vishalmehta1991

vishalmehta1991

vishalmehta1991

vishalmehta1991

vishalmehta1991

(A) Initialize a bit mask indicating which samples are contained in each node
(B) Initialize a “node map” indicating which nodes are present at each level
(C) ForEach(tree_level)
    1. Find the node id of all data samples, using the bit mask
    2. Compute the possible splits for all bins, all columns and all nodes
    3. Find the best split for each node
    4. Update the bit mask and sparse node map to feed the next level

vishalmehta1991

|  ######cuML######                 |  ######Sklearn######                 |
|                                   |                                      |
|  from cuml import                 |  from sklearn.ensemble import        |
|  RandomForestClassifier as cuRF   |  RandomForestClassifier as sklRF     |
|                                   |  import multiprocessing as mp        |
|                                   |                                      |
|  # cuml Random Forest params      |  #sklearn Random Forest params       |
|  cu_rf_params = {                 |  skl_rf_params = {                   |
|     ‘n_estimators’: 25,           |      ‘n_estimators’: 25,             |

vishalmehta1991

from cuml.dask.ensemble import RandomForestClassifier as cuRF_mg
# cuml Random Forest params
cu_rf_params = {
    ‘n_estimators’: 25,
    ‘max_depth’: 13,
    ‘n_bins’: 15,
    ‘n_streams’: 8
}

# Start by setting up the CUDA cluster on the local host