Closed kastnerkyle closed 9 years ago
Hi Kyle
Thanks for pointing that out. I agree that the resampling makes the results non comparable to the *NADE results but on other hand directly comparable to the IWAE results by Burda (thats why we choose the resampling procedure).
I think I'll just add a note highlighting the issue and a switch to choose between the binarized dataset provided by Hugo Larochelle and the resampled MNIST dataset.
btw it gives 1-2 NATS to do the resampling every epoch.
That all sounds nice. Since there are so many different "binarized" MNISTs it will be nice to compare the difficulty in modeling a few different types.
On Sat, Oct 10, 2015 at 9:37 AM, Casper Sønderby notifications@github.com wrote:
Hi Kyle
Thanks for pointing that out. I agree that the resampling makes the results non comparable to the *NADE results but on other hand directly comparable to the IWAE results by Burda (thats why we choose the resampling procedure).
I think I'll just add a note highlighting the issue and a switch to choose between the binarized dataset provided by Hugo Larochelle and the resampled MNIST dataset.
btw it gives 1-2 NATS to do the resampling every epoch.
— Reply to this email directly or view it on GitHub https://github.com/casperkaae/parmesan/issues/8#issuecomment-147089983.
Agreed, thats left for future work :)
The bernoulli resampling is happening every epoch - this matches the IWAE paper but is a bug, since resampling is effectively data augmentation the result cannot be compared to VAE on a fixed binarization (such as that provided by Hugo Larochelle here http://www.dmi.usherb.ca/~larocheh/mlpython/_modules/datasets/binarized_mnist.html).
In particular this line: https://github.com/casperkaae/parmesan/blob/master/examples/iw_vae.py#L289
They mention this issue at the bottom of the github repo https://github.com/yburda/iwae