keras_DCGAN_transfer_learning

Building a Keras DCGAN using a pre-trained Torch generator

Trying to use the power of Deep Learning to generate new images, but not willing to invest in hindend GPU? Transfer learning could be a solution.

keras_dcgan.py builds and trains a DCGAN model (https://arxiv.org/abs/1511.06434). You can choose to either use the generator from the original paper, or a similar FloyedHub implementation (https://github.com/ReDeiPirati/dcgan).

In the latter case, you can choose to copy the weights of the pre-trained model. Those you can download from here https://www.floydhub.com/redeipirati/datasets/dcgan-300-epochs-models/1/netG_epoch_299.pth, but for convenience I have already include the model in the weights directory.

Note that the FloyHub model is built with Torch (see torch_dcgan.py). To translate this to Keras Torch2Keras.py:

1. builts the Torch model and loads its weights
2. builts a Keras model with the same architecture
3. extracts weights as numpy arrays from the first model and feeds them to the second
4. saves weights of Keras model in weights sub-directory

The converter works only for the given architecture, for a more general one try https://github.com/nerox8664/pytorch2keras

The FloyHub pre-trained model is trained on the `Labeled Faces in the Wild' database (http://vis-www.cs.umass.edu/lfw/). The corresponding Keras implementation gives:

To apply this to your project see main.py. I am currentely working on switching from 'faces in the wild' to 'simpsons faces'.