yolov5-face
yolov5-face copied to clipboard
depth and width multiple
Hi, I was interested in your stem block, so I added the stem block module in the ultralytics yolov5 training pipeline. All went fine when depth multiple = 0.33 and width multiple 0.5 with SGD optimizer with lr = 0.02, the default hyperparameter setting for yolov5. It works great, and small objects mAP went up.
But when I increased the depth and width multiple to 0.67 and 0.75, then model doesn't seem to be learning. mAP decreases in each epoch. So I have tried changing the optimizer to ADAM instead of SGD and tried bunch of different learning rates, but none of them seems to work.
Are there any reasons for this?