v1.1.5
Bugs fixed in v1.1.5
- Issue with large models and CUDA, previous version loaded all models at once using CUDA memory very quickly. FIX: Models are initialized and stored on CPU after creation, moved back to CUDA when training/inferencing. This will have impact on inferencing performance since models are being moved back and forward from CPU<->GPU but this only affects very large models.
Features added in v1.1.5
- Easier installation with conda-forge packaging, can now simply install with conda install vprtempo -c conda-forge
- Added options to plot metrics, similarity matrices, and additionally output precision and recall in json file
- Modified model names so that they're clearer, also rather than model names simply being network architecture dims more information such as the datasets themselves included so multiple models with the same architecture can be trained across different datsets without overwriting them.