1.When I train the model(anyone) according to README , I found it's very difficult for the model with data augmentation to converge, and the model without data augmentation converged well.
From tensorboard visualization, it seems that the model can only learn background information with data augmentation, and the LOSS is about 40.
I am confused is there anything I should pay attention to and can anyone show me the convergence curve?

2.Using the checkpoints offered by the author, the result on FlyingChair is good.
But when testing the model in SINTEL dataset, I found the error is extremely large which is about 14, and can’t reach the paper’s result which is about 4.
Specifically, the resolution of SINTEL image is 1024x436, so I resized it to 1024x448 to adapt the model. I’m not sure whether it’s a problem.
Anyone have any better solution?
Thanks a lot!