Ross Wightman

@byronyi as per @tmabraham's comment, I've been chugging along with some updated code that works with PyTorch XLA, it's on a diffeferent branch of this repository https://github.com/rwightman/pytorch-image-models/tree/bits_and_tpu/timm/bits#readme I've been juggling...

@ngfuong there's no interval to skip some subset of epochs for eval & save ... but you can use `--checkpoint-history 20` to say keep the last 20 epochs of eval...

@xonobo I looked at this briefly, the mobiledet CPU, GPU and TPU backbone could be added to the efficientnet generator w/ an extra block type (Tucker) and some tweaks, but...

@nahidalam I was not planning to add that one as it doesn't look very compelling compared to many recen models in that size range, any particular reason?

@SleepyTT curious what cloud filesystem / setup you use that doesn't support hard links? I'm not aware of many modern setups, including file share protocols that don't support them these...

Mentioned in #607, yes, plan is to add feature extraction but in a way that's generic for all non-CNN archs (so the various vision transformers and the new MLP-Mixer nets)....

@bryant1410 given it's not the cleanest workaround, and actually could have some crappy performance impact on large datasets (iterating through the whole dataset to reach the end), what are the...

@jheek ah, k... so in thise case yeah, the empty dict acts as a sufficient 'not-initialized' truthy value that the rest of my lazy logic should still work.

@jheek I thought this would be a quick and easy fix but ended up going down a rabbit hole. The idea doesn't work. You cannot restore a FrozenDict with state...

@jheek I created an EmaState dataclass and got that working in a less hacky fashion. Still have a bit of an issue, allowing training to start with ema active and...