tpcp 0.13.0

Pipeline and Dataset helpers for complex algorithm evaluation.

tpcp - Tiny Pipelines for Complex Problems

A generic way to build object-oriented datasets and algorithm pipelines and tools to evaluate them.

Easily install tpcp via pip:

pip install tpcp

Or add it to your project with poetry:

poetry add tpcp

Why?

Evaluating Algorithms - in particular when they contain machine learning - is hard. Besides understanding required concepts (cross validation, bias, overfitting, ...), you need to implement the required steps and make them work together with your algorithms and data. If you are doing something "regular" like training an SVM on tabular data, amazing libraries like sklearn, tslearn, pytorch, and many others, have your back. By using their built-in tools (e.g. sklearn.evaluation.GridSearchCV) you prevent implementation errors, and you are provided with a sensible structure to organize your code that is well understood in the community.

However, often the problems we are trying to solve are not regular. They are complex. As an example, here is the summary of the method from one of our recent papers:

• We have continuous multi-dimensional sensor recordings from multiple participants from a hospital visit and multiple days at home
• For each participant we have global metadata (age, diagnosis) and daily annotations
• We want to train a Hidden-Markov-Model that can find events in the data streams
• We need to tune hyper-parameters of the algorithm using a participant-wise cross validation
• We want to evaluate the final performance of the algorithm for the settings trained on the hospital data -> tested on home data and trained on home data -> tested on home data
• Using the same structure we want to evaluate a state-of-the-art algorithm to compare the results

None of the standard frameworks can easily abstract this problem, because here we have none-tabular data, multiple data sources per participant, a non-traditional ML algorithm, and a complex train-test split logic.

With tpcp we want to provide a flexible framework to approach such complex problems with structure and confidence.

How?

To make tpcp easy to use, we try to focus on a couple of key ideas:

• Datasets are Python classes (think of pytorch.datasets, but more flexible) that can be split, iterated over, and queried.
• Algorithms and Pipelines are Python classes with a simple run and optimize interface, that can be implemented to fit any problem.
• Everything is a parameter and everything is optimization: In regular ML we differentiate training and hyper-parameter optimization. In tpcp we consider everything that modifies parameters or weights as an optimization. This allows to use the same concepts and code interfaces from simple algorithms that just require a grid search to optimize a parameter to neuronal network pipelines with hyperparameter tuning.
• Provide what is difficult, allow to change everything else: tpcp implements complicated constructs like cross validation and grid search and, whenever possible, tries to catch obvious errors in your approach. However, for the actual algorithm and dataset you are free to do whatever is required to solve your current research question.

Should you use tpcp?

Datasets

Yes - the object-oriented Datasets have proven themselves to be a really nice and flexible way to encapsulate Datasets with data from multiple modalities. There is a clear path of integrating lazy-loading, load-cashing, data filtering, or pre-processing on loading. From our experience, even if you ignore all the other tpcp features, Datasets can greatly simplify how you interact with your data sources and can serve as a self-documenting API for your data.

Learn more (Examples)

Other projects using Datasets:

• gaitmap-datasets
• cold-face-test-analysis
• carwatch-analysis

Parameter Optimization and Cross Validation

Maybe - All parameter optimization features in tpcp exist to provide a unified API, in case other specific frameworks are too specialised. In cases where all your algorithms can be abstracted by sklearn, pytorch (with the skorch wrapper), tensorflow/Keras (with the scikeras wrapper), or any other framework that provides a nice scikit-learn API, you will get all the features tpcp can provide with much less boilerplate by just using sklearn and optuna directly. Even, if you need to implement completely custom algorithms, we would encourage you to see if you can emulate a sklearn-like API to make use of its vast ecosystem.

This will usually work well for all algorithms that can be abstracted by the fit-predict paradigm. However, for more "traditional" algorithms with no "fit" step or complicated optimizations, the run (with optional self_optimize) API of tpcp might be a better fit. So if you are using or developing algorithms across library domains, that don't all work well with a sklearn API, then Yes, tpcp is a good choice.

Learn more: General Concepts, Custom Algorithms, Parameter Optimization, Cross Validation

Dev Setup

We are using poetry to manage dependencies and poethepoet to run and manage dev tasks.

To set up the dev environment including the required dependencies for using and developing on tpcp run the following commands:

git clone https://github.com/mad-lab-fau/tpcp
cd tpcp
poetry install -E optuna -E torch # This might take a while

Note, that this will fail at the moment for Python 3.11., as pytorch is not yet available for this version. You can safely skip the -E torch flag if you don't need the pytorch features. All tests that require pytorch will be skipped automatically.

Afterwards you can start to develop and change things. If you want to run tests, format your code, build the docs, ..., you can run one of the following poethepoet commands

CONFIGURED TASKS
  format         
  lint           Lint all files with Prospector.
  check          Check all potential format and linting issues.
  test           Run Pytest with coverage.
  docs           Build the html docs using Sphinx.
  bump_version   

by calling

poetry run poe <command name>

If you installed poethepoet globally, you can skip the poetry run part at the beginning.

Contribution

The entire development is managed via GitHub. If you run into any issues, want to discuss certain decisions, want to contribute features or feature requests, just reach out to us by opening a new issue.