Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
The implementation of "TabR: Unlocking the Power of Retrieval-Augmented Tabular Deep Learning"
| Date | Stars |
|---|---|
| 2026-07-31 | 332 |
| 2026-08-05 | 332 |
| 2026-08-06 | 332 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# TabR: Unlocking the Power of Retrieval-Augmented Tabular Deep Learning<!-- omit in toc -->
> [!IMPORTANT]
> Check out the new tabular DL model: [TabM](https://github.com/yandex-research/tabm)
This is the official implementation of the paper "TabR: Unlocking the Power of Retrieval-Augmented Tabular Deep Learning" ([arXiv](https://arxiv.org/abs/2307.14338)).
Table of Contents:
- [The main results](#the-main-results)
- [How to reproduce the results](#how-to-reproduce-the-results)
- [Set up the environment](#set-up-the-environment)
- [Software](#software)
- [Data](#data)
- [Environment variables](#environment-variables)
- [Quick test](#quick-test)
- [Tutorial](#tutorial)
- [Reproducing other results](#reproducing-other-results)
- [Understanding the repository](#understanding-the-repository)
- [Code overview](#code-overview)
- [Running scripts](#running-scripts)
- [Technical notes](#technical-notes)
- [Adding new datasets and metrics](#adding-new-datasets-and-metrics)
- [How to add a new dataset](#how-to-add-a-new-dataset)
- [How to optimize a custom metric](#how-to-optimize-a-custom-metric)
- [How to add a new task type](#how-to-add-a-new-task-type)
# The main results
After setting up the environment, use [this notebook](notebooks/results.ipynb) to browse the main results (for now, you can scroll to the last cell to get an idea of what it looks like).
# How to reproduce the results
## Set up the environment
### Software
For this project, we highly recommend using a conda-like environment manager instead of pip to get things right for the libraries that use CUDA, especially for Faiss.
The available options:
- [mamba](https://mamba.readthedocs.io/en/latest/installation.html) is a fast replacement for conda
- (we used this) [micromamba](https://mamba.readthedocs.io/en/latest/installation.html#manual-installation) can be used to avoid any conflicts with your current setup: it is a single binary which does not require any "installation" (see the [documentation](https://mamba.readthedocs.io/en/latest/user_guide/micromamba.html))
- [conda](https://docs.conda.io/en/latest/miniconda.html) is a valid option, but setting up the environment can become extremely slow (or even impossible)
Then, run the following commands (replace `micromamba` with `mamba` or `conda` if needed):
```shell
git clone https://github.com/yandex-research/tabular-dl-tabr
cd tabular-dl-tabr
micromamba create -f environment.yaml
micromamba activate tabr
```
If the `micromamba create` command fails, try using `environment-simple.yaml` instead of `environment.yaml`.
If your machine does not have GPUs, use `environment-simple.yaml`, but replace `faiss-gpu` with `faiss-cpu` and remove `pytorch-cuda`.
### Data
(***License:** we do not impose any new license restrictions in addition to the original licenses of the used dataset.
See the paper to learn about the dataset sources*)
Navigate to the repository root and run the following commands:
```
wget https://huggingface.co/datasets/puhsu/tabular-benchmarks/resolve/main/data.tar -O tabular-dl-tabr.tar.gz
tar -xvf tabular-dl-tabr.tar.gz
```
After that, the `data/` directory should appear.
### Environment variables
**When running scripts, the environment variable `CUDA_VISIBLE_DEVICES` must be explicitly set**. So we assume that you do run the following command first before running other commands:
```
export CUDA_VISIBLE_DEVICES="0"
```
## Quick test
To check that the environment is configured correctly, run the following command and wait for the training to finish (in this experiment, hyperparameters and results are extremely suboptimal, this is needed only to test the environment):
```
python bin/ffn.py exp/debug/0.toml --force
```
The last line of the output log should look like this:
```
[<<<] exp/debug/0 | <date & time>
```
## Tutorial
Here, we reproduce the results for MLP on the California Housing dataset (in the paper, this dataset is Excerpt of 16,528 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:935933dd50bc2a0f, topic:deep-learning, topic:pytorch