tensorflow-yolov3 icon indicating copy to clipboard operation
tensorflow-yolov3 copied to clipboard

Published 20 hours ago • verified publisher icon YunYang1994

predicting zero objects always

Open ask142857 opened this issue 6 years ago • 5 comments

I am clearly observing decrease in train loss and test loss while training but when I'm running evaluate.py it just prints ground truth , i have put some print statements and realized that for each image the predict function in evaluate is giving list of zero size implying it has detected zero objects. Help me fix this issue

ask142857 avatar Jun 28 '19 11:06 ask142857

hello, anybody here ? Can anybody fix this issue on the basis of such descriptions ?

YunYang1994 avatar Jun 28 '19 12:06 YunYang1994

I have run the model on a random rbg image to check whether it's giving any output from the evaluate.py . It has given the same empty array as output from the predict function . Is everything alright with evaluate.py?

ask142857 avatar Jun 28 '19 12:06 ask142857

me too

lway-hb avatar Jul 02 '19 07:07 lway-hb

Me too. I am using my own dataset which has 2 classes. Still predice zero objects.

I did not modify any value in config.py except __C.TRAIN.ANNOT_PATH and __C.YOLO.CLASSES. I used yolov3_coco_demo as my initial weight and the loss on the beginning was 2716. Is the loss normal?

dvlee1024 avatar Jul 08 '19 07:07 dvlee1024

I got this problem, then change self.trainable: True in evaluate.py. In my case, it's solved

kelvinliu04 avatar Aug 27 '20 07:08 kelvinliu04