Question about SYNTHIA checkpoint/training

[greeneggsandyaml]

Hello authors,

Thank you for your paper! I am running into issues while trying to train a model on the SYNTHIA dataset. I am using the provided checkpoint pretrained on SYNTHIA. When I try to use --num_classes 16 for training, I get an error that the pretrained model has the wrong number of classes:

size mismatch for layer5.conv2d_list.0.weight: copying a param with shape torch.Size([19, 1024, 3, 3]) from checkpoint, the shape in current model is torch.Size([16, 1024, 3, 3]).
... {other size mismatches} ...
size mismatch for layer6.conv2d_list.3.bias: copying a param with shape torch.Size([19]) from checkpoint, the shape in current model is torch.Size([16]).

It seems I am doing something wrong -- what training flags do you use for SYNTHIA training?

Thanks!

[greeneggsandyaml]

Hi, I just wanted to follow up on this.

[xyyue]

Got the same error. Why this is happening? Can the authors comment more on this?

[ljjcoder]

I also got the same error? Did anyone fix it ?