[TensorFlow] Conditional Generative Adversarial Nets (CGAN)

TensorFlow implementation of Conditional Generative Adversarial Nets (CGAN) with MNIST dataset.

Architecture

Training algorithm

The algorithm for training CGAN [1].

CGAN architecture

The architecture of CGAN [1].

Graph in TensorBoard

Graph of CGAN.

Results

Training Procedure

Loss graph in the training procedure.
Each graph shows loss of the discriminator and loss of the generator respectively.

Test Procedure

From random noise without conditions

z:2	z:64	z:128

From random noise with conditions

z:2	z:64	z:128

Latent space walking with conditions

Class-0 (z:2)	Class-1 (z:2)	Class-2 (z:2)	Class-3 (z:2)	Class-4 (z:2)

Class-5 (z:2)	Class-6 (z:2)	Class-7 (z:2)	Class-8 (z:2)	Class-9 (z:2)

Environment

• Python 3.7.4
• Tensorflow 1.14.0
• Numpy 1.17.1
• Matplotlib 3.1.1
• Scikit Learn (sklearn) 0.21.3

Reference

[1] Mehdi Mirza and Simon Osindero. (2014). Conditional Generative Adversarial Nets. arXiv preprint arXiv:1411.1784.