Aditya Oke
Aditya Oke
Yes 😀
I think we should discuss more over this. Datasets is really tricky especially when it comes to object detection etc. For the Torchvision models, we expect VOC format. And for...
Let's atleast support some parts for now. Increasing this priority slightly.
This is really tricky. Let me explain in bit detail.
This is actually something we should look in a longer run for major refactor. For Classification Torchvision provides CNNs (backbones) trained over imagenet. We extended this to use any backbone,...
In short. We need to support the following - [x] All Non FPN backbones on imagenet and other weights (through torchvision backbones) - [x] Resnet FPN Based backbones through imagenet....
Hmm, let me start an initial refactor. This refactor is little tricky.
These two are super hard to support. > Resnet FPN Backbones on other weights. 1. Resnet FPN Backbones on other weights is somehow not possible due to `hardcoding` by torchvision...
The above PR, reduces this urgency by sometime. There can be better solution but we need training for most weights.
I will be happy to take losses :) dice loss first.