Training, error: can't pickle Environment objects when num_workers > 0

[annalina111]

Hi, anybody else run into an issue when trying to run either of the training scripts, due to num_workers being set to 8 when instantiating a DataLoader? When i've changed this to 0 (line 112 and 113 in train_rave.py and corresponding lines in train_prior.py), I can run the script fine, but I'm wondering if I can get checkpoint outputs more quickly if I can solve this..

here is a relevant github discussion:
pytorch/vision#689
there seems to be an understood solution here but I am having trouble figuring out how to implement these insights within the RAVE code..

[vigliensoni]

Same here!

[bjorkfan59]

I'm having the same issue! If I change num_workers to be 0, the program recommends that I use 8 (because my computer has 8 cores), so it should be able to run fine with 8 workers. Not really sure what the problem might be. Also, when I try to run with 0 workers, I run into a new error:

ValueError: n_components=128 must be between 0 and min(n_samples, n_features)=64 with svd_solver='full'

I'm running on a macbook without a GPU. I'm trying to train on a fairly small dataset just to get things going, a folder with 130 .wav files that are all just a few seconds each. Any advice?

[bjorkfan59]

Ah. I should have read the setup page. The issue is probably that I was trying to run this on mac lol