Lennard Kiehl

Hi @gigony, thanks for the quick reply! My use case is basically using cuCIM/CuImage to generate random patches on the fly from ~10.000 TIFF files. Usually one would pre-generate those...

Thanks @jakirkham! I will have a look at dask_image. But as you seem to be one of the main authors, you might already know the answers to my questions: 1....

Interesting, thanks for your input! I'll have to familiarize myself with dask then and see how it could possibly improve/speed up my approach. But so far I was actually pretty...

Lazy-loading is such a simple and great idea! The overhead is also minimal. Thanks, I will try it out :) Should I close this issue then or do you still...

@guarin Can you explain why the implementation of the loss does not use `proto_labels`/`support_labels` (which are one-hot vectors) like the original does [here](https://github.com/facebookresearch/msn/blob/4388dc1eadbe3042b85d3296d41b9b207656e043/src/losses.py#L38-L42)? They are initialized [here](https://github.com/facebookresearch/msn/blob/4388dc1eadbe3042b85d3296d41b9b207656e043/src/msn_train.py#L219-L227). Maybe I overlooked...

Ok, so as I understand it, the original implementation has the option for label smoothing the `proto_labels`/ `support_labels` in which case they would be different from the identity matrix, which...