Code for our paper "Disentangling shared and private latent factors in multimodal Variational Autoencoders" which was published in the proceedings of MLCB 2023.
Here is a brief overview of the codebase:
- In experiments/core.py, we provide the implementation and training script for various versions of MVAEs, MMVAEs and MoPoE-VAEs.
- In multimodalVAE/MVAE.py and multimodalVAE/MMVAE.py, you can find the main model classes.
- We use hydra framework to configure the experiments and W&B for tracking. The main file containing the wrapper for training and evaluation is experiments/experiment.py.
Below are examples how to train various versions of multimodal VAEs. Note that model.name can take values in [MVAE, MoPoE-VAE, MMVAE, MMVAE++].
Data generated using the GP model (loaded via dataset.name=toy_GP). As this is a synthetically generated dataset, we can additionally provide dataset.dim_z="2+2+2".
python experiments/experiment.py --multirun model.name=MMVAE,MMVAE++,MVAE,MoPoE-VAE dataset.name=toy_GP dataset.n_genes_top=10 dataset.n_genes_bottom=20,50,100,250 +dataset.n_cpgs_bottom=100 dataset.dim_z="2+2+2" trainer.lr=1e-3
BRCA expression and methylation data (loaded via dataset.name=BRCA_ER_exprs_and_meth)
python experiments/experiment.py --multirun model.name=MMVAE,MMVAE++,MVAE,MoPoE-VAE dataset.name=BRCA_ER_exprs_and_meth dataset.n_genes_top=1000 dataset.n_genes_bottom=500 +dataset.n_cpgs_top=1000 +dataset.n_cpgs_bottom=25000 trainer.lr=1e-4 trainer.n_epochs=500
Single cell RNA-seq and ATAC-seq experiment (loaded via dataset.name=MOFA_RNA_ATAC)
python experiments/experiment.py --multirun model.name=MMVAE,MMVAE++,MVAE,MoPoE-VAE dataset.name=MOFA_RNA_ATAC dataset.n_genes_top=10 dataset.n_genes_bottom=10,100,1000 +dataset.n_cpgs_bottom=5000 trainer.lr=1e-4 trainer.n_epochs=100