I am currently going through the “train_faster_rcnn.py” code to finetune the model on my custom dataset. I wonder how can i modify the anchor ratios and scales values ? as my dataset has objects of various different sizes.
Also i am thinking to play with the “first_stage_max_proposals” , i.e. number of proposals passed from stage 1 to stage 2 ? how can i modify/fine tune these parameters.
You would have to implement your FasterRCNNTrainTransform class.
Take a look at the code here to see how the current version is implemented: https://gluon-cv.mxnet.io/_modules/gluoncv/data/transforms/presets/rcnn.html#FasterRCNNDefaultTrainTransform
I believe the number of proposals are defined the in the actual Faster-RCNN hybridblock code. You can find that here: https://gluon-cv.mxnet.io/_modules/gluoncv/model_zoo/rcnn/faster_rcnn/faster_rcnn.html#FasterRCNN