@AlexeyAB
Hi,
I am trying to perform object detection on VisDrone Dataset 2020. I have 12 classes and set my max_batches to 8000(I Know that's pretty low but owing to time contsraints and others) and filters in conv. layers to 51. I am using Google colab GPU. My issue is that the Average loss fluctuates between 300 to 500, say it is 312 for an iteration then it goes up like 350, then 380 and upwards suddenly reducing again to 340 and 310 and then shoots up and my mAP@0.5 is always at 0.35% there is no observed increase or decrease. So I stopped it mid-way, so can you suggest any way to reduce the average loss and increase mAP@0.5. I'm thinking of adjusting the anchor box dimensions but I'm not sure how to go about it. Any other ideas are also welcome.
I am attaching some screenshots to give some more context.
@AlexeyAB Hi, I am trying to perform object detection on VisDrone Dataset 2020. I have 12 classes and set my max_batches to 8000(I Know that's pretty low but owing to time contsraints and others) and filters in conv. layers to 51. I am using Google colab GPU. My issue is that the Average loss fluctuates between 300 to 500, say it is 312 for an iteration then it goes up like 350, then 380 and upwards suddenly reducing again to 340 and 310 and then shoots up and my mAP@0.5 is always at 0.35% there is no observed increase or decrease. So I stopped it mid-way, so can you suggest any way to reduce the average loss and increase mAP@0.5. I'm thinking of adjusting the anchor box dimensions but I'm not sure how to go about it. Any other ideas are also welcome. I am attaching some screenshots to give some more context.