Model trained in 0.4.1 on CUDA 9.0 is not working in 0.4.0 on same CUDA

1

I am working on a Siamese structured model, it was trained in PyTorch-0.4.1. I have implemented the inference in the same version of PyTorch, that works. But inferring on a system with PyTorch-0.4.0 gives the following issue:

Unexpected key(s) in state_dict: “bn1.num_batches_tracked”, “bn2.num_batches_tracked”, “bn3.num_batches_tracked”, “bn4.num_batches_tracked”, “bn5.num_batches_tracked”.

I don’t have a clue on what is wrong. Upgrading from 0.4.0 to 0.4.1 will solve, but can I solve without the upgrade.

Please help.

2

Loading checkpoints is not forward compatible. So please either upgrade, or for this particular case remove those keys as they don’t matter (only for this particular case).

1 Like
3

Facing similar issue, trying to load a model trained on 0.4.1 from a 0.4.0 environment.