Joshua Z. Zhang
Joshua Z. Zhang
Checkout the demo here, https://github.com/zhreshold/mxnet-yolo/blob/master/video_demo.py Should be quite similar
The name is not necessary the operator name, print the layers to see the real name first
Check out random generator seed on wiki, you will get a better idea how seed control the randomness.
Fresh training? python train.py --num-class 2 --class-names 'a,b' should work. Let me know if you have other issues.
#124 Please see this, and try if it solves the problem. If so, you guys are loading the pretrained model.
batchsize=32, lr-steps=80,160,240, increase end-epoch, better train on single gpu. YMMV
I don't have the logs now, what mAP did you guys get, have you tried some other network like resnet or inception?
I guess something is broken. Will update later.
Recently people have successfully trained a densenet based ssd from scratch, which indicates that with properly designed backpropagate routes, we can skip the pretraining step. The paper is called DSOD:...
@yanhn CNN forward time is not affected by convergence. The time difference resides in NMS operation. Good model returns a lot of background regions which have very low scores, and...