ColloSSL: Collaborative Self-Supervised Learning for Human Activity Recognition

July 2, 2022 ยท View on GitHub

ColloSSL (pronounced colossal) is a technique for collaborative self-supervised contrastive learning among a group of devices by utlizing the time-synchronocity of their data.

gsl_architecture_revised-1

This repo is a Tensorflow implementation of the ColloSSL paper.

@article{jain2022collossl,
  title={ColloSSL: Collaborative Self-Supervised Learning for Human Activity Recognition},
  author={Jain, Yash and Tang, Chi Ian and Min, Chulhong and Kawsar, Fahim and Mathur, Akhil},
  journal={Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies},
  volume={6},
  number={1},
  pages={1--28},
  year={2022},
  publisher={ACM New York, NY, USA}
}

Environment setup

The code works with the latest docker image of tensorflow. Run bash run_container.sh to start a docker container, make changes in the script according to your filesystem. Subsequently, run pip install -r requirements.txt to install extra dependencies in the docker container.

Dataset files

You can find the pre-processed dataset files here

Directory Structure

We create a directory for each train_device as follows:

args.working_directory / args.train_device / args.exp_name / args.training_mode

e.g., /mnt/data/gsl/runs/thigh/my_exp/single/

Inside each directory, there are three subdirs:

  • /models
  • /logs
  • /results

Each hyperparam runs is assigned a run_name and all logs, models and result files share the same name.

Running instruction

The scripts/ directory has example scripts for each type of experiment. Before running any script, you need to change the working_directory and dataset_path in the scripts.

  • collossl_single_run.sh - Single ColloSSL run for a particular train_device and eval_device with a fixed hyperparameter settings
  • collossl.sh - Multiple ColloSSL runs for all devices in the dataset. Runs happen across multiple-gpus and are automatically scheduled one after other. Please refer to hparam_tuning_mp.py to change/add any hyperparameters.
  • supervised.sh - Supervised baseline for each device in the dataset. Runs happen across multiple-gpus and are automatically scheduled one after other. Please refer to hparam_tuning_mp.py to change/add any hyperparameters.
  • supervised_all_devices.sh - Supervised baseline using all device data during training.
  • other_ssl_baselines.sh - Running configurations for other baselines in the paper. Runs happen across multiple-gpus and are automatically scheduled one after other. Please refer to hparam_tuning_mp_ssl_baseline.py to change/add any hyperparameters.

The results of all the runs are stored in args.working_directory / args.train_device / args.exp_name / args.training_mode/ logs /result_summary.csv file which can later be plotted using plot_results.py.

Scripts also generate plots of completed runs in scripts/args.exp_name directory. Example plots are shown in /results directory

Steps to retrieve a certain model (manually)

  1. Open Tensorboard with --logdir=<args.working_directory/args.train_device/args.exp_name/args.training_mode/logs/hparam_tuning_*>
  2. Go to the Scalars tab and pick the run of your choice (e.g., the one with the best F1 score). Copy its run_id.
  3. (Optionally) You can now go to the HParams Table view and find the hyperparams corresponding to it. Unfortunately, the ID in Hparams is system generated and not the same as run_id. We will have to match the runs based on other metrics, e.g., F-1 score.
  4. The model and result file for the selected run should be in args.working_directory/args.train_device/args.exp_name/args.training_mode/models/run_id.hdf5 and args.working_directory/args.train_device/args.exp_name/args.training_mode/results/run_id.txt

Models

License