Open offchan42 opened 6 years ago
I trained yolov2.cfg on my own dataset for object detection and I got some OK predictions(about 50% accuracy). The avg loss stopped decreasing at around 2. I trained yolov2-voc.cfg on the Pascal VOC dataset with the initialization of pre-trained weights(yolov2-voc.weights) and the avg loss stopped decreasing around 3. I didn't look at the code but I think you are right that loss calculation of darkflow is different from darknet.
I got 0.3 loss using darknet
and the mAP was about 50% to 60%. So it was quite good for me.
Maybe the loss is dataset-specific. I don't know.
hi off99555, you can give us more details pls . how many images learning rate epochs batch subdivision.... thx
I got 4,000 images, everything else is set to default. I had done the training job a long time ago.
I trained yolov2.cfg on my own dataset for object detection and I got some OK predictions(about 50% accuracy). The avg loss stopped decreasing at around 2. I trained yolov2-voc.cfg on the Pascal VOC dataset with the initialization of pre-trained weights(yolov2-voc.weights) and the avg loss stopped decreasing around 3. I didn't look at the code but I think you are right that loss calculation of darkflow is different from darknet.
yes, me too. my loss calculated around 1.6 to 2.5 but the result show quite decent performance.
What is the ideal loss value for
darkflow
during training? Because there is no option to evaluate loss on a validation set, knowing this value should be useful.Is it similar to this repository? https://github.com/AlexeyAB/darknet It stated that we should stop training when the average loss is around
0.060370
. But I'm skeptical because I suspect that the loss calculation mechanism might be different.