Home

Awesome

Deep InfoMax Pytorch

Pytorch implementation of Deep InfoMax https://arxiv.org/abs/1808.06670

Encoding data by maximimizing mutual information between the latent space and in this case, CIFAR 10 images.

Ported most of the code from rcallands chainer implementation. Thanks buddy! https://github.com/rcalland/deep-INFOMAX

Pytorch implementation by the research team here

Current Results (work in progress)

airplaneautomobilebirdcatdeerdogfroghorseshiptruck
Fully supervised0.77800.89070.62330.56060.68910.64200.79670.82060.86190.8291
DeepInfoMax-Local0.61200.69690.40200.42260.49170.58060.68710.58060.68550.5647

alt_text

Figure 1
Top: a red lamborghini, Middle: 10 closest images in the latent space (L2 distance), Bottom: 10 farthest images in the latent space.

Some more results..

alt_text

alt_text

alt_text

alt_text