cuda.ml
cuda.ml copied to clipboard
mlverse
R interface for cuML
Is there a way to load a treelite serialized model directly (instead of a xgboost / lightgbm model). It seems possible in the Python interface via `load_from_treelite_model` (https://docs.rapids.ai/api/cuml/stable/api.html#cuml.ForestInference.load_from_treelite_model). My use...
I was trying to install this package with `install.packages("cuda.ml")` or `devtools::install_github("mlverse/cuda.ml")` but I got an error when building the package, ``` /usr/bin/ld: cannot find /lib64/libpthread.so.0 /usr/bin/ld: cannot find /usr/lib64/libpthread_nonshared.a collect2:...
I would expect the cuda toolkit version in the conda step to match the cuda toolkit version in the install step -- lines 316 and 288 should match.
While integration with `{parsnip}` is done, I still need to verify everything works well with other tidymodels packages such as `{tune}` for hyperparameter tunning. It is possible that everything will...
At the momet there are some copy of data when passing a trained cuML models back to the R process from the GPU. Some of these copies may be reasonably...
RAPIDS cuML, similar to scikit-learn, by default uses 5-fold cross validation to calibrate class probabilities: see https://github.com/rapidsai/cuml/blob/57a6ff7ecae6a4b6e2e9db97dc664be453310111/python/cuml/svm/svc.pyx#L394-L396
i.e., the following: - Working with a pre-existing version of libcuML - Working with some pre-built version of libcuML downloaded automatically and bundled together with the rest of {cuda.ml}
e.g., `hardhat::validate_predictors_are_numeric(processed$predictors)`, among others