deploy

BiLoc 1-bit GPU runtimes (Oxford)

Prebuilt 1-bit GPU runtimes for the BiLoc Oxford model, with the bit-packed Oxford weights. Two variants are available: fused and eqk.

Requirements

• Linux x86_64, Python 3.9
• PyTorch 2.8.0+cu128
• An NVIDIA GPU of compute capability 12.0 (sm_120)
• The BiLoc Python environment (the same as for test.py)

Usage

Run from the biloc/ directory:

import deploy

rt = deploy.build_runtime('fused')   # or 'eqk'
pose = rt.infer(image)               # image: [B, 3, 5, 32, 512] float32 CUDA tensor -> pose: [B, 3, 6]

image is a batch of normalized range images as produced by the BiLoc data loader; pose is the normalized pose of every frame, the same as pred_pose of the PyTorch model.

Evaluation

cd biloc
python -m deploy.evaluate --dataroot ../data --variant fused   # or --variant eqk

--dataroot is the directory that contains Oxford/.

Latency

cd biloc
python -m deploy.benchmark --dataroot ../data                  # fused and eqk
python -m deploy.benchmark --dataroot ../data --with-diffloc   # also the real-valued DiffLoc

Run either script with --help for all options.