'nan' detection result from my own faster rcnn model

[guoqiang0148666]

Thank you for your excellent work very much !

a few month ago, l proposed a issue about my own faster rcnn model in failing to inference, l reflected the problem to your open ai lab QQ team (the response is that the conv layer after roi_pooling layer fail to get the previous layer's data) and found it be fixed in new commit (0e3c30 'fix tflite bugs' Fri Sep 11:33:45 2019), my own model run ok.

however, l found a new question that the detection result is nan (bbox_delte_data, score are nan, but roi_data is ok), l suspected that it may be caused by my proposed previous question, l am looking forward to your reply, any suggestion will be welcome.