We have formally entered a streak of productive weeks! The
WGAN PR is now complete and under review, and should be ready to merge anytime next week. We have totally refactored the
GAN module according to
enable_if<> paradigm, allowing us to choose the variant of
GAN at compile time.
With the above changes, we have further reduced the training time to under ~7 hours with all the variants! We tested our implementations, again on the full
MNIST dataset, and obtained the following results:
We're still planning out the optimal startegy for implementing
DualOptimizer class. Thankfully, Marcus has offered his help (as always) regarding the same. We have also planned to complete the
Spike and Slab RBM work left over from Kris'
GSoC last year. We have been very fortunate to cover a lot up until now, and probably we can cover a lot more this month as well.
Generated by 1.8.13