0.17.3 SSD Test net output mAP is 0

[cvhuang]

Hi, Is anyone have used 0.17.3 training ssd model. I trained a ssd net in 0.17.2, it works and mAP is 0.85. I copied the *prototxt, python and dataset files to 0.17.3 and retrained it, but the Test net output mAP is 4.53444e-05(almost 0). The train mbox_loss is around 3.5.
Any hints?
ps: I used model parameter file which was trained in 0.17.2 to inference in 0.17.3. They have same result.