With a RTX 2080ti, the performance is much worse. Getting around 1 it/s while with lstein/stable-diffusion I get ~10 it/s.
Could be that there is a bunch of low hanging fruits for optimization though, and it can reach higher performance, the amount of people who spent time doing any sort of optimizations are obviously much lower for this TensorFlow port (1 person) vs the pytorch implementation (too many to count manually).
Edit: it's unclear to me if it's actually running on the GPU at all.
Nice effort, though it'll be better to have a requirements.txt, or even better, a proper Anaconda envionment.yaml file or other virtual env to specify dependencies and their versions.
It's great lots of people are making their code available, but this is really something that always seems to come up for most public python ML repos. It just makes adoption and usage so much harder and frustrating for users.
Separately, how does "converting weights" from eg: Pytorch to Keras work? Are there public converters, is some custom code involved, etc?
3 comments
[ 1.5 ms ] story [ 20.5 ms ] threadCould be that there is a bunch of low hanging fruits for optimization though, and it can reach higher performance, the amount of people who spent time doing any sort of optimizations are obviously much lower for this TensorFlow port (1 person) vs the pytorch implementation (too many to count manually).
Edit: it's unclear to me if it's actually running on the GPU at all.
It's great lots of people are making their code available, but this is really something that always seems to come up for most public python ML repos. It just makes adoption and usage so much harder and frustrating for users.
Separately, how does "converting weights" from eg: Pytorch to Keras work? Are there public converters, is some custom code involved, etc?