mssibyl 0.1.1.dev0

Time Series Forecasting Tool for Machine Learning Researchers with Nerual Process

Sibyl: A Time-Seires Forecasting Package

Sibyl is a package for forecasting time series data with PyTorchLightning which is a lightweight PyTorch wrapper for Machine Learning Researchers.

Sibyl is released and maintained by Microsoft Substrate Connectivity Team.

Models

• Seq to Seq
• Long-short term memory
• Neural process
• Attentive neural process

Installation

Sibyl is avaialbe on PyPi, feel free to get it via pip:

pip install mssibyl

Quick Start

UPDATING ...

import pandas as pd
from mssibyl.models import BaseModel
from mssibyl.trainer import BaseTrainer
from mssibyl.datasets.utils import make_timeline

df = pd.read_csv("./data.csv")
df.columns = ["ds", "y"]

df_with_future = make_timeline(df, freq="H", periods=300)

args = {
    "lookback": 48,
    "forecast_horizon": 1,
    "data": df_with_future,
    "target": "y",
}

sibyl_model = BaseModel(args)
trainer = BaseTrainer(
    max_epochs=100, gpus=1, show_progress_bar=False, early_stop_callback=True
)
trainer.fit(sibyl_model)
preds = sibyl_model.predict(interval=True, include_history=True)

preds.to_csv("res.csv", index=False)

Changelog

Version 0.1 (coming)

• Initial release

License

Sibyl is licensed under the MIT license