xuebinqin
commented
3 years ago Hi, Zhou, Thanks for your interests. Please refer to the detailed training settings in our paper. For the training of the SOD dataset, we followed other papers without using the validation process. What we did before is plotting the -log(loss) of the training process. If the training curve is flat, it should be ok to stop (In the model development process, we actually validate and evaluated all the hundreds of intermediate model weights of many different models, we found there was no obvious overfitting in terms of the F-measure. So don't worry about that if you are training on DUTS-TR and test on the rest datasets mentioned in our paper. We guess that's because all the test datasets are from different groups at different time, whose differences are significant enough.). In your case, if you want to get better results on SOD task, you can also pick one of the other SOD datasets, e.g. MSRA and so on, as your validation set. Then select the best model based on the validation metrics such as F-measure, MAE, etc. We also gave our suggested iteration numbers (around 400K on DUTS-TR dataset) in the paper. All the best.
shantzhou
hello, I couldn't find the val part in u2net_train, how to know the quantity of the model without val。