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
Note: results are questionable.. re-testing with fixes at the moment.
| airplane | automobile | bird | cat | deer | dog | frog | horse | ship | truck | |
|---|---|---|---|---|---|---|---|---|---|---|
| Fully supervised | 0.7780, 0.8907, 0.6233, 0.5606, 0.6891, 0.6420, 0.7967, 0.8206, 0.8619, 0.8291 | |||||||||
| DeepInfoMax-Local | 0.6120, 0.6969, 0.4020, 0.4226, 0.4917, 0.5806, 0.6871, 0.5806, 0.6855, 0.5647 |
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..
