Open kristinagrig06 opened 3 years ago
Hey @kristinagrig06 @tatevikh, very cool project! Can I collaborate in this issue? I am new to open source but I have experience in software development (also in a team). I also have experience with pytorch and tensorflow.
Hey @MoritzWillmann, welcome to Hub and Hacktoberfest, and thank you for your willingness to contribute! Do you have a proposed solution in mind for this issue?
Hey @dhiganthrao, I checked out some of your code for .numpy()
yesterday and I think I'll first implement it similar to that. I read in pytorch forums though that transforming a buffer to torch.Tensor
directly can be slow so I'd benchmark it against a call torch.from_numpy(np.frombuffer(...))
. Do you have any thoughts on this? I didn't look into tensorflow yet, but I think it'll be similar.
I think for now .pytorch() and .tensorflow() can call .numpy() underneath. @AbhinavTuli ?
Hey @farizrahman4u, I was just about to suggest that. I implemented it both ways yesterday and there's no recognisable performance difference. The "no numpy"-version needs much more implementing work though due to datatypes. Also it seems like there will be a major change in how pytorch handles storage for different datatypes coming up soon. It should get easier then...
@MoritzWillmann sounds good.
hey i want to solve this issue please assign this issue to me
🚨🚨 Feature Request
If your feature will improve
HUB
Currently, only Datasets can be used to create PyTorch Dataloaders, TensorFlow Datasets. Need to add functionality to convert Hub tensors to PyTorch/TensorFlow compatible format.
Difficulty: Medium