======================================================================================

As far as I know, there is no light-weight implementation of DCGAN based on caffe.

Inspired by DeePSiM implementation, a few lines of python code can train the dcgan model quickly without any hack in caffe core lib (Dosovitskiy has already done this. However, I think the code could be merged back to master branch).

You will need to compile the deepsim-caffe-branch. And make sure your PYTHONPATH point to it.

The deepsim-caffe only support cudnn-4.0. If disable the cudnn engine and replace some convolution layers with the master branch, a latest cudnn and cuda will work fine.

For face generator, please prepare celebA dataset as the link said. Then make a train list file and put it in the data.prototxt.

Just typing

python train.py

To view the model result by

python generate.py generator.prototxt snapshots_test/4000/generator.caffemodel

The visualizations of the models at iteration 3000 and 4000 are as following: