mohcinemadkour / Semisupervised-learning-with-DCGANs

Using DCGANs to learn useful representations during the adversarial training process and using the learned features to classify images with relatively little training data

Semisupervised-learning-with-DCGANs

It is a work in progress

==> Using DCGANs to learn useful representations during the adversarial training process and using the learned features to classify images with relatively little training data

==>Main aim of this project : To use the power of unsupervised representation learning with DCGANs to build discriminative image classifiers which can be trained with relatively little training data as compared to a fully supervised paradigm using a semisupervised learning approach.

==>Heavily influenced by :
Paper - Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

Code - Keras Implementation of DCGANs

Currently works for Keras with Tensorflow Backend. Will add theano support soon.

==>Steps involves in the process,
1. Train the Deep Convolutional Generative Adversarial Networks with the datset in an unsupervised manner.
2. Reuse the Discriminator for semi supervised classification on the MNIST/CIFAR10 datasets.
   -> This is done by freezing the weights of the discriminator network while supervised training.
  -> Currently working with one FC softmax layer on top of the leaned feauture layers

==> The trained weights are stored in their respective HDF5 files.

==>Instructions for running: For working with the MNIST dataset,

Training the model - python mnist_gan.py --mode train --batch_size 128 --epoch_num 200

To generate Images - python mnist_gan.py --mode generate --batch_size 64 --nice

To run the semisupervised classifier - python mnist.py

For working with the CIFAR 10 Dataset,

Training the model - python cifar_gan.py --mode train --batch_size 128 --epoch_num 200

To generate Images - python cifar_gan.py --mode generate --batch_size 64 --nice

To run the semisupervised classifier - python cifar.py

For working with the google SVHN dataset,

The code assumes that the 'train_32x32.mat' and 'test_32x32.mat' files are present in the current project directory.

Training the model - python svhn_gan.py --mode train --batch_size 128 --epoch_num 100

To generate Images - python svhn_gan.py --mode generate --batch_size 64 --nice