Training curve for RTS

[chenshihfang]

Hi, thanks for your RTS work.

I found that the loss for initial filter iteration occasionally overfitting when training. (~110 epochs) Is this usual? Due to the big GPU memory requirement for RTS, so I change to batch size to 5 for training, but I think reducing the batch may not lead to network overfitting.

BTW, is it OK to reduce memory Instance Memory / Segmentation Memory sizes for reducing GPU memory when training? I guess the high GPU memory occupied may be because of this memory mechanism.