DeepPulse

PyTorch implementation of pulse measurement neural networks.

Python scripts to train and test different PPG calculating neural networks. Network architectures used in research:

• PhysNet 3D (https://github.com/ZitongYu/PhysNet)
• SpatioTemporal PhysNet (https://arxiv.org/pdf/1905.02419v1.pdf)
• HR-CNN (https://github.com/radimspetlik/hr-cnn)
• Global Context Block (https://arxiv.org/pdf/1904.11492.pdf)
• Dilated convolution

Usage

To train or test network use appropriate script from listed below. PyTorch Dataloader and Dataset are prepared to work with sets containing waveform information from pulseoximeter. Scripts require lists of sequences to train and test the network. Sequences should be placed in <dataset>\<sequence_num> directory, where <dataset> should be set as root_dir = <dataset> in training and testing scripts.

Each image of the sequence should be cropped to only subjects face using utils\face_crop.py

HR-CNN

training script hr_cnn_train.py
validation script hr_cnn_valid.py
network implementation and global context block hr_cnn.py
network implementation using dilated convolution hr_cnn_dil.py

PhysNet 3D

training script PhysNet_train.py
validation script PhysNet_train.py
base network implementation PhysNet.py
network with GCBlock PhysNetGlobal.py
network implementation using dilated convolution PhysNetDil.py

Spatio-temporal PhysNet

training script PhysNet_train.py
validation script PhysNet_test.py
network implementation PhysNet_SpaTemp.py

GradCam

Script grad_cam.py contains function to visualise activation maps using GradCam method adapted to 3D convolution.

Data

Datasets used are acquired from:

• PURE (https://www.tu-ilmenau.de/en/neurob/data-sets-code/pulse/)
• VIPL-HR (https://vipl.ict.ac.cn/view_database.php?id=15)
• PFF (https://ieeexplore.ieee.org/document/8272721)

Requirements

• scipy
• torch
• opencv-python
• torchvision
• numpy
• pandas