This project has two purposes:
Suppose you have installed gbdtmo. If not, please refer GBDT-MO.
Find the path of gbdtmo.so and modify Lib_path in cfg.py.
Lib_path = "path to gbdtmo.so"Import gbdtmo and load the shared library file
from gbdtmo import load_lib, GBDTSingle, GBDTMulti
import cfg
LIB = load_lib(cfg.Lib_path)Build a gbdtmo instance. You must setup the output dimension.
out_dim = 10
params = {"max_depth": 5, "lr": 0.1}
booster = GBDTMulti(LIB, out_dim=out_dim, params=params)Setup your dataset, train and predict. Data in the first tuple is used for training. Data in the second tuple is used for validation which can be omitted. Items in tuples must be a numpy array.
booster.set_data((x_train, y_train), (x_valid, y_valid))
booster.train(num_rounds)
preds = booster.predict(x_valid)For more information, refer the Python scripts or our documentation.
Get the performance of non-sparse gbdtmo for a specific dataset via
python run_peformance.py dataset gbdtmo
Get the running time of non-sparse gbdtmo and gbdtso for each round via
python run_time.py dataset
Get the performance of sparse gbdtmo for a specific dataset via
python run_sparse.py dataset -time 0
Get the running time of sparse gbdtmo for each rounds via
python run_sparse.py dataset -time 1
See help of those scripts for more details. Results will be recorded in log/. Please refer test.py to see how to parse them. We provide datasets mnist, mnist_reg, yeast and Caltech101. For nus-wide, you should download it into dataset/ from here and run loader.py to pre-process it.