---
license: mit
tags:
  - phylogenetics
  - biology
---

# DeepPhylo checkpoints

Trained models for *Deep Generative Models for Phylogenetic Inference with
Complex Evolutionary Processes*. These are the exact checkpoints behind the
paper's NELBO bar chart and its RF-vs-sites / RF-vs-taxa figures.

Five methods (BU, MO, TD, PF2, QF) on three simulated datasets: `jc` (DNA, Jukes-Cantor), `cat` (amino
acids, CAT mixture), `indels` (DNA, GTR with indels and realignment).

| Directory | Method | Dataset | Step | Size |
|---|---|---|---|---|
| `BU_JC` | Bottom up | JC | 26,000 | 224 MiB |
| `MO_JC` | Middle out | JC | 26,000 | 224 MiB |
| `TD_JC` | Top down | JC | 22,000 | 224 MiB |
| `PF2_JC` | Phyloformer 2 (baseline) | JC | 16,000 | 299 MiB |
| `QF_JC` | QuartFormer (baseline) | JC | 56,000 | 26 MiB |
| `BU_CAT` | Bottom up | CAT | 26,000 | 224 MiB |
| `MO_CAT` | Middle out | CAT | 26,000 | 224 MiB |
| `TD_CAT` | Top down | CAT | 22,000 | 224 MiB |
| `PF2_CAT` | Phyloformer 2 (baseline) | CAT | 16,000 | 299 MiB |
| `QF_CAT` | QuartFormer (baseline) | CAT | 60,000 | 26 MiB |
| `BU_Indels` | Bottom up | Indels | 26,000 | 224 MiB |
| `MO_Indels` | Middle out | Indels | 24,000 | 224 MiB |
| `TD_Indels` | Top down | Indels | 22,000 | 224 MiB |
| `PF2_Indels` | Phyloformer 2 (baseline) | Indels | 14,000 | 299 MiB |
| `QF_Indels` | QuartFormer (baseline) | Indels | 50,000 | 26 MiB |

Each directory holds `model.ckpt` and the `config.yaml` of the run that produced
it, which is what the code reads to rebuild the model.

## Usage

Clone [DeepPhylo](https://github.com/AlanNawzadAmin/DeepPhylo), then place a
directory here under `checkpoints/` and name it in `run=`:

```bash
python sample.py run=BU_JC sampling.path=out.pkl
```

See the repository README for greedy vs. consensus decoding.

`MANIFEST.tsv` lists the training run id, step, size and SHA-256 of every file.