Deep Latent Variable Path Modelling (DLVPM) is a method for path/structural equation modelling utilising deep neural networks. The aim of the method is to connect different data types together via sets of orthogonal deep latent variables (DLVs). Full documentation for this package can be found here. This work has now been published in the paper.

Backend-agnostic with Keras 3: DLVPM is implemented on Keras 3 and runs backend-agnostically on either TensorFlow or PyTorch; select your preferred backend by installing the corresponding extra (e.g., tf-cpu, tf-gpu, tf-apple, torch-cpu, torch-apple, or preinstall CUDA PyTorch for torch-gpu) and, if needed, set KERAS_BACKEND=tensorflow or KERAS_BACKEND=torch. The high-level Keras API (model.fit, model.evaluate, etc.) is unchanged across backends.

NOTE: A pure PyTorch version of this project, contributed by Johannes Scherer, is available on the separate pytorch branch.

If you find this project valuable, please consider giving it a star on GitHub. Your support helps others discover the project and motivates continued development!

The gif above shows a plot of model training in progress on a three factor DLVPM model linking different omics and Imaging data types from Lung Cancer patients. The dataset this model is trained on is included in this package.

Installing deep-lvpm (keras 3 / multi-backend) — keras3 branch

TL;DR: This branch uses Keras 3 (multi-backend). You must install one backend (TensorFlow or PyTorch) via the provided extras. Then install from the keras3 branch URL.

---

Choose a backend

TensorFlow (default)

• CPU: tf-cpu
• NVIDIA GPU on Linux (bundled CUDA wheels): tf-gpu
• Apple Silicon (M-series): tf-apple

PyTorch

• CPU (Linux/Windows/macOS Intel): torch-cpu
• Apple Silicon (M-series): torch-apple
• NVIDIA GPU: torch-gpu (empty extra; preinstall CUDA wheels from pytorch.org first)

Python

• Python 3.11 or 3.12

---

Conda (recommended)

Create and activate an environment:

conda create -n dlvpm-k3 python=3.11 -y
conda activate dlvpm-k3

TensorFlow backends

CPU (Linux/Windows/Intel Mac):

pip install "git+https://github.com/alexjamesing/Deep_LVPM.git@keras3#egg=deep-lvpm[tf-cpu]"

NVIDIA GPU (Linux, bundled CUDA):

pip install "git+https://github.com/alexjamesing/Deep_LVPM.git@keras3#egg=deep-lvpm[tf-gpu]"

Apple Silicon (M-series):

pip install "git+https://github.com/alexjamesing/Deep_LVPM.git@keras3#egg=deep-lvpm[tf-apple]"

PyTorch backends

CPU (Linux/Windows/macOS Intel):

pip install "git+https://github.com/alexjamesing/Deep_LVPM.git@keras3#egg=deep-lvpm[torch-cpu]"

Apple Silicon (M-series):

pip install "git+https://github.com/alexjamesing/Deep_LVPM.git@keras3#egg=deep-lvpm[torch-apple]"

NVIDIA GPU (CUDA):

# 1) Install CUDA-enabled PyTorch from https://pytorch.org (per your CUDA/driver), e.g.:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
# 2) Then install deep-lvpm without pulling CPU wheels:
pip install "git+https://github.com/alexjamesing/Deep_LVPM.git@keras3#egg=deep-lvpm[torch-gpu]"

---

venv

Create and activate a virtual environment:

python3 -m venv dlvpm-k3
source dlvpm-k3/bin/activate   # Windows: dlvpm-k3\Scripts\activate

Then choose one of the same install lines as above (e.g., ...[tf-cpu], ...[tf-gpu], etc.).

---

Verifying the backend

python -c "import keras, os; print('KERAS_BACKEND=', os.getenv('KERAS_BACKEND')); print('Detected backend:', keras.backend.backend())"

If needed, set the backend explicitly and re-run:

export KERAS_BACKEND=tensorflow   # or: torch

Notes

• tf-gpu is Linux-only and uses tensorflow[and-cuda] (no separate CUDA toolkit install required).
• torch-gpu extra is intentionally empty to avoid pulling CPU wheels from PyPI; always install CUDA-enabled PyTorch first.
• If multiple backends are installed, Keras will pick one; use KERAS_BACKEND to force your choice.

Tutorials & Metrics

Three turnkey tutorials ship with the toolbox and default to the TensorFlow backend. Launch them with:

• python -m deep_lvpm.tutorial.tutorial_mnist_tf – associate MNIST images with labels and visualise the latent space.
• python -m deep_lvpm.tutorial.tutorial_tcga_tf – integrate five TCGA lung cancer modalities using residual encoders.
• python -m deep_lvpm.tutorial.tutorial_siamese_tf – train a Siamese encoder on CIFAR-10 and probe the embeddings linearly.

All tutorials report the expanded StructuralModel.evaluate metrics (total_loss, cross_metric, mse_loss, and redundancy) introduced in this release so you can monitor both cross-view alignment and within-view redundancy.