Transfer Learning on YOLOv4-CSP

[srikhetramohanty]

What layers need to be freezed to perform transfer learning on YOLOv4-csp effectively? Currently I have frozen all the layers except the 3 conv layers before YOLO layers. ~31k trainable parameters on my custom dataset with ~400 training images and 1 class. The loss doesnt seem to reduce and the precision recall remains at 0 till ~200-300 epochs. But when i train from scratch on the same data, it achieves decent performance of ~0.65 mAP(0.5). Any way out? @WongKinYiu Thanks in advance.