Getting -1 for map using VOC07+12 Trainval about fpn HOT 6 OPEN

I encounted this issue too, I checked the weights I trained, there are lots of nan in the scope of res3_5, Top-Down, RPN, Fast-RCNN, is it just a hyperparameter problem?
I also found that I didn't restore all weights in Resnet50.npy, especially batch-normalization layer's weights 'mean, variance', should them be restored to graph variables beta, gamma?

from fpn.

The same problem here, which version of tensorflow you are using?

from fpn.

there might some difference between these models
ie, maybe caffe pretrain model requires input normalization. while the origin one doesn't

from fpn.

I used the tensorflow 1.0.1,does this version lead to the problem? @imyhxy

from fpn.

@Alina-liu I am not sure, but I used the tensorflow 1.8.

from fpn.

The same problem here, has it solved? thx @Alina-liu @imyhxy

from fpn.