discriminator can be controlled.

[BenjaminMidtvedt]

I think it's looking good now, nice work! I'm good to merge