Another wonderful week passed by and we have exciting news to share! Both the
Batch Support and
DCGAN PRs are now merged, and we have begun our work on
Dual Optimizer and
Wasserstein GAN, which would be our priority for the coming two weeks.
After all the time spent optimizing and training the network, we have been able to reduce the training time of ~80 hours with singular inputs, to less than 10 hours with a mini-batch of 50 inputs. This is atleast 1.5x faster than the corresponding training time of
Tensorflow on CPU! Here are some results that we obtained:
We're currently in the process of deciding the optimal structure for the
Dual Optimizer class, and implementing
Gradient Penalty and
Weight Clipping modules for
WGAN. After these issues are sorted, we'd be free of our main goals by Phase II evaluations!
Generated by 1.8.13