Get Started¶
IntelX needs Python 3.11 and an NVIDIA GPU. The simulator casts its rays in
numba @cuda.jit kernels, so the GPU is required. Inference falls back to CPU,
the simulator does not.
1. Install¶
git clone https://github.com/IRVLUTD/IntelX.git
cd IntelX
python3.11 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
pip install -e .
requirements.txt is the frozen environment this release was tested in. It
takes torch 2.11 and torchvision 0.26 in their CUDA 12.8 builds from the
PyTorch index, which run on every GPU from Turing (sm_75) to Blackwell
(sm_120), and numba-cuda, which brings the CUDA compiler the kernels are
built with, so no system CUDA toolkit is needed. On an older GPU, install a
torch build that supports it first, then pip install -e ., which resolves
the rest from pyproject.toml, the loose set of packages this tree imports.
2. Add the Checkpoint¶
Download checkpoint_best.pth
(293.4 MB) and put it in weights/. Its checksum and what it holds are on the
Checkpoint page.
3. Get the Maps¶
Three maps ship with the repository, one of each kind. The
Box folder holds
every test, out-of-distribution and real map; the contents of its maps/
folder go in data/maps/. The 18 training maps and the coverage dataset (about
1 TB) are not public yet; collect_data regenerates the dataset from the
training maps. Maps and Data has the layout.
4. Run IntelX¶
python -m intelx.cli.run_intelx # IntelX on Small Office 3, drawn live
python -m intelx.cli.run_method --method nearest --map ood/aws_maps/hospital.pgm
python -m intelx.cli.run_sweep_all --selftest # the benchmark's self-test, no GPU
The first explores Small Office 3 until 99% of the free space is covered and
writes the video to results/run_intelx/. Running IntelX covers
the settings, any single method and the paths.
5. Check the Tree¶
The tests read the source rather than importing it, so none of them needs torch, CUDA or a checkpoint. The Code Map lists what each one checks.
Next¶
- The Benchmark: every method from the same spawns, and the MapEx service the prediction baselines need.
- Training the Regressor: collect coverage data, train the U-Net and evaluate it.