fundamentalvision / Deformable-DETR

Deformable DETR: Deformable Transformers for End-to-End Object Detection.

Apache License 2.0

3.15k stars 513 forks source link

Easy Finetuning and Inference #181

Open robinnarsinghranabhat opened 1 year ago

robinnarsinghranabhat commented 1 year ago

Contribution of this PR:

Support while finetuning Deformable-DETR on custom-data. Checkout : how_to.md
Provided a Inference Notebook
Following Error : ImportError:` cannot import name '_NewEmptyTensorOp' from 'torchvision.ops.misc'
Additional guidelines for easy setup for gcloud gpu instance.

robinnarsinghranabhat commented 1 year ago

Potentially relevant to the following issues:

49

97

117

116

nikhilchh commented 1 year ago

Hey @robinnarsinghranabhat I was trying to find the inference notebook but couldn't find it in your repo. Can you point me to that please. Thanks :)

nikhilchh commented 1 year ago

In deformable-DETR code they don't add extra +1 for background because of loss has been designed. (There are some discussions about it in the other issues). But you have mentioned to add +1 in your instruction for fine tuning.

nhattan417 commented 7 months ago

More information at: fundamentalvision/Deformable-DETR#78 and fundamentalvision/Deformable-DETR#72. (TL;DR due to the difference between Softmax+CE and Sigmoid+BCE)