Hello,
I trained a pretrained faster RCNN with resnet50 FPN and I would try to do the same thing with a resnet101.
Is there a way to train a pretrained faster RCNN with resnet101 FPN as backbone ?
And ideally to change the anchor boxes sizes (with a pretrained mode on resnet 101 or resnet 50)? Because I know that my objects are supposed to have always the same shape.
anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
aspect_ratios = ((0.5, 0.75, 1.0),) * len(anchor_sizes)
anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)
# get the number of input features
in_features = modelr.roi_heads.box_predictor.cls_score.in_features
# define a new head for the detector with required number of classes
modelr.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
modelr.rpn.anchor_generator = anchor_generator #adding in customized anchors