Retrieving the logits for the class predictions for the Faster RCNN model in torchvision


I have followed the Object Detection Finetuning Tutorial to the train torchvision’s Faster RCNN model on my own data. I have found that the model only returns the score for the predicted category of each bounding box, but I would like to retrieve the logits behind each prediction (before softmax).

Is there a way of accessing those values? The only approach I could imagine is to copy all relevant source code, modify the respective return statements and then manually copy all the weights over to my “new” model. If anybody has an idea for a less ugly approach it would be highly appreciated!



Have you found a solution already?



Does anyone have an answer for this? I am also interested.

Any updates? :slight_smile: @ptrblck