What exactly is happening in segtrain? I tried to train a model from Scratch on roughly 300 images. The model page gives 66% as a result after 11 epochs and stops there. And the results are terrible. I train 11 epochs with the same material on the gpu and the results are ok, not fantastic but workable:
- We know that the automatic evaluation does not correspond to manual evaluation. Could we rather do a fix number of epochs and save every 10th or every 50th model as Ben and I are doing on our gpus?
- Put as little as possible into the testset in order to keep the trainingset as big as possible.