tsfast 0.1.4

A description of your project

TSFast

Description

tsfast is an open-source deep learning library for time series analysis and system identification tasks. Built on PyTorch & fastai, it offers efficient deep learning models and utilities.

tsfast is an open-source deep learning package that focuses on system identification and time series analysis tasks. Built on the foundations of PyTorch and fastai, it provides efficient implementations of various deep learning models and utilities.

Key Features

• Specialized Data Handling for Time Series:
  • Employs SequenceBlock (built on fastai.TransformBlock) for robust sequence data processing pipelines.
  • Includes a range of transforms tailored for sequences, such as SeqSlice, SeqNoiseInjection, and Normalize adapted for time series tensors.
  • Features advanced data loading with TbpttDl (for Truncated Backpropagation Through Time), and factories for WeightedDL and BatchLimitDL.
• Predefined Datasets & Helpers: Offers easy-to-use benchmark datasets (e.g., create_dls_silverbox from identibench) for rapid prototyping and experimentation.
• Tailored Time Series Models: Provides implementations of Recurrent Neural Networks (RNNs, including DenseNet_RNN, ResidualBlock_RNN), Convolutional Neural Networks (TCNs, CausalConv1d), and combined architectures (CRNN, SeperateCRNN) specifically designed for sequence modeling. Includes building blocks like SeqLinear and stateful batch normalization.
• Integrated fastai Learner: Features RNNLearner, TCNLearner, CRNNLearner, etc., extending fastai’s Learner for streamlined model training, equipped with custom time-series losses (e.g., fun_rmse, nrmse) and callbacks (e.g., TbpttResetCB, ARInitCB, SkipFirstNCallback).
• System Identification & Prediction:
  • Supports simulation (prediction based on inputs) and N-step ahead forecasting.
  • Includes specialized models and callbacks for system identification tasks like FRANSYS (FranSys, FranSysCallback) and AR models (AR_Model, ARProg).
  • Provides an InferenceWrapper for easier model deployment and prediction.
• Hyperparameter Optimization: Integrates with Ray Tune via HPOptimizer for efficient hyperparameter searching.

Installation

You can install the latest stable version from pip using:

pip install tsfast

For development installation:

git clone https://github.com/daniel-om-weber/tsfast
cd tsfast
pip install -e '.[dev]'

Quick Start

Here is a quick example using a benchmark dataloader. It demonstrates loading and visualizing data, training a RNN, and visualizing the results.

from tsfast.basics import *
dls = create_dls_silverbox()
dls.show_batch(max_n=1)

lrn = RNNLearner(dls)
lrn.fit_flat_cos(1)

<style> /* Turns off some styling */ progress { /* gets rid of default border in Firefox and Opera. */ border: none; /* Needs to be in here for Safari polyfill so background images work as expected. */ background-size: auto; } progress:not([value]), progress:not([value])::-webkit-progress-bar { background: repeating-linear-gradient(45deg, #7e7e7e, #7e7e7e 10px, #5c5c5c 10px, #5c5c5c 20px); } .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar { background: #F44336; } </style>

epoch	train_loss	valid_loss	fun_rmse	time
0	0.006716	0.006925	0.015471	01:55

lrn.show_results(max_n=1)

<style> /* Turns off some styling */ progress { /* gets rid of default border in Firefox and Opera. */ border: none; /* Needs to be in here for Safari polyfill so background images work as expected. */ background-size: auto; } progress:not([value]), progress:not([value])::-webkit-progress-bar { background: repeating-linear-gradient(45deg, #7e7e7e, #7e7e7e 10px, #5c5c5c 10px, #5c5c5c 20px); } .progress-bar-interrupted, .progress-bar-interrupted::-webkit-progress-bar { background: #F44336; } </style>

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

Citation

If you use tsfast in your research, please cite:

@Misc{tsfast,
author = {Daniel O.M. Weber},
title = {tsfast - A deep learning library for time series analysis and system identification},
howpublished = {Github},
year = {2024},
url = {https://github.com/daniel-om-weber/tsfast}
}