I’m trying to train an rfdata model on a custom dataset. I’d like to train two separate models - one optimized for inference at 640p and another at 1280p.
All of my dataset images have the same native resolution, so I’m wondering if it’s possible to specify custom width and height values during training to match that.
Right now, I can set the resolution hyperparameter to an integer (for example, 640), but I’m not sure how that actually works - does it automatically make the images square, or is there a way to keep different width and height values?
Hi @Amanda_Marino,
Great question! You can specify the training resolution using the resize preprocessing step. Specifically, we recommend implementing the stretch to resizing step.
Out of curiosity, which RF-DETR variant are you training?
I’m doing my training in python using the rfdetr sdk. I’m trying to train the small model, RFDETRSmall(). I have not considered doing any preprocessing myself.
Does setting the resolution hyperparameter to 640 automatically resize my data using stretching? Would they be resized to [640x640]? My original data is [3840 x 2160].