mxnet implement for Conditional Wasserstein GAN
You only need to change the is_wgan flag to test wgan or dcgan results. change is_mlp to test mlp result.
generate result after 30 epochs.
wgan:
dcgan:
mlp wgan:
note:
- because we want to try mlp result, so I flatten the input image to a vector, then append condition one-hot vector to the vector. If you only want to try Convolution ops, you will not need to add condition in this way
- wgan seems not better than dcgan, maybe something wrong. If you know what happen, please let me know.
- you can learn a lot about how to construct data iterator, how to backward without output layer
Code borrows from mxnet gan example, wgan metric from WGAN