Thank you for sharing the excellent work! I have tried to implement PASCAL VOC results with SEAM pseudo-mask. I have trouble in training. When I am training the code for about 8 iterations, the loss turns to NaN. Is there any issue or modification in a code?