Closed hadipash closed 4 months ago
This PR introduces the extraction of the training environment initialization, as it is utilized across the project. I have also added a graph caching feature, which is exceptionally helpful during the hyperparameters tuning stage. Here are the numbers (tested on SDv2.1 and pokemon-blip dataset): | Device | Environment | First epoch time without caching | With caching |
---|---|---|---|---|
910 | MS2.1 | 15+ minutes | 4~ minutes | |
910* | MS2.2.1 | 17+ minutes | 5+ minutes |
@HaoyangLee the graph cache is fixed now?
@HaoyangLee the graph cache is fixed now?
Haven't used that feature for a while. Not sure about it in the latest MS version. @hadipash 's testing result looks nice.