h2o4gpu
h2o4gpu copied to clipboard
h2oai
H2Oai GPU Edition
* Would be good if a user can specify early stopping with a particular metrics, i.e., `AUC`, `Logloss`, `RMSE`, etc.
E.g., predictions or clusters should be saved on GPU and default for _ptr methods is to only have those things on the GPU.
Include side by side examples of scikit and h2o4gpu. Can start of with comparing GLM in scikit vs GLM in h2o4gpu, XGBoost, and K-Means in scikit vs K-Means in h2o4gpu...
Single model GLM is pretty slow using POGS. Implement a new one in the XGBoost repository. While implementing look into options to distribute it (maybe using Rabbit?).
https://www.tensorflow.org/install/install_linux http://www.python36.com/how-to-install-tensorflow-gpu-with-cuda-9-2-for-python-on-ubuntu/ Power, e.g.: https://developer.ibm.com/code/howtos/install-tensorflow-on-power Point is power needs to have tf compiled, and currently tf 1.8-1.9 only pip installs for cuda9.0. If need to support other archs/installs, need to...
Ideally it should running also on GPU. But even without it naive implementation https://github.com/h2oai/h2o4gpu/blob/master/src/interface_py/h2o4gpu/solvers/factorization.py#L309-L311 can be optimized to use less memory.
Current state of C/C++ examples: - no documentation - not checked with CI Do we want to keep maintaining them?
* Current implementation only supports RMSE * In case of more sophisticated function eq (1) can be replaced with Taylor approximation https://arxiv.org/pdf/1808.03843.pdf * It's worth to implement POC to check...