IncompatibleKeys(missing_keys=[], unexpected_keys=...)

[vietvo89]

Hi Raff

I followed your instructions to test your pre-trained model as below.

from MalConvGCT_nocat import MalConvGCT
mlgct = MalConvGCT(channels=256, window_size=256, stride=64,)
x = torch.load("malconvGCT_nocat.checkpoint")
mlgct.load_state_dict(x['model_state_dict'], strict=False)

I got a bug:
_IncompatibleKeys(missing_keys=[], unexpected_keys=['convs_1.0.weight', 'convs_1.0.bias', 'convs_atn.0.weight', 'convs_atn.0.bias'])

It seems there are something wrong and mismatch.

Thanks

[vietvo89]

Hi Raff

I have resolved this bug. Here is my solution. It seems that if you miss model.eval(), it will get wrong. :)) If you think it is right, please update in order that other guys know how to run it and save their time.

device = torch.device('cuda:0' if torch.cuda.is_available() else "cpu")
mlgct = MalConvGCT(channels=256, window_size=256, stride=64,)
mlgct = mlgct.to(device)
if device == 'cuda':
    mlgct = torch.nn.DataParallel(mlgct, device_ids=0)
x = torch.load("malconvGCT_nocat.checkpoint")
mlgct.load_state_dict(x['model_state_dict'], strict=False)
mlgct.eval()

[RaffEdwardBAH]

What version of PyTorch are you using, and can you try again with 1.4.0? I think the save/loading did change a bit around 1.6/1.7, and this work was done before that.

[vietvo89]

I am using 1.3.1. Yeah, mismatched versions may cause this issue.

Thanks