yolov7
yolov7 copied to clipboard
WongKinYiu
training is so slow
When I train on my own data set, the speed is significantly slower than yolov5. I have 8 a40. Is it because of ota loss? #136
Mainly because of paste_in I think, currently we implement paste_in by a for loop, it very slow when CPU is not powerful.
When is the training code of pose open source? There is a problem with yolov5pose's code. I rewrote it myself, On my dataset, the position of the key points is normal, but the recall has reached the bottleneck. @WongKinYiu
I still have a question, what is the difference between OTA and simOTA, is that one of the negative samples takes all the negative samples, and the other is the rest of k? @WongKinYiu
For keypoint detection: merge this into this and modify this for training. You could update your code on it, I just add YOLOR support.
SimOTA uses top K, OTA uses optimization. I alos use this, but it has some problems. OTA also use dynamic_k_estimation, I think the difference of them is negative sample loss?
keypoints detection：When I use oksloss, several fixed points will be offset when predicting. Have you encountered this problem? @WongKinYiu
姿势的训练代码什么时候开源？yolov5pose的代码有问题。我自己重写了，在我的数据集上，关键点的位置是正常的，但是recall已经到了瓶颈。@WongKinYiu
请问你是如何修改的edgeai-yolov5， 训练数据关键点位置不正确的问题
姿势的训练代码什么时候开源？yolov5pose的代码有问题。我自己重写了，在我的数据集上，关键点的位置是正常的，但是recall已经到了瓶颈。@WongKinYiu
请问你是如何修改的edgeai-yolov5， 训练数据关键点位置不正确的问题
你可以参考;https://github.com/elyha7/yoloface