Hi, I’m taking part in a satellite image competition using fastaiV2, and the library really helps me rank top N with a wonderful coding experience.
My question is the following. In my training, I crop the (256, 256) sized image from (1024, 1024) sized image to meet the GPU memory limitation and still validate and calculate metrics on (1024, 1024). I have to set dls.valid.bs=1 since the image is too big. Sadly, it takes about 8 minutes in the validation phase of every epoch .
Can I change the learner’s validation frequency, like validating every 10 epochs rather than every epoch? This could greatly accelerate the training process.