qxgb 0.1.0

QXGBoost: Quantized XGBoost for FPGA deployment

qxgb

QXGBoost: quantized XGBoost Classifier for FPGA Deployment.

qxgb provides QXGBClassifier inheriting from xgboost.XGBClassifier with quantization aware training and registers the Alkaid tracer plugin for converting trained (Q)XGBoost boosters into RTL designs with Alkaid.

Installation

pip install qxgb

Quick Start

from qxgb import QXGBClassifier

model = QXGBClassifier(
    scale=4,
    bias=2,
    n_estimators=32,
    max_depth=3,
)

model.fit(X_train, y_train)
y_pred = model.predict(X_test)

hw_bst = model.ibooster() # Integer booster

from alkaid.converter import trace_model
from alkaid.trace import FVArray, trace

inp = FVArray.new(16).quantize(0, 8, 0).as_new()
_, out = trace_model(hw_bst, inputs=inp, mode='mux')
comb = trace(inp, out)

hw_pred = comb.predict(X_test) # == y_pred

Examples

Notebook examples are in examples/:

• jsc.ipynb
• mnist.ipynb
• nid.ipynb

License

LGPLv3. See LICENSE.