Practical overview
When it fits
Use NeuroCard when investigating correlation-aware cardinality estimates or reproducing its documented training and evaluation experiments, particularly on the JOB configurations.
Who it is for
- Database researchers studying learned cardinality estimation
- Engineers evaluating join-size estimates for query optimization
- Practitioners reproducing the repository's documented JOB benchmark experiments
Documented uses
- Estimate row counts for multi-table join queries.
- Train or evaluate neural cardinality estimators using registered experiment configurations.
- Compare estimation quality using Q-error metrics and inspect density-model quality using train_bits.
Requirements & limits
The repository is public and licensed under Apache-2.0.
- The documented quick start uses Ubuntu or Debian, build-essential, Conda and the IMDB CSV dataset.
- Registered configurations distinguish checkpoint-only evaluation from training followed by evaluation.
- The documented JOB-M large Transformer configuration has 107 million parameters and notes a four-GPU launch option or reduced batch size.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare the NeuroCard JOB-light test
Set up the documented environment and dataset, then launch the short test run.
Before you start
- Ubuntu or Debian with build-essential installed
- Conda
- NeuroCard repository files
- IMDB CSV dataset, either downloaded with the provided script or already available
Steps
Install build-essential with the system package manager.
Create and activate the Conda environment using environment.yml.
From the repository, download the IMDB CSV files with scripts/download_imdb.sh, or provide existing CSVs under datasets/job.
Run python run.py --run test-job-light.
Expected outcome
The README describes this command as a short test run.
Guide limitations
- The README identifies the dataset download as 1.2 GB.
- If provided CSV files lack headers, the README points to scripts/prepend_imdb_headers.py.
Run a registered training or evaluation experiment
Launch a repository experiment configuration for training and evaluation or checkpoint evaluation.
Before you start
- NeuroCard Conda environment and repository setup
- A configuration name registered in experiments.py
- The dataset and checkpoint required by the selected configuration
Steps
Choose a configuration registered in experiments.py.
For checkpoint evaluation, choose a reload configuration; for training and evaluation, choose a normal configuration.
Launch it with python run.py --run .
Expected outcome
The script runs the selected training-and-evaluation workflow or, for reload configurations, evaluation only; the README identifies Q-error and train_bits as metrics.
Guide limitations
- Configuration details and hyperparameters are statically declared in experiments.py.
- The README notes that JOB-M large Transformer training uses a four-GPU launch option or a reduced batch size.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- neurocard/neurocard · Official homepage title · retained in its source language