HI, Mr. Chen, I downloaded the latest version of your implementation, however, I encountered some vital issues when training the network.
The software fails to generate predicted boxes, and the problem seems to raise from the predict() function in faster_rcnn.py. Specifically in: prob = (F.softmax(at.totensor(roi_score), dim=1)). By the way, the roi_score here also seems to be abnormal.
the probability (prob) that is later sent into function _suppress() is too small, so that in function _suppress(), the boolean result fromkeep = nms(cls_bbox_l, prob_l,self.nms_thresh) is thus always False, since no predicted anchor has a score greater than 0.7 (i.e. ThreshHold).
At first, I thought this results from the fact that the network is good enough to predict good boxes, however, I found that the Test mAP is unexpectedly low compares to your old version. The mAP for the first 4 epoch is 0.0017, 0.012, 0.0044, 0.0045.
Could you please give me some instructions on how to figure this problem? Thanks.
@chenyuntc
HI, Mr. Chen, I downloaded the latest version of your implementation, however, I encountered some vital issues when training the network.
The software fails to generate predicted boxes, and the problem seems to raise from the
predict()function infaster_rcnn.py. Specifically in:prob = (F.softmax(at.totensor(roi_score), dim=1)). By the way, the roi_score here also seems to be abnormal.the
probability (prob)that is later sent into function_suppress()is too small, so that in function_suppress(), the boolean result fromkeep = nms(cls_bbox_l, prob_l,self.nms_thresh)is thus alwaysFalse, since no predicted anchor has a score greater than 0.7 (i.e. ThreshHold).At first, I thought this results from the fact that the network is good enough to predict good boxes, however, I found that the Test mAP is unexpectedly low compares to your old version. The mAP for the first 4 epoch is 0.0017, 0.012, 0.0044, 0.0045.
Could you please give me some instructions on how to figure this problem? Thanks.