[P] Oddly thresholded confidence scores on scaled yolov4 csp

All object detections on the scaled yolov4 csp model have a confidence below 0.5, while it should range from 0 to 1. Does anything come to mind as to what the problem might be?

Info:

* I'm using a branch of the [author's PyTorch repo](https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp)
* Predictions are otherwise pretty good in terms of bbox placement
* I'm training on a single gpu
* Darknet coco weights are converted to ".pt" PyTorch weights for training
* A custom dataset is …

confidence machinelearning