I am trying to train my Custom Object in Yolo Darknet with class ->2 ,filter(21),collected around 5000 images of size (1106x620), set Learning rate .01 batch is 64, ans sub division is 16.
Observation- after 500 - 3000 of iteration , i can see average loss is in between 7.2 to 7.4
Please let me know how should i improve my average loss ?
Loss might be stuck in between but It might not be accuracy criteria if you are doing object detection.
try to calculate Mean Average Precision(MAp) using the command (if you are using original YoLo repository)
./darknet detector map data/obj.data yolo-obj.cfg backup\yolo-obj_7000.weights
Try decreasing the learning rate to 0.001 or 0.0001. Also, check the pretrained weights you are using, if there are suitable for your custom use case.
Try decreasing learning rate and batch size. Check if labels class, x, y, width, height are correct. If this does not work there might be a problem about your labels and/or data.
