The validation_set is put into the training list.

[Xun-Yang]

I found that you put both the training set and validation set into the training list, which leads to a higher performance on the two datasets.

[chenjoya]

In the original repo, it also does this. We test on the test set not on the val set. What is the problem??

[Xun-Yang]

Thanks for the reply.

The problem is: In the original repo, its val set is just used to test the model (using model.eval(), the model parameters are not updated.) Please check the on_update() in the train.py (line 182) in the original repo.

In your repo, you directly use the val_set for training.

[chenjoya]

Thank you for pointing out this. We will check it now. The result will be also updated.

[tujun233]

Dear @ChenJoya , I wanna make sure whether the result and code have been updated? I saw cfg.DATASETS.TRAIN still is ["tacos_train", "tacos_val"], but in /tan/data/datasets/init.py line 36, dataset = datasets[0], does it means that the code only use tacos_train? Thank you!