Unexpected (for some) keys in load_state_dict

vision
1

I try to load a saved model like this:

state = torch.load(args.ckpt)
model = GeneratorSeg(args.size, args.latent, args.n_mlp, image_mode=args.image_mode,
                         channel_multiplier=args.channel_multiplier, seg_dim=args.seg_dim
                         ).to(device)
new_state_dict = OrderedDict()
    for k, v in checkpoint['g_ema'].
        print(k) # I can see there are "convs.12.conv.weight", "convs.12.conv.blur.kernel" keys
        new_state_dict[k] = v

  model.load_state_dict(new_state_dict)
  model.eval()

But I get this error:
Unexpected key(s) in state_dict: “convs.12.conv.weight”, “convs.12.conv.blur.kernel”,
It is a bit confusing, when I print the keys, I can see that there are such keys in checkpoints.
Does anybody have any idea?

2

The key mismatch errors are raised e.g. if you’ve changed the model architecture or generally the attributes names of the parameters or buffers after saving the state_dict.
In your case the “convs.12.conv.weight”, “convs.12.conv.blur.kernel” are unexpected in the state_dict and cannot be mapped to the current model. I.e. if you check the parameter names of the model you should not be able to find them using these keys:

for name, param in model.named_parameters():
    print(name)

and they might have been moved to another attribute name etc.