EIR
A toolkit for training deep learning models on genotype, tabular, sequence, image, array and binary data.
Install / Use
/learn @arnor-sigurdsson/EIRREADME
Supervised modelling, sequence generation, image generation, array output and survival analysis on genotype, tabular, sequence, image, array, and binary input data.
WARNING: This project is in alpha phase. Expect backwards incompatible changes and API changes between minor versions.
What's New
- July 2025: More docs added
- June 2025: Overhaul, expand and restructure documenation
- April 2025: More scaling related tutorials
- Scaling Diffusion: Image Generation from Text - Exploring artistic styles from historical collections
- Supervised Fine Tuning of GPT Style Model - Fine tuning a pre-trained model for instruction following
- March 2025: Scaling tutorials added
- Streaming Data: Training with FineWeb - Train models via continuous data streaming
- Scaling Compute: Training a BabyGPT - Examples of training GPT-style models
Table of Contents
- Install
- Usage
- Use Cases
- Features
- Supported Inputs and Outputs
- Related Projects
- Citation
- Acknowledgements
Install
Installing EIR via pip
pip install eir-dl
Important: The latest version of EIR requires Python 3.13. Using an older version of Python will install an outdated version of EIR, which will likely be incompatible with the current documentation and might contain bugs. Please ensure you are using Python 3.13.
Installing EIR via Container Engine
Here's an example with Docker:
docker build -t eir:latest https://raw.githubusercontent.com/arnor-sigurdsson/EIR/master/Dockerfile
docker run -d --name eir_container eir:latest
docker exec -it eir_container bash
Usage
Please refer to the Documentation for examples and information.
Use Cases
EIR allows for training and evaluating various deep-learning models directly from the command line. This can be useful for:
- Quick prototyping and iteration when modelling on new datasets.
- Establishing baselines to compare against other methods.
- Fitting on data sources such as large-scale genomics, where DL implementations are not commonly available.
If you are an ML/DL researcher developing new models, etc., it might not fit your use case. However, it might provide a quick baseline for comparison to the cool stuff you are developing, and there is some degree of customization possible.
Features
-
General
- Train models directly from the command line through
.yamlconfiguration files. - Training on genotype, tabular, sequence, image, array and binary input data, with various modality-specific settings available.
- Seamless multi-modal (e.g., combining text + image + tabular data, or any combination of the modalities above) training.
- Train multiple features extractors on the same data source, e.g., combining vanilla transformer, Longformer and a pre-trained BERT variant for text classification.
- Support for checkpointing and continued training, as well as pretraining and transferring parts of trained models to new tasks.
- Train models directly from the command line through
-
Supervised Learning
- Supports continuous (i.e., regression) and categorical (i.e., classification) targets.
- Multi-task / multi-label prediction supported out-of-the-box.
- Model explainability for genotype, tabular, sequence, image and array data built in.
- Computes and graphs various evaluation metrics (e.g., RMSE, PCC and R2 for regression tasks, accuracy, ROC-AUC, etc. for classification tasks) during training.
-
Sequence Generation
- Supports various sequence generation tasks, including basic sequence generation, sequence to sequence transformations, and image to sequence transformations. For more information, refer to the respective tutorials: sequence generation, sequence to sequence, image to sequence and tabular to sequence.
-
Image Generation
- Image generation is supported. For more information, refer to the respective tutorials: Building a Simple Image Autoencoder, Image Colorization and Super-Resolution and Guided Diffusion for Image Generation.
-
Array Output
- Supports array output tasks, such as building simple autoencoders for tasks like MNIST Digit Generation.
-
Time Series
- Time series inputs and outputs is possible, such as Transformer-based Power Consumption Prediction and Stock Price Prediction Using Transformers, One-shot and Diffusion Models.
-
Survival Analysis
- Time-to-event prediction is supported as an output type, demonstrated through Patient Survival Prediction using Free Light Chain Data and Survival Analysis Using Cox Proportional Hazards Model.
-
Scaling
- Support for streaming data training enabling models to train on datasets too large to fit in memory or with real-time data.
- Capabilities for scaling compute resources to train larger language models like GPT variants.
-
Many more settings and configurations (e.g., augmentation, regularization, optimizers) available.
Supported Inputs and Outputs
| Modality | Input | Output | |------------|:-----:|:------:| | Genotype | ✓ | † | | Tabular | ✓ | ✓ | | Sequence | ✓ | ✓ | | Image | ✓ | ✓ | | Array | ✓ | ✓ | | Binary | ✓ | | | Survival | n/a | ✓ |
† While not directly supported, genotypes can be treated as arrays. For example see the MNIST Digit Generation tutorial.
Related Projects
- EIR-auto-GP: Automated genomic prediction (GP) using deep learning models with EIR.
Citation
If you use EIR in a scientific publication, we would appreciate if you could use o
