GT encoder has poor traLoss and traTarLoss

[Maozhentan2024]

Hello,@xuebinqin thank you very much for your outstanding contributions and sharing on DIS. I encountered a problem while researching and training the GT encoder. The trained GT encoder has poor traLoss and traTarLoss, and the training was interrupted at epoch=10, generating: GTENCODER-gpu_itr_90000_traLoss_2.5166_traTarLoss_0.568_valLoss_0.3852_valTarLoss_0.0118_maxF1_0.9978_mae_0.0009_time_0.035759.pth. In contrast, the model trained without hypar["interm_sup"] = False has higher segmentation accuracy. I would like to ask what might be the reason for this?

[Maozhentan2024]

When training without the GT encoder supervision, the trained model is gpu_itr_47500_traLoss_0.7772_traTarLoss_0.1048_valLoss_0.395_valTarLoss_0.0206_maxF1_0.9952_mae_0.0022_time_0.067231.pth, which surprisingly performs better than the GT encoder. Here are the comparison images below.