Metadata-Version: 2.4
Name: flashEE
Version: 0.1.1
Summary: FlashEukaryoticEncoder -- a small, fast protein language model for eukaryotic proteins.
Author: Maciej Szczesny
License: MIT
Project-URL: Source, https://github.com/skurl/QuantisedEncoder
Keywords: protein,language-model,bioinformatics,eukaryotes,embeddings,quantisation
Requires-Python: >=3.9
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: torch
Requires-Dist: numpy
Requires-Dist: huggingface_hub
Dynamic: license-file

# flashee

**FlashEukaryoticEncoder** - a 4.94M-parameter protein language model trained on eukaryotic
proteins. Per-protein and per-residue embeddings, ESM-shaped. 2.2x faster than ESM2-8M with a 2.7x shallower memory slope, at 66% of the parameters.

```bash
pip install flashee
```

```python
from flashee import FlashEE

plm = FlashEE.load("flashee.pth")                      # local checkpoint, or a Hub repo id
emb = plm.embed(["MQIFVKTLTGKTITLEVEPSDTIENVK..."])    # [N, d] per-protein (mean-pooled)
res = plm.embed(seqs, per_residue=True)                # list of [Li, d]
```

From the command line:

```bash
flashee embed proteins.fasta --ckpt flashee.pth -o embeddings.npz
```

Weights are on Huggingface: [szchesny/flashee](https://huggingface.co/szchesny/flashee).
`FlashEE.load("szchesny/flashee")` pulls them; a local `.pth` path works too.

Source, full numbers and the quantisation investigation:
**https://github.com/skurl/flashEE**

MIT licensed. Dependencies: torch, numpy, huggingface_hub.
