pytorch / captum

Model interpretability and understanding for PyTorch
https://captum.ai
BSD 3-Clause "New" or "Revised" License
4.8k stars 485 forks source link

With the release of Detectron2 will we see an example on that code base? #139

Closed RaymondCM closed 4 years ago

RaymondCM commented 4 years ago

Detectron2 has been released https://github.com/facebookresearch/detectron2. It would be very nice to see an example using the default models there.

NarineK commented 4 years ago

Thank you for the questions, @RaymondKirk ! Do you want to understand the predictions or do you want to use attributions as an edge detector, how do you see it to be used with Detectron2 ?

RaymondCM commented 4 years ago

@NarineK Detectron2 provides a simple way to run popular detection and segmentation models, the forward pass is generally wrapped in a general trainer function, so isn't too trivially used with this package. I think it would help a lot of people understand how to use this package if an example was made with a Detecron2 model (or even a standard PyTorch model), and would really enhance this repo! So my question was mostly wondering if there's any plans for including more advanced deep detectors RCNNs etc in the examples?

fmassa commented 4 years ago

@RaymondKirk if the abstractions around detectron2 makes it harder to use with Captum, you could try considering using the detection models in torchvision, which can be imported and used without a trainer / config system.

RaymondCM commented 4 years ago

@fmassa Thank you for your input, I'll close this issue as it was merely a sugestion for enhancement.