Error(s) in loading state_dict for CMAPolicy:

[volkancirik]

Hello,

I'm downloaded the pretrained model and try to run the test_set_inference.yaml. I receive this error.

Unexpected`` key(s) in state_dict: "critic.fc.weight", "critic.fc.bias"

Is there a mismatch between the code and the pretrained model? If so, could you update the model link? Or do you have other suggestions? Thanks in advance!

[jacobkrantz]

The pretrained models load correctly for me on the master branch. For the rxr-habitat-challenge branch, the baseline models inherit from a base policy class that does not have critic.* layers. These layers were never used during training/evaluation of the baselines and can be safely ignored:

https://github.com/jacobkrantz/VLN-CE/blob/52a478259ae41068eecb751a0cc49307384e460b/vlnce_baselines/common/base_il_trainer.py#L76

can be changed to

self.policy.load_state_dict(ckpt_dict["state_dict"], strict=False)

[rpartsey]

Hi @jacobkrantz! Thanks for opensourcing your code. I follow your instructions in README.md but still have the same error as @volkancirik described above.

@volkancirik have you figured out how to fix it?

UPD: I set self.policy.load_state_dict(ckpt_dict["state_dict"], strict=False) and then faced with the RuntimeError: 'lengths' argument should be a 1D CPU int64 tensor, but got 1D cuda:0. To fix the RuntimeError I followed advice here https://github.com/pytorch/pytorch/issues/43227 and converted lengths to cpu lengths.cpu() (in torch/nn/utils/rnn.py file).

After that, I finally evaluated CMA_PM_DA_Aug.pth checkpoint without errors

run python run.py --exp-config vlnce_baselines/config/paper_configs/cma_pm_da_aug_tune.yaml --run-type eval

and got the following results

Episodes evaluated: 1839
Average episode distance_to_goal: 7.610544
Average episode success: 0.287113
Average episode spl: 0.265512
Average episode ndtw: 0.496594
Average episode path_length: 8.303312
Average episode oracle_success: 0.356172
Average episode steps_taken: 88.323545

[cshizhe]

Do you have an idea why the performance is much lower than the result listed here?