I used this web(https://www.jianshu.com/p/a672f702e596) to train faster rcnn(I input command line" python ./tools/train_net.py --gpu 0 " on terminal). It has a problem below the picture. How to solve it?
It seems the dimension doesn’t align in the ROI layer.
A common issue is that you have a different class number compared to the pre-train model.
You can remove the final fully connected layer (fc8?) and retrain it from a random initial directly.