iranroman/deepwavetorch

DeepWave: A Recurrent Neural-Network for Real-Time Acoustic Imaging (PyTorch implementation)

DeepWave: A Recurrent Neural-Network for Real-Time Acoustic Imaging (PyTorch)

This repository contains a PyTorch implementation of the DeepWave model originally published at NeurIPS 2019

| paper | original code |

Get started with inference

Main building blocks:

• SphericalChebConv: Spherical Chebyshev graph convolutions
• BackProjLayer: project correlation matrix into image form (intensity map form)
• ReTanh: Rectified hyperbolic tangent action function
• DeepWave: the actual model architecture

DeepWave (PyTorch) architecture:

Deepwave: input=S, trainable={mu, D, tau}
    y <- BackProjLayer(S) 
conv4 <- SphericalChebConv(I_init) + y
conv4 <- ReTanh(conv4)
conv3 <- SphericalChebConv(conv4) + y
conv3 <- ReTanh(conv3)
conv2 <- SphericalChebConv(conv3) + y
conv2 <- ReTanh(conv2)
conv1 <- SphericalChebConv(conv2) + y
conv1 <- ReTanh(conv1)
conv0 <- SphericalChebConv(conv1) + y
I_out <- ReTanh(conv0)

Many of the operations used in this implementation were borrowed from the repository deepsphere-pytorch.

Installation

First time setup

Create a Python virtual environment

python3 -m venv /path/to/your_new_venv

Start your_new_venv

source /path/to/<your_new_venv>/bin/activate

• Clone DeepWaveTorch (this repo!).

git clone [email protected]:adrianSRoman/DeepWaveTorch.git 

• Initialize submodules

git submodule init

• Start ImoT_tools: plotting library used to nicely visualize DeepWave's output

cd ImoT_tools
pip install -r requirements.txt
python3 setup.py develop

• Start DeepWave