The incremental_learning module provides tools and classes for implementing and managing online learning models. It includes functionalities for processing data, managing shared memory buffers, and training/testing models in real-time.

• execution_modes_buffers

  • Handles different execution modes for training and testing models using shared memory buffers.

• inter_process_communication

  • Manages communication between processes, including socket communication for sending and receiving data.

• online_data_processing

  • Contains functions for processing online data, including generating observations and features from raw data.

• online_models

  • Implements various online learning models and their training/testing routines.

• ping_pong_buffers

  • Manages ping-pong buffers for sharing data among independent processes using shared memory.

There are several scripts available to demonstrate different ways to use the incremental learning module:

• all_models_train.py

  • This script trains all available online models using a predefined dataset and evaluates their performance.

• online_traces_processing.py

  • This script processes power and performance traces read from local files and generates a pickle file containing a dataframe with all processed observations.

• open_matplotlib_processing.py

  • This script opens and processes data using Matplotlib for visualization, allowing for real-time plotting of model performance.

• postprocessing_model_data.py

  • This script processes the online models' output data and generates error metrics figures. It includes functions for visualizing model data with and without time information.

• runtime_adaptive_models.py

  • This script adapts the models in real-time based on incoming data, adjusting parameters to improve performance dynamically. Needs to be run on the board (remember to copy the incremental_learning module.)

For run-time usage on the board, this module requires the FPGA Workload Manager. This manager leverages the incremental learning module for workload modeling and management.

This project is described in detail in the accompanying paper. You can read the paper here.