This repository contains the source code and data for the MobiCom'24 paper Scarf.

• Matlab: we use MATLAB_R2022a.
• A Linux machine with a GPU that is able to run ResNet is required. We use Linux 22.04 and CUDA version 12.3. We have tested NVIDIA GTX 1080, RTX 4070 and RTX 4090.
• The version of required Python packages we use are listed in requirements.txt
• Install required Python packages

chmod +x setup.sh
./setup.sh

• The Matlab scripts are located in matlab/. Each script corresponds to a figure in the paper. To execute a script, open it in Matlab and run all the sections to get the figures.
• We provide pre-processed data (permittivity, lightness, and oven-based volumetric water content) associated with the scripts.
• The scripts contains the mathematical models described in the papar.

• The Pytorch scripts are located in resnet/.
• We provide representative pre-trained ResNet models for Table 4 in the paper. Specifically, we provide a model for only using machine learning and another model for using a combination of mathematical model and machine learning. The models are located in resnet/pretrained_models/.
• We provide the testing dataset for evaluating the pre-trained models. The dataset is in resnet/testing_data/.
• To evaluate the performance of the pre-trained models, run the following:

cd resnet
python3 run_ml_model_test.py
python3 run_math_and_ml_model_test.py

• The scripts will print the correlation, R2, and MSE for the given testing dataset. NOTE: the results may look better than those in Table 4, this is because the testing dataset here is only a random sample, while Table 4 has averaged over 5 random samples.