Meta-training for Human-Activity Recogniton (HAR)

These codes implements OML (Online aware Meta-Learning) and MAML-Rep approaches to HAR datasests (PAMAP2, HAPT, UCIHAR, and DSADS)

This approach is based on the paper:

Meta-Learning Representations for Continual Learning by Khurram Javed and Martha White

Paper : https://arxiv.org/abs/1905.12588

The original OML code can be found at
https://github.com/khurramjaved96/mrcl

The code provided by the authors were used as a base to support:

1. new scenarios of continual learning: nc (new classes)
2. HAR/time series datasets
3. pipeline to experiment execution, configurable to different combination of parameteres
4. set of plots, including stats per class, iteration, average, etc.

Main files:

File	Description
meta-training.py	RLN training
meta-testing.py	meta-testing
run_meta-testing.py	runs meta-testing with several encoders
meta-training_batch.py	runs batch encoders

Main directories

Folder	Description
configs	parametrization files
datasets	classes and codes for dataset and benchmark preprocessing
model	classes for model setting and meta-learning
oml	oml model setting
utils	main experiments' common classes with various purposes

Execution examples

meta-training - encoder training - RLN

python meta-training.py --dataset=pamap2 --scenario=nc --steps=25000 --plot --reset --new_seed --runs=5 --random --model 'oml'

meta-testing

python meta-testing.py --model 'maml' --path </home/> --plot --plot_file plot_meta-testing.py --runs 20 --classes_schedule 2 --new_seed --reset_weights --iid

**batch **

python meta-testing_batch.py --dataset=pamap2 --steps=50 --runs=5 --lr=0.001 --plot --scenario nc --new_seed --augmentation [Jitter,Scale,Perm,TimeW,MagW]

notes

• Parametrization description can be found in ../configs/classification/class_parser* files
• Examples of sh file to run meta-training in run_encoders_nc.sh
• Examples of sh file to run batch baseline in run_meta-testing_batch.sh