vbpi-torch

Pytorch Implementation of Variational Bayesian Phylogenetic Inference

Dependencies

• Biopython
• bitarray
• dendropy
• ete3
• PyTorch

You can build and enter a conda environment with all of the dependencies built in using the supplied environment.yml file via:

conda env create -f environment.yml
conda activate vbpi-torch

Preparation

Unzip DENV4_constant_golden_run.trees.zip in the rooted/data/DENV4 directory.

Running

These steps will reproduce the experiments in the preprint A Variational Approach to Bayesian Phylogenetic Inference.

• The evidence lower bounds will be saved to a *_test_lb.npy file.
• The KL divergences will be saved to a *_kl_div.npy file (if --empFreq is turned on).
• The trained model will be saved to a *.pt file.

In the unrooted/ folder

python main.py --dataset DS1 --psp --empFreq
python main.py --dataset DS1 --psp --nParticle 20 --gradMethod rws --empFreq
python main.py --dataset flu100 --psp
python main.py --dataset flu100 --psp --supportType mcmc -cf 100000 --maxIter 400000

One can also load the checkpoints for testing (e.g., KL computation)

python main.py --dataset flu100 --psp --supportType mcmc --empFreq --test --datetime "20xx-xx-xx xx:xx:xx.xxxxxx"

where the value for --datetime is the datetime for the saved model that you want to test.

See more concrete examples here: ds1.ipynb, flu100.ipynb.

In the rooted/ folder

python main.py --dataset DENV4 --burnin 2501 --coalescent_type constant --clock_type strict --init_clock_rate 1e-3 --sample_info --psp --empFreq
python main.py --dataset HCV --burnin 251 --coalescent_type skyride --clock_type fixed_rate --init_clock_rate 7.9e-4 --psp

See more concrete examples here: denv4.ipynb, hcv.ipynb.