rf · CLIO Core

What it does

Trees are fitted one after another with the decision tree engine. Because the bootstrap and the feature subsets are hashed rather than sampled from a stream, the forest is the same for any paging, block count or node count.

Rows

n_features feature columns followed by the class label. clio_ml.COLUMNS["rf"] is 1.

Python

python

import numpy as np, clio_ml as cm
rng = np.random.default_rng(1)
X = rng.normal(size=(4096, 7)).astype(np.float32)
y = (X[:, 0] + X[:, 1] > 0).astype(np.float32)
data = np.column_stack([X, y]).astype(np.float32)

r = cm.fit_rf(data, num_trees=4, tree={"max_depth": 3, "num_classes": 2})
print(len(r.model.trees))
print(cm.predict("rf", r, data[0, :7]))

fit_rf(X, **options). The nested tree options are passed as a dict under tree.
Options (RandomForestOptions): num_trees (8), seed (1), features_per_node (0 = ceil(sqrt(nf))), tree (a DecisionTreeOptions: num_classes 2, max_depth 4, num_bins 32, threads 256, row_base 0).
Result RandomForestResult: model (RandomForestModel with trees, a list of DecisionTreeModel) and stats.
predict("rf", r, x) returns the voted class.

C++

cpp

struct RandomForestOptions {
  u32 num_trees = 8;
  u32 seed = 1;
  u32 features_per_node = 0;  // 0 = ceil(sqrt(nf))
  DecisionTreeOptions tree;   // depth, bins, classes, threads, row_base, cluster
};
struct RandomForestModel {
  std::vector<DecisionTreeModel> trees;
  std::vector<float> Flatten() const;
};
struct RandomForestResult { RandomForestModel model; VecStats stats; };
RandomForestResult FitRandomForest(Dataset &ds, const RandomForestOptions &opt);
RandomForestResult FitRandomForestPlain(const float *data, u64 n_rows, u32 nf,
                                        u64 elems_per_page, u32 nblocks,
                                        const RandomForestOptions &opt);
RandomForestResult FitRandomForestReference(/* same arguments */);
u32 RandomForestPredict(const RandomForestModel &m, const float *x);

The ChiMod is module ml_rf.