Alexei Baevski

Results 13 comments of Alexei Baevski

why is the ddp_backend set to "simple" in that checkpoint? this is not one the values we support (we only support c10d and no_c10d). where did the checkpoint you are...

@shruti-bh do you know why ddp_backend is set to this value?

hey, if variances are jumping up and down, that looks like a collapse and you may want to lower you learning rate. i dont have variance plots for the nlp...

i think this looks reasonable, please submit a PR!

still a problem in any version > 0.17 it seems, including 1.1.1. const enums are supposed to be transpiled into their value, but instead are kept as is which causes...

Nope, putting that into tsconfig as below does not work. Tried both 1.1.1 and 0.19.1 "awesomeTypescriptLoaderOptions": { "forkChecker": true, "useBabel": true, "target": "es6", "disableFastEmit": false }, just to be clear...

Same here, it is really annoying. Also does not happen 100% of the time, but most of the time

it wont have a much of an effect, but you have to match the feature extractor to the normalization setting normalize in dataloader -> layer norm in feature extractor no...

looks like some driver, cuda or pytorch problem, have you tried other versions? what if you run validate.py and point at your checkpoint? does it work if you train with...