Skip to content

Installation Guide

kreview is built to be a high-performance evaluation environment. It depends on DuckDB, XGBoost, SHAP, and scientific Python libraries.

Requirements

  • Python ≥ 3.10
  • pip (latest)

Environment Setup

[!IMPORTANT] The report is a single self-contained HTML rendered directly by kreview report — no external tooling (Quarto was removed in the #79 report redesign).

The easiest way to run kreview without managing external dependencies is to use our pre-built Docker container (hosted on GHCR). It ships with Python 3.12 and all ML libraries: 

docker pull ghcr.io/msk-access/kreview:latest
docker run -v /your/data:/data ghcr.io/msk-access/kreview:latest \
  label --cancer-samplesheet /data/cancer.csv ...   # stages are driven by Nextflow
For more complex execution commands (e.g., binding multiple access paths), see the Docker Operations Guide.

Option 2: Local Install (Pip)

Then clone the repository. kreview is developed entirely using nbdev, so the source notebooks (nbs/) act as the active execution environment.

git clone https://github.com/msk-access/kreview.git
cd kreview

2. Install the Package

Install the core package with all runtime dependencies:

pip install -e .

For GPU model evaluation (TabPFN, TabICL and fine-tuned variants), install with GPU extras: 

pip install -e '.[gpu]'
This adds tabpfn, tabicl, shapiq, and torch as dependencies.

Install with all linting, testing, CI, and documentation tools bundled: 

pip install -e '.[all]'

Optional Sub-Packages

If you only need specific toolchains instead of the full all suite: - Jupyter Only: pip install -e '.[jupyter]' - Testing Only: pip install -e '.[test]' - Docs Only: pip install -e '.[docs]' - Multimodal selectors (GrootCV/Leshy): pip install -e '.[arfs]' — required for --multimodal-selection grootcv|leshy (the containers ship it by default).

[arfs] and [legacy-boruta] are mutually exclusive

The deprecated boruta_shap selector needs pip install -e '.[legacy-boruta]', which cannot coexist with the arfs extra: BorutaShapPlus pins numpy<=2.0.0 while arfs 3.0 requires numpy>=2.0.2 — pip refuses any environment containing both. legacy-boruta exists only to reproduce pre-#96 runs.

3. Install Git Hooks

Development Hook

If you are contributing code, install the pre-commit hooks to automatically strip Jupyter notebook metadata and run linters before each commit:

nbdev-install-hooks
pre-commit install

4. Verification Check

To quickly verify that the CLI was successfully mapped:

kreview --help

You should see an output tree listing run, label, features-list, and report. If so, you're clear to proceed to Configuration!

Listing Registered Features

You can immediately verify all 26 feature evaluators are discoverable: 

kreview features-list