# caveman

Audio reconstruction model, that "uncompresses" lossy mp3s.

## Installing dependencies

There is the pyproject file, so you can use that.

Or, if you prefer to keep your sanity, use uv:

```
uv sync
```

## Training

Adjust the location of your data directories via the variables in the cavemanml.py file. Any compression type is accepted, as long as your audio is 44100kHz.

```
uv run cavemanml.py
```

Obviously the original dataset (FMA) is not provided with this repo.

## Inference

For now, the only inference this code can do, is the example. You can adjust the input and output file in the code, though. Theoretically, the model should take on anything. To use the provided checkpoint:

```
uv run cavemanml.py ./checkpoint70.pth
```