size mismatch when fine-tuning model on KITTI dataset

[ToABetterDay]

Hi, thanks for your sharing. When I tried to fine tune the model on KITTI dataset using command: python -m maploc.train experiment.name=OrienterNet_MGL_kitti data=kitti experiment.gpus=1 data.loading.train.batch_size=2 training.finetune_from_checkpoint='"experiments/OrienterNet_MGL_reproduce/orienternet_mgl.ckpt"'

It generates such error: size mismatch for model.map_encoder.encoder.adaptation.0.0.weight: copying a param with shape torch.Size([9, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([8, 64, 1, 1]). size mismatch for model.map_encoder.encoder.adaptation.0.0.bias: copying a param with shape torch.Size([9]) from checkpoint, the shape in current model is torch.Size([8]).

I want to ask what size should be correct, I tried change configuration in orienternet.yaml but it doesn`t work.

[BPsoda]

I guess this is caused by a configuration conflict between the experiment OrienterNet_MGL_reproduce and checkpoint orienternet_mgl.ckpt. It seems that the orienternet_mgl.ckpt is the downloaded official checkpoint. In its configuration, the unary_prior is set to true, so the map_encoder also produces position prior as the 9th channel. However, your configuration (the default configuration) set unary_prior to false, so the map encoder only outputs 8 channels. Result in size mismatch. To solve this problem, I guess you could set "unary_prior": true in orienternet.yaml. But you will probably meet other conflicts. I fine-tuned on a model trained from scratch by myself.

[ToABetterDay]

Thank you! After set unary_prior to true, there is no size mismatch and the fine tune works.