Home

Overview

startorch is a Python library to generate synthetic time-series. As the name suggest, startorch relies mostly on PyTorch to generate the time series and to control the randomness. startorch is built to be modular, flexible and extensible. For example, it is easy to combine multiple core sequence generator to generate complex sequences. The user is responsible to define the recipe to generate the time series. Below show some generated sequences by startorch where the values are sampled from different distribution.

uniform	log-uniform
sine wave	Wiener process

Motivation

Collecting datasets to train Machine Learning models can be time consuming. Another alternative is to use synthetic datasets. startorch provides modules to easily generate synthetic time series. The user is responsible to define the recipe to generate the time series. The following example shows how to generate a sequence where the values are sampled from a Normal distribution.

from startorch.sequence import RandNormal

generator = RandNormal(mean=0.0, std=1.0)
batch = generator.generate(seq_len=128, batch_size=4)

It is possible to combine multiple generators to build a more complex generator. The example below shows how to build a generator that sums multiple the output of three sine wave generators.

from startorch.sequence import (
    Add,
    Arange,
    Constant,
    RandLogUniform,
    RandUniform,
    SineWave,
)

generator = Add(
    (
        SineWave(
            value=Arange(),
            frequency=Constant(RandLogUniform(low=0.01, high=0.1)),
            phase=Constant(RandUniform(low=-1.0, high=1.0)),
            amplitude=Constant(RandLogUniform(low=0.1, high=1.0)),
        ),
        SineWave(
            value=Arange(),
            frequency=Constant(RandLogUniform(low=0.01, high=0.1)),
            phase=Constant(RandUniform(low=-1.0, high=1.0)),
            amplitude=Constant(RandLogUniform(low=0.1, high=1.0)),
        ),
        SineWave(
            value=Arange(),
            frequency=Constant(RandLogUniform(low=0.01, high=0.1)),
            phase=Constant(RandUniform(low=-1.0, high=1.0)),
            amplitude=Constant(RandLogUniform(low=0.1, high=1.0)),
        ),
    )
)
batch = generator.generate(seq_len=128, batch_size=4)

API stability

While startorch is in development stage, no API is guaranteed to be stable from one release to the next. In fact, it is very likely that the API will change multiple times before a stable 1.0.0 release. In practice, this means that upgrading startorch to a new version will possibly break any code that was using the old version of startorch.

License

startorch is licensed under BSD 3-Clause "New" or "Revised" license available in LICENSE file.