DSC-180B-Hardware-Acceleration

Table of Contents

• DSC-180B-Hardware-Acceleration
  • Table of Contents
  • Environment Setup
  • Data Acquisition
  • Models
  • Simulation and Profiling
  • Config File

Environment Setup

Please refer to the Setup Guide for instructions on how to build and run the Docker image and container.

Python Environment Requirement

Version

• Python 3.9 or higher

Required Libraries

• scikit-learn
• numpy
• Pytorch (Local) - Note: We need access to a GPU for PyTorch; otherwise, set enable_GPU=False in the config.py file. The CUDA version should match the PyTorch version. Specific information can be found at PyTorch Getting Started.
• torchvision

Data

The dataset is too large to fit in GitHub. The original dataset can be downloaded from PAMAP2 Physical Activity Monitoring. Unzip the PAMAP2_Dataset you downloaded, which includes a readme.pdf with a specific description of the data. Then, copy and paste the Optional and Protocol directories to /data in our repository. Thus, in /data/Optional, there should be 5 files called subject1xx.dat; and 9 files in /data/Protocol called subject1xx.dat. Run python generate_data.py, which will generate three files: small_sample, feature, and target, which are cleaned and ready for deep learning models. The small_sample file contains the first 10 examples of data, which can be read by code faster and will be used in the simulation step.

Note: small_sample file is already in the repository. The above steps make the data generating process reproducible. outfile and target are too large and take times to generate, an alternative way to get them is to download from google drive(target and feature) ,and put them in root directory

Models

Run python xxx_train.py to train the corresponding model, print the model accuracy on the test set, and save the model in /models.

Note: Training models take a long time, you can reproduce the models by running above code. Or, you can download models here, and put them in /models. Or run this command to download large model python download_model.py, this command require gdown installed in python environment.

Simulation

In real-life situation, a trained model is saved on a device. Sensors on device store data on a hard disk. Model will read data on memory or harh disk and then input sequences of data into the model, then the model predicts the current human activity.

We use the PyTorch profiler to track the above program operations and save the profiling data in /profiling/environmentName/modelName/. The JSON file can be loaded by Chrome tracer (chrome://tracing) and generate a visualization.

Use following command to simulate all model on current environment.

python run_simulation.py --env "environment name" --gpu "True/False" --other "True/False" --trial #int

For example, current environment is "env10" with GPU, and it's third time we run it:

python run_simulation.py --env env10 --gpu True --other False --trial 3

Config

Data generating parameters and model parameters are stored in configs.py. Details are explained in the config file.