Ross Wightman

@xwang233 it's definitely related to specific pytorch and cuda / cudnn releases, I have to stick to ones that work to get through without issues, recently I found that the...

@michaelklachko That's my understanding of how it'd work (with a simplified model like that). You'd then always be using `clean=True` in your example for validation and production. As you say...

I did a first pass implementing option 2. It was minimal code, seems to be working, measured at somewhere between 15-20% throughput hit on my older Pascal (2 x 1080ti)...

@michaelklachko SplitBatchNorm is in master now. I've done limited experiments with it at this stage. https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/split_batchnorm.py

@sab148 The only part of the AdvProp I have so far is the SplitBatchNorm. So yes, you must generate adversarial examples with a CNN, possibly using a variety of attacks,...

@amaarora likely some missing details (as there always are), but that is the gist of it yes. Similar to the AugMix training in terms of batch construction and possible splitbn...

@yorklyb it's a bug, last feature info needs to include up to this line https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/ghostnet.py#L174 I need to check more closely to see what the current state and fix it,...

Thanks for the contribution. I agree that config files can make more sense than the sprawl of args that the scrips here have developed. I had a different (eventual) plan...

Also, the defaults that were placed in the config files in this PR are different and break the standard ImageNet training hparam starting point. I'll keep this PR open as...

I don't have resources to dedicate to train and evaluate new hybrid models, creating them with timm is quite straightfoward with existing building blocks. If anyone trains worthwhile hybrids with...