Heikki Arponen

also recommend commenting out all nonpytools stuff... or try to install it on e.g. Ubuntu VM. I'll try to sort it out meanwhile

oops I guess I forgot to add that in the dependencies... the module is actually "python-midi" so just pip install that EDIT: no it's in the readme...

Thanks for reporting this! Actually cython + windows are known not to play along well, so I'd suspect that's the culprit here... not sure though. You actually probably won't need...

Just a FYI I'm experiencing similar slowdowns with transforming keypoints taking about 5x longer than heavy augmentations for images... I was a bit horrified to find out that the keypoint...

using albumentations nowadays most of the time, and haven't had much need for keypoint augs since.

Are you using the `seed_everyhting` function (recommended in PyTorch Lightning) and `ddp` parallel training? That might set the same seed on each process => exact same shuffling... I had a...

Actually if you do multi-node ddp, then you need to use seed_everything to ensure the distributed models are initialized to be identical, unless there's a very recent update that fixes...

@tchaton sure, although I really can't say how much time I can spend on that... FYI everything is working fine for me with Pytorch Lightning + webdataset, after having some...

Well I kinda gave up on torch-xla... things were just mysteriously slow and it's pretty hard to debug all the issues. So using CUDA only now.

You should open, decode and resize each image to same shape, e.g. [512, 512, 3], then store all 1.2M images into a [1.2M, 512, 512, 3] size TensorStore. Then accessing...