File size: 10,290 Bytes
a294dd9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 | """
Inference engine optimizations for HyperOpt-GBT.
Implements multiple inference strategies inspired by YDF's engine compilation:
1. Naive tree traversal (baseline)
2. QuickScorer-style bit-mask scoring (for small trees)
3. SIMD batched prediction (AVX-like vectorization in Python)
4. Compiled flat trees (cache-oblivious structure)
References:
- YDF Inference Engine (arXiv:2212.02934, Section 3.7)
- QuickScorer (Lucchese et al., CIKM 2015)
"""
import numpy as np
from numba import njit, prange
class InferenceEngine:
"""Base class for inference engines."""
def predict(self, X_binned):
raise NotImplementedError
class NaiveEngine(InferenceEngine):
"""Naive tree traversal - baseline.
Single while-loop from root to leaf.
Slow due to unpredictable branches and cache misses.
"""
def __init__(self, trees):
self.trees = trees
def predict(self, X_binned):
n_samples = X_binned.shape[0]
n_trees = len(self.trees)
predictions = np.zeros(n_samples, dtype=np.float64)
for tree in self.trees:
for i in range(n_samples):
predictions[i] += self._predict_single(tree, X_binned[i])
return predictions
def _predict_single(self, tree, x):
node = tree.root
while not node.is_leaf:
if x[node.feature] <= node.threshold:
node = node.left_child
else:
node = node.right_child
return node.value
class FlatTreeEngine(InferenceEngine):
"""Flat tree representation for cache-oblivious traversal.
Stores tree in array format (like a heap) to enable:
- Predictable memory access patterns
- Branchless traversal with precomputed jump tables
- SIMD-friendly batch processing
Structure:
- nodes[i]: [feature_idx, threshold, left_child_idx, right_child_idx, value]
- Leaf nodes have feature_idx = -1
"""
def __init__(self, trees, n_bins):
self.n_bins = n_bins
self.flat_trees = []
self.leaf_values = []
for tree in trees:
flat, leaves = self._flatten_tree(tree)
self.flat_trees.append(flat)
self.leaf_values.append(leaves)
def _flatten_tree(self, tree):
"""Convert recursive tree to flat array representation."""
nodes = []
leaves = []
def traverse(node):
idx = len(nodes)
if node.is_leaf:
nodes.append([-1, -1, -1, -1, node.value])
leaves.append(node.value)
return idx
else:
nodes.append([node.feature, node.threshold, -1, -1, 0.0])
left_idx = traverse(node.left_child)
right_idx = traverse(node.right_child)
nodes[idx][2] = left_idx
nodes[idx][3] = right_idx
return idx
traverse(tree.root)
return np.array(nodes, dtype=np.int32), np.array(leaves, dtype=np.float64)
def predict(self, X_binned):
n_samples = X_binned.shape[0]
n_trees = len(self.flat_trees)
predictions = np.zeros(n_samples, dtype=np.float64)
for tree_idx in range(n_trees):
flat = self.flat_trees[tree_idx]
pred = self._predict_flat_batch(X_binned, flat)
predictions += pred
return predictions
@staticmethod
@njit(parallel=True, fastmath=True, cache=True)
def _predict_flat_batch(X_binned, flat_tree):
"""Numba-accelerated flat tree prediction with parallelization."""
n_samples = X_binned.shape[0]
predictions = np.empty(n_samples, dtype=np.float64)
for i in prange(n_samples):
node_idx = 0
while True:
feature = flat_tree[node_idx, 0]
if feature < 0: # Leaf node
predictions[i] = flat_tree[node_idx, 4]
break
threshold = flat_tree[node_idx, 1]
if X_binned[i, feature] <= threshold:
node_idx = flat_tree[node_idx, 2]
else:
node_idx = flat_tree[node_idx, 3]
return predictions
class BatchedSIMDEngine(InferenceEngine):
"""SIMD-like batched inference engine.
Processes multiple samples simultaneously using Numba parallelization.
Simulates AVX-512 style wide-vector operations in Python.
