Issue with ToTensor Transform and Multiprocessing in PyTorch Lightning

[DaniMlk]

Hi,

I’m encountering an issue with the ToTensor transform when using multiprocessing within PyTorch Lightning. I’ve implemented a custom dataset that reads data, applies transformations, and places the processed data into a queue using multiprocessing. Then, in the getitem method, I fetch the data from the queue.

The process works flawlessly outside of the PyTorch Lightning environment, but when I integrate it into PyTorch Lightning and attempt to apply the ToTensor transform, the data suddenly disappears with the error message 'no more image', which also disappears shortly after. Interestingly, when I move the transform application directly to the getitem function, it works fine.I also dug into the transform and noticed the issue is when it tries to change the data type.