Implementation of a few types of Generative Adversarial Networks in Keras.

• Train an ACGAN on the MNIST dataset using class priors. Inputs are (100) dimensional noise vectors along with (10) dimensional class vectors (conditioning) and the outputs are generated (1,28,28) images
• Samples of generated outputs at various epochs in ACGAN-MNIST/Run1/Results
• Trained generator and discriminator models in ACGAN-MNIST/Run1/Models
• Generated samples in ACGAN-MNIST/Samples
• The samples generated are much sharper than other generative methods such as variational autoencoders etc.
• Some cherry picked generations below

| | | | |

| | | | |

• Train a DCGAN on the MNIST dataset without any class priors. Inputs are (100) dimensional noise vectors and the outputs are generated (1,28,28) images
• Samples of generated outputs in DCGAN-MNIST/GeneratedOutputs
• Trained generator and discriminator models in DCGAN-MNIST/TrainedModels
• Generated digit samples below

| | | | |

| | | | |

Note: Training the GAN for longer would give us much better results, which are not as blurry.