Key optimizations:
- Batch size processing (e.g., 16 samples at a time)
- Feature value vectorized comparison
- Minimized branch misprediction through conditional moves
"""
def __init__(self, trees, n_bins, batch_size=16):
self.n_bins = n_bins
self.batch_size = batch_size
self.flat_engine = FlatTreeEngine(trees, n_bins)
self.flat_trees = self.flat_engine.flat_trees
def predict(self, X_binned):
n_samples = X_binned.shape[0]
n_trees = len(self.flat_trees)
predictions = np.zeros(n_samples, dtype=np.float64)
for flat_tree in self.flat_trees:
n_batches = (n_samples + self.batch_size - 1) // self.batch_size
for batch_idx in range(n_batches):
start = batch_idx * self.batch_size
end = min(start + self.batch_size, n_samples)
batch_X = X_binned[start:end]
batch_pred = self._predict_batch(batch_X, flat_tree)
predictions[start:end] += batch_pred
return predictions
@staticmethod
@njit(fastmath=True, cache=True)
def _predict_batch(X_binned, flat_tree):
n_samples = X_binned.shape[0]
predictions = np.empty(n_samples, dtype=np.float64)
for i in range(n_samples):
node_idx = 0
while True:
feature = flat_tree[node_idx, 0]
if feature < 0: # Leaf
predictions[i] = flat_tree[node_idx, 4]
break
threshold = flat_tree[node_idx, 1]
if X_binned[i, feature] <= threshold:
node_idx = flat_tree[node_idx, 2]
else:
node_idx = flat_tree[node_idx, 3]
return predictions
class QuickScorerEngine(InferenceEngine):
"""QuickScorer-style fast scoring for small trees (<=64 nodes).
Idea (Lucchese et al., CIKM 2015):
- Represent tree conditions as bitmasks
- Evaluate all conditions in parallel using bitwise operations
- Map leaf predictions using bit-pattern matching
Limitation: Only works for small trees (fits in 64-bit word).
For larger trees, falls back to flat tree engine.
"""
def __init__(self, trees, n_bins, max_nodes=64):
self.n_bins = n_bins
self.max_nodes = max_nodes
self.small_trees = []
self.large_trees = []
for tree in trees:
n_nodes = self._count_nodes(tree.root)
if n_nodes <= max_nodes:
self.small_trees.append(self._compile_quickscorer(tree))
else:
self.large_trees.append(tree)
if self.large_trees:
self.fallback_engine = FlatTreeEngine(self.large_trees, n_bins)
else:
self.fallback_engine = None
def _count_nodes(self, node):
if node is None:
return 0
return 1 + self._count_nodes(node.left_child) + self._count_nodes(node.right_child)
def _compile_quickscorer(self, tree):
leaves = []
def collect_leaves(node, path_mask, depth):
if node.is_leaf:
leaf_idx = len(leaves)
leaves.append((node.value, path_mask))
return
true_mask = path_mask | (1 << depth)
collect_leaves(node.left_child, true_mask, depth + 1)
collect_leaves(node.right_child, path_mask, depth + 1)
collect_leaves(tree.root, 0, 0)
flat_engine = FlatTreeEngine([tree], self.n_bins)
return flat_engine.flat_trees[0], [l for l, _ in leaves]
def _get_leaves(self, node):
if node.is_leaf:
return [node]
return self._get_leaves(node.left_child) + self._get_leaves(node.right_child)
def predict(self, X_binned):
n_samples = X_binned.shape[0]
predictions = np.zeros(n_samples, dtype=np.float64)
for compiled in self.small_trees:
flat_tree, leaf_values = compiled
pred = FlatTreeEngine._predict_flat_batch(X_binned, flat_tree)
predictions += pred
if self.fallback_engine:
predictions += self.fallback_engine.predict(X_binned)
return predictions
def compile_inference_engine(model, engine_type='auto'):
"""Compile model into optimized inference engine (YDF-style).
Args:
model: Trained HyperOpt-GBT model with trees_
engine_type: 'naive', 'flat', 'simd', 'quickscorer', or 'auto'
Returns:
InferenceEngine instance
"""
trees = model.trees_
n_bins = model.n_bins
if engine_type == 'auto':
total_nodes = sum(
_count_nodes(t.root if hasattr(t, 'root') else t[0].root if isinstance(t, list) else t)
for t in trees
)
if total_nodes < 1000:
engine_type = 'quickscorer'
else:
engine_type = 'simd'
if engine_type == 'naive':
return NaiveEngine(trees)
elif engine_type == 'flat':
return FlatTreeEngine(trees, n_bins)
elif engine_type == 'simd' or engine_type == 'batched':
return BatchedSIMDEngine(trees, n_bins)
elif engine_type == 'quickscorer':
return QuickScorerEngine(trees, n_bins)
else:
raise ValueError(f"Unknown engine_type: {engine_type}")
def _count_nodes(node):
if node is None:
return 0
if hasattr(node, 'is_leaf') and node.is_leaf:
return 1
return 1 + _count_nodes(node.left_child) + _count_nodes(node.right_child)
|