search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.75k stars 7.96k forks source link

Doubt about resnet cfg #3870

Open linghu8812 opened 5 years ago

linghu8812 commented 5 years ago

Hello!

I want to use resnet as backbone of YOLOv3. But I have found that there is a shortcut between layer 5 and layer1, layer17 and layer13, and so on. The size of layer 5 is 104x104x256, while the size of layer 1 is 104x104x64, and the size of layer 17 is 52x52x512, the size of layer 13 is 104x104x256. After the layer 17, it print a stderr. image refer to https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py, I doubt there might be something wrong with the resnet50.cfg, layer1 cannot have a shortcut with layer5.

AlexeyAB commented 5 years ago

@linghu8812 Hi, Yes, probably there are mistakes.

If somebody have a time to fix it and retrain on ImageNet for about 1M iteration, it would be nice.