# Get started

## Install

TimeNet needs Python 3.11 or newer. The core install stays lean; the CLI and PyTorch loader are
extras you opt into.

!!! info "Coming soon"
    The `timenet` packages aren't on PyPI yet. Install instructions land here once they ship.

To work on TimeNet or author connectors, clone the repo and sync with uv:

```bash
git clone https://github.com/OpenTSLM/TimeNet.git
cd TimeNet
uv sync --all-groups --all-extras
```

## Load a dataset

!!! info "No public registry yet"
    There's no hosted registry to pull from, so build the offline `timenet/hello-world` dataset
    into a local registry first. It needs no network and comes from `timenet-connectors`.

```bash
timenet-curate build timenet/hello-world
```

That writes into your local registry, which is where the [`TimeNet`](client.md) client looks by
default. Load the dataset:

```python
import pandas as pd
from timenet.client import TimeNet

dataset = TimeNet().load("timenet/hello-world")
dataset.describe()                                # identity, counts, a quick preview

# Each channel converts to Arrow or NumPy, so it drops straight into pandas:
series = dataset.samples[0].time_series[0]
df = pd.DataFrame({series.channel: series.to_numpy()})
print(df.head())
```

!!! tip "pandas is optional"
    The DataFrame step uses pandas (`uv add pandas`). It isn't a TimeNet dependency. Drop it for a
    pure-NumPy workflow.

`load` reads the dataset into a [`TimeFDataset`](timef-dataset.md) with lazy per-series values;
`to_arrow()` / `to_numpy()` on a [`TimeSeries`](timef-dataset.md) pull the values on demand. See
[Client](client.md) for search, version pinning, PyTorch, and the CLI, and
[Connectors](connectors.md) / [Curation](curation.md) to build your own datasets.