Nicolas Granger

I have pushed this idea further with a realistic scenario and it seems that torch.rpc doesn't scale well for a dataloader usage. I observed a ton of CPU overhead both...

The model from the paper and the data pre-processing part are not overly complicated at first sight, but the prediction algorithm (prefix search) might require some work. I'll try to...

It seems there are some precision issues on real world data (TIMIT speech). I need to investigate that first. When I get it to work reliably I think I will...

The latests commit should fix most precision issues, but there is still some divergence at some point. The final output layer values (before softmax) explode at some point in the...

Ok, the damn error is fixed now, both loss and gradient are now in line with tensorflow's implementation. Thanks for reading through this, I have corrected the pickle line. I...

I wanted something that would remain robust even with optimizations off, especially because I was debugging it myself ;-) so the handwritten logsumexp is safer when properly implemented. For the...

I think french IGN maps look a bit similar with a bit more contrast, you can have a [live example here](https://www.geoportail.gouv.fr/carte?c=2.2697018925436296,48.51931139893438&z=15&l0=GEOGRAPHICALGRIDSYSTEMS.MAPS.SCAN-EXPRESS.STANDARD:WMTS(1)&l1=GEOGRAPHICALGRIDSYSTEMS.MAPS::GEOPORTAIL:OGC:WMTS(1)&l2=GEOGRAPHICALGRIDSYSTEMS.MAPS.SCAN25TOUR.CV::GEOPORTAIL:OGC:WMTS(1)&permalink=yes).

@Eitetsu0 fractional scaling is not supported by xwayland AFAIK and the point of this PR is just to expose the scaling value as a config option instead of letting sway...

As far as I can tell, this is the same bug as #3345.

Here are some additional random suggestions, but I don't think I'd use that myself because it might prove distracting: - A rehearsal mode to generate a timing file for each...