Get started¶

Install¶

TimeNet needs Python 3.11 or newer. The core install stays small. The CLI and the PyTorch loader are extras. You can add them.

uv add timenet                # core: TimeF format, reader/writer, registry
uv add 'timenet[cli]'         # add the timenet console command
uv add 'timenet[torch]'       # load_torch; reuses your torch, or pulls the default build

pip install timenet
pip install 'timenet[cli]'
pip install 'timenet[torch]'

uv tool install 'timenet[cli]'       # the `timenet` command
uv tool install timenet-connectors   # `timenet-build` (connector authors)
# or, with pipx:  pipx install 'timenet[cli]'

The torch extra accepts any torch build. If you already have a CUDA torch (for example, for training), you keep it as-is. For a small CPU-only torch, install it from the PyTorch CPU index first:

pip install torch --index-url https://download.pytorch.org/whl/cpu

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

git clone https://github.com/OpenTSLM/TimeNet.git
cd TimeNet
make sync  # install the dev environment (workspace + extras)

Load a dataset¶

timenet-build build timenet/hello-world

The build writes into your local registry. The TimeNet client looks there by default. Now load the dataset:

import pandas as pd
from timenet.client import TimeNet

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

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

load reads the dataset into a TimeFDataset with lazy per-series values. to_arrow() and to_numpy() on a TimeSeries pull the values on demand. See Client for search, version pinning, PyTorch, and the CLI. See Connectors and Build to build your own datasets.