Micro-LM

A plain C library for machine learning on edge devices. It requires configuration files produced by Desk-LM.

Micro-LM currently implements the following ML algorithms:

• Linear SVM
• Decision Tree
• Random Forest
• K-NN
• TripleES, Holt-Winters Triple Exponential Smoothing for time series

Each algorithm provides both classification and regression, for binary and multiclass problems. SVM supports only ordinal multiclass classification. Holt-Winters Triple Exponential Smoothing supports only regression.

We are extending the library to other algorithms, also unsupervised. Your voluntary contribution is welcome.

Usage

The Micro-LM files available in this repository must be compiled together with the .c and .h files produced by Desk-LM, which performs model training and cross-validation, and creates .c and .h files that store the best model's parameters.

You can compile the code as an executable or as a static library, using gcc/g++ for a Microcontroller or a desktop (e.g., thorugh Eclipse CDT or Visual Studio Code).

Configuration must be performed in ELM.h, where the user has to specify some #define, such as:

• The algorithm: SVM, DT, RF, KNN or TripleES
• DS_TEST, if you want to test performance in a dataset, instead of doing one shot estimations. Used only by: knn, decisionTree, svm, randomForest.
• REGRESSION, if you want to perform a regression. Default is classification (no regression). Used only by: knn, decisionTree, svm, randomForest. TripleES performs only regression

ELM.h exposes the following functions:

• float* preprocess(float* X), where X is the sample vector. It returns the pointer to the processed vector. Use that pointer for the classification / regression function.

• int (*pClassf)(float[] X_processed), where X_processed is the sample vector. pClassf points to the classification function of the selected algorithm. It returns the value of the estimated class for the input.

• float (*pRegress)(float[] X_processed), where X_processed is the sample vector. pRegress points to the regression function of the selected algorithm. It returns the estimated value for the given input.

• HW_TripleExpoSmoothing(int arrayD[], int vlen, double alpha, double beta, double gamma,int slen, int n_preds, double scaling_factor), for Holt-Winters time series

Test.h exposes the following function:

• void RunTest(), for the whole dataset testing. The algorithm is automatically selected by setting the algorithm constant in ELM.h (see above). This function prints the accuracy rate on Console. N.B. Preprocess is built in this function.

Data type

float 32 data are used

Algorithm template

Inside this repository there are ExampleAlgoTemplate.h and ExampleAlgoTemplate.c. Those files should be used as base schema to develop more Machine Learning Algorithm that are fully compatible to the ELM.h structure.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Contributing

Please see CONTRIBUTING.md

Credits:

• Alessio Capello for Random Forest implementation and refactoring;
• Gabriele Campodonico for TripleES implementation;
• Laura Pisano for TripleES implementation;

Reference article for more infomation

F., Sakr, F., Bellotti, R., Berta, A., De Gloria, "Machine Learning on Mainstream Microcontrollers," Sensors 2020, 20, 2638. https://www.mdpi.com/1424-8220/20/9/2638

References

https://medium.com/open-machine-learning-course/open-machine-learning-course-topic-9-time-series-analysis-in-python-a270cb05e0b3