Very high loss while training TAO yolov4

In short, for 4.0.0 or 4.0.1 version, to improve the mAP, we fix issues in the yolov4 structure, the loss function, and etc.

You can find some difference in the spec file, for example in 4.0.x version,
loss_loc_weight: 1
loss_neg_obj_weights: 1
loss_class_weights: 1

1 Like