Comment on Cifar10 Classification and Pretrained Models

[Kyle Sargent]

Hello! I have really enjoyed reading your paper and want to do some experiments with the invertible network on Cifar-10. However, I think there are some small mistakes with the CIFAR-10 config file, namely that it refers to MNIST and has the wrong image size (28 instead of 32). After rectifying this, I get another error, namely

File "/home/kyle/mintnet/models/cnn_classification.py", line 122, in forward
    center2 = center2.permute(0, 2, 1, 3, 4, 5).contiguous().view_as(self.weight2)
RuntimeError: shape '[16, 16, 1, 1]' is invalid for input of size 2304

I'm guessing some other parameter in this config file is not correct, but I couldn't determine which one from the paper. Would you be able to help me figure it out?

Additionally, would you happen to have checkpoints for the trained CIFAR10 model lying around? At least for now, I don't actually need to train a model myself, I just want to use a trained one.

Thanks in advance!

[Ali Sadeghian]

We also have the same issue, it would be great if you publish the configs that reproduce the paper's results for CIFAR?

[Yang Song]

The config files have been fixed. I have tested it again on my computer and it should work fine now. I have also provided a link to pre-trained checkpoints in README.

Sorry for the late reply, not sure why I was not notified by issues...