Human Activity Recognition (HAR) has been developed for a long time. Many highly accurate models were proposed. However, one major obstacle towards the real-life smart home is the ability to adapt to change.
This project aims to develop a continual learning model that is able to learn a new task but also retain previously learned tasks. The key technique used here is generative replay. Two agents are working together: one generative model is to mimic all trained lesson while one fully-connected neural network is to gradually lean to capture structured knowledge.
See technical detail here
It is an part of CS5099 Dissertation at University of St. Andrews.
- Prepare python environment
python3 -m venv env - Enter to the env
source env/bin/activate - Install library
pip install -r requirements.txt
Run python run_main.py --results-dir="YOUR RESULT DIR" --data-dir="DATASET"
DATASET must be one of these options ["casas", "housea", "pamap", "dsads"]
Other available arguments can be seen by running python run_main.py --help
You could run run_*.py for a different experiment (The same arguments can be applyied in any experiments). You need to run visdom if you want to run run_plot.py.
- Run
python3 -m visdom.serverin one terminal - Open browser and then enter url given by the visdom server
- Open anohter terminal then run your model with
--visdom(only work withrun_plot.py) - See the interactive report
- implement XdG
- consider diversity as well as accuracy
- self-verifying => select samples with probability rather than yes/no result
- GAN-ralated
- implement Minibatch or Feature mathcing to avoid mode collapse
- try Unrolled loss https://arxiv.org/abs/1611.02163
- label smoothing https://arxiv.org/abs/1606.03498
- TTUR Two Time-Scale Update Rule https://arxiv.org/abs/1706.08500
- different learning rates for the generator and discriminator
- 0.0004 for the discriminator and 0.0001 for the generator
- automated parameter selection
- fANOVA
- try other GANs
- try hierarchy model
- experiments with task evolution
- experiments with noise; robustness to noise
Pakawat Nakwijit; An ordinary programmer who would like to share and challange himself. It is a part of my 2018 tasks to open source every projects in my old treasure chest with some good documentation.
This project is licensed under the terms of the MIT license.