I’m recently doing research on the topic of semantic segmentation.
I’ve found out I have accuracy drop by updating from MXNet 1.1 to 1.2. (Of course training with same code and appropriate learning rate)
I’ve checked the release note of MX1.2, but still have no idea which will causing this issue.
I’ve also check the checkpoint restore mechanism. It seems normal.
The training code use Xavier initializer, deconvolution for bilinear upsampling, and SGD optimizer.
Can someone give me a hint or direction on how to investigate on this issue?