SMART-HAR: Strategic Meta-learning with Adaptive Replay and Network Tuning for lifelong learning in HAR

Main files:

File	Description
smart_offlin.py	RLN training
smart_online.py	online learning
run_online.py	runs online learning with several RLNs and baselines
run_offline.sh	examples of RLNs training with different parametrizations

Main directories

Folder	Description
configs	experiments' setting classes
datasets	classes and functions for dataset preprocessing and benchmark generation
model	classes for model setting and meta-learning
utils	main experiments' common classes with various purposes

Execution examples

**Meta-training - encoder training - RLN with replay generation and augmentation **

python smart_offline.py --dataset=ucihar --scenario=nic --steps=2000 --plot --reset --replay --replay_strategy "herding,random,exemplar" --new_seed --runs=5 --random --model 'oml' --main_folder 'smart' --network_id 'har_1layer' --augmentation ['Jitter','Scale','Perm','TimeW','MagW']

Online learning

python smart_online.py --path /path to encoder --folder_id smart --plot --reset_weights --name continual --model oml --replay --replay_strategy random --replay_update --encoder_update --encoder_replay --encoder_strategy replay --encoder_classes all --encoder_ML --runs 5

Notes

• Parametrization available and their descriptions can be found in ../configs/class_parser* files.
• Examples of sh file to run meta-training in run_offline.sh.
• Additional scripts for plotting are available to be customized.

References

This project incorporates the following third-party implementations:

• Meta-Learning Algorithm and experiment class:
  The core meta-learning module in /model/, /oml/ are adapted from khurramjaved96.
  We extended and adapt the original implementation to support HAR data, new :