How to extract features from pre-trained FasterRCNN pytorch?

I need some help.
I would like to use the pre-trained model Faster from pytorch package :

import torchvision
import torch
from torchvision.models.detection import fasterrcnn_resnet50_fpn
model = fasterrcnn_resnet50_fpn(pretrained=True)

Faster RCNN has 2 outputs : (label, bbox) for each Region that it has selected. I would like to extract, for each region, the vector from the last fully connected layer. The vector where I can find all the probabilities of belonging to the different possible categories. So the vector just before the softmax …

For example, with Coco dataset (91 categories), I would like to have one vector [1,91] for each region.

I have read a lot of solution for VGG, Resnet … but nothing about pre trained faster model.
Can someone help me ?