Size mismatch when importing pretrained vqa model

❓ Questions and Help

Hello,

I have encountered a size mismatch problem when loading the state dictionary of the pretrained vqa model.

In particular, when running python -u habitat_baselines/run.py --exp-config habitat_baselines/config/eqa/il_vqa.yaml --run-type eval

I receive the following error RuntimeError: Error(s) in loading state_dict for VqaLstmCnnAttentionModel: size mismatch for cnn.encoder_seg.weight: copying a param with shape torch.Size([40, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([41, 512, 1, 1]). size mismatch for cnn.encoder_seg.bias: copying a param with shape torch.Size([40]) from checkpoint, the shape in current model is torch.Size([41]). size mismatch for cnn.score_pool2_seg.weight: copying a param with shape torch.Size([40, 16, 1, 1]) from checkpoint, the shape in current model is torch.Size([41, 16, 1, 1]). size mismatch for cnn.score_pool2_seg.bias: copying a param with shape torch.Size([40]) from checkpoint, the shape in current model is torch.Size([41]). size mismatch for cnn.score_pool3_seg.weight: copying a param with shape torch.Size([40, 32, 1, 1]) from checkpoint, the shape in current model is torch.Size([41, 32, 1, 1]). size mismatch for cnn.score_pool3_seg.bias: copying a param with shape torch.Size([40]) from checkpoint, the shape in current model is torch.Size([41]).

Could it be that the problem is in the pretrained vqa model? The size of these cnn-related tensors does not match when compared to the ones specified in il models (41 vs 40)…

[mukulkhanna]

Thanks @nilinykh for pointing this out. It seems the VQA checkpoint provided in the README is an outdated one. I’ll share the latest one here in a few hours, and also open a new PR to update the README.

[mukulkhanna]

Hi @rpartsey, I just created a PR for the same (#942). Thanks for pointing this out!

We can close this issue as soon as it’s merged.