MetaSense: Few-Shot Adaptation to Untrained Conditions in Deep Mobile Sensing

This is the PyTorch Implementation of MetaSense (MetaSense: Few-Shot Adaptation to Untrained Conditions in Deep Mobile Sensing, ACM SenSys '19).

[ Paper ] [ Website ] [ Video ]

Tested Environment

We tested our codes under this environment.

• OS: Ubuntu 18.04.4 LTS
• GPU: TITAN RTX 24G
• GPU Driver Version: 440.82
• CUDA Version: 10.2

Installation Guide

1. Prepare an appropriate machine. We recommend to use a GPU-installed Ubuntu machine.
2. Download or clone our repository
3. Set up a python environment using conda (see below)
4. Download datasets (see below)
5. Run the code (see below)

Python Environment

We use Conda environment. You can get conda by installing Anaconda first.

We share our python environment that contains all required python packages. Please refer to the ./metasense.yml file

You can import our environment using conda:

conda env create -f metasense.yml -n metasense

Reference: https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file

Datasets

We used four datasets:

• Individual-Condition Human Activity Recognition (ICHAR) - ours
• Individual-Condition Speech Recognition (ICSR) - ours
• Heterogeneity Human Activity Recognition (HHAR)
• Daily and Sports Activities (DSA)

We organized DSA and HHAR datasets and preprocessed data (e.g., min-max normalization) for our convenience. For details, please refer to our paper.

Download Datasets

We provide our preprocessed datasets in a csv format [here]

• ichar_minmax_scaling_all.csv: minmax-scaled (0-1) ICHAR dataset used in our experiment
• icsr_minmax_scaling_all.csv: minmax-scaled (0-1) ICSR dataset used in our experiment
• hhar_minmax_scaling_all.csv: minmax-scaled (0-1) HHAR dataset used in our experiment
• dsa_minmax_scaling_all.csv: minmax-scaled (0-1) DSA dataset used in our experiment
• ichar_original_all.csv: ICHAR dataset before scaling (for different purposes)
• icsr_original_all.csv: ICSR dataset before scaling (for different purposes)

To run our codes, you first need to download at least one of the datasets. After that, make a directory for the datasets:

$cd .               #project root
$mkdir dataset

and locate them in the ./dataset/ directory.

How to Run

Please refer to the ./run.sh file. It contains running commands for all methods, datasets, and number of shots.

If you want to verify your installation, you can do simply with the following command:

$python main.py --dataset metasense_activity_scaled --method MetaSense --tgt PH0007-jskim --epoch 200 --log_suffix run_test_5shot_0.1  --src rest --train 1 --model MetaSense_Activity_model --nshot 5 --lr 0.1 

CAUTION: For the TrC method, it requires a pre-trained Src model. Please make sure you have trained a Src model with the same --log_suffix argument before training a TrC instance.

Misc.

1.TensorBoard

We support visualization (confusion matrix, train/valid/test loss and accuracy, etc.) and logging with TensorBoard.

Running train examples will automatically generate events files for TensorBoard.

You can check them by learning, for instance,

$tensorboard --logdir log/ --port 5000 --bind_all

2. Checkpoint

The code will save the validation best model (i.e., checkpoint) in the logging directory.

For instance, LOG_DIR/cp/cp_best.pth.tar

3. Result

After the training, it will generate a result.txt file that provides the validation best accuracy and the test accuracy.

For instance, LOG_DIR/result.txt

Note that the result of PN, MAML, and MetaSense includes accuracy from given 1 to 10 shots in a single result.txt file, while the result of other methods includes only one accuracy for a specified shot through the argument.

Citation

@inproceedings{metasense,
    author = {Gong, Taesik and Kim, Yeonsu and Shin, Jinwoo and Lee, Sung-Ju},
    title = {MetaSense: Few-Shot Adaptation to Untrained Conditions in Deep Mobile Sensing},
    year = {2019},
    isbn = {9781450369503},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3356250.3360020},
    doi = {10.1145/3356250.3360020},
    booktitle = {Proceedings of the 17th Conference on Embedded Networked Sensor Systems},
    pages = {110–123},
    numpages = {14},
    keywords = {few-shot learning, deep learning, meta learning, human activity recognition, mobile sensing},
    location = {New York, New York},
    series = {SenSys ’19}
}