I have some experience working with pytorch would like it to know if it would helpful developing a plugin for DALI for the same.
Hi,
our current PyTorch integration consist of the iterators:
https://docs.nvidia.com/deeplearning/dali/user-guide/docs/plugins/pytorch_tutorials.html
and our ExternalSource operator can accept PyTorch tensors.
If you have any ideas on how to make our PyTorch integration better or how to extend it we are happy to discuss them here or see and review any PRs.
Hi @Hemantr05 ,
One idea that comes to my mind is to extend the DALI PyTorch iterator to produce sparse tensors. You can check GluonIterator SPARSE_TAG and DENSE_TAG for a reference.
@JanuszL would there happen to be a discussion forum or a platform where the above can be discussed ?
@klecki Would developing a computational graph of the same be helpful?
@Hemantr05 - GitHub is our platform. I think we can discuss it here, in this issue.
Could you elaborate more on what exactly you have in mind by a computational graph of the same?
@JanuszL
Creating a computational graph for DALI which shows the workflow or a visualization toolkit like tensorboard.
Creating a computational graph for DALI which shows the workflow or a visualization toolkit like tensorboard.
Plugin DALI to tensorboard would be very nice but I'm afraid that this may be a too big task for the beginning without knowing much about DALI internals. DALI is mostly a C++ library with a thin Python API layer on top. If you are more familiar with Python development then extending PyTorch iterator would be a good task for the beginning. If you are fine with working with the native code then go ahead with the tensorboard.
@JanuszL Will look into PyTorch Iterator for now.
Hello @JanuszL could you please assign me this issue?
I have created https://github.com/NVIDIA/DALI/issues/2365, but I cannot assign it to you (it appears that I can assign only people from the Nvidia organization). Anyway, please feel the owner of the task.
Oh alright! @JanuszL thanks!
Hi, @JanuszL Apologies for the delay.
Will resolve this by the end of the week.
Sure, not problem. We are not pushing.