If I try to load pytorch 0.3 or 0.2 version in pytorch 0.4, batchnorm gives the track running status error.
Are torchvision pretrained models trained everytime when new release occurs or do the transfer weights ‘in some way’ across versions?
If I try to load pytorch 0.3 or 0.2 version in pytorch 0.4, batchnorm gives the track running status error.
Are torchvision pretrained models trained everytime when new release occurs or do the transfer weights ‘in some way’ across versions?
This is relatively rare and most likely considered a bug. In your case PyTorch master has a bugfix to help BatchNorm load older snapshots, don’t tell anyone, but you might try just copypasting from that.
Best regards
Thomas
I’m also struggling with models trained on older vesions (0.4). Wish there was a solution for that.