Contributing to torchvision


I have finished creating three datasets that are not available in torchvision, and that I think it might be interesting to have them available. I have take a look on how to contribute and it says I should first post about it in order to discuss the implementation details.


1 Like

Could you post your current working branch?
CC @fmassa


For the moment I have only uploaded tinyImagenet (still preparing the others to be as closed as possible to the other datasets in torchvision). Files are selfcontent and prepared at least for torchvision 0.2.0. What I have done is just copy the new functions from 0.3.0 that do not appear in 0.2.2.

I have finished uploading the three datasets, though I will add some comments we can discuss about them. Btw, there are design choices to be done. For the moment two of the three datasets has been split into two datasets (one for classification and one for object detection). This is because images have different shapes and for classification a preprocess is done to avoid heavy computations during training. We can discuss about this, I have just made it as an initial approximation