When trying to create a pre-trained layer with xresnet34 and xresnet18 I get the error:
copying a param with shape torch.Size([64, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([64, 64, 3, 3]).
Currently we only have available the weights for the 1 channel version?
The model loads correctly for resnet50/101
xresnet only has pretrained xresnet50 at the moment I believe
def _xresnet(pretrained, expansion, layers, **kwargs):
# TODO pretrain all sizes. Currently will fail with non-xrn50
So we have an XResNet50 pretrained now, but it’s only available in v2 correct?
Yes. But I think you could copy the weights over since it’s just pytorch weights?
Awesome, that’s a great idea and should work. I’ll test it tomorrow - thanks for the idea!