While loading the pre-trained FCOS-3D weights for training, the code complains about incompatible keys.
Upon analyzing it, most of the "incompatible" keys are as expected (eg, tpv_aggregator, tpv_head related keys missing in the loaded model and bbox_head keys missing in the defined model).
However, one set of items still did not make sense to me. For the image_backbone (and possibly for image_neck) batch norm layers, the running_mean, running_var, and num_batches_tracked keys are present in the checkpoint, but not in my model definition. I am not sure if this is intended -- a quick Google search says that these parameters are critical for training and inference, even if the BN parameters are frozen.
While loading the pre-trained FCOS-3D weights for training, the code complains about incompatible keys.
Upon analyzing it, most of the "incompatible" keys are as expected (eg,
tpv_aggregator,tpv_headrelated keys missing in the loaded model andbbox_headkeys missing in the defined model).However, one set of items still did not make sense to me. For the
image_backbone(and possibly forimage_neck) batch norm layers, therunning_mean,running_var, andnum_batches_trackedkeys are present in the checkpoint, but not in my model definition. I am not sure if this is intended -- a quick Google search says that these parameters are critical for training and inference, even if the BN parameters are frozen.Could you shed some light on this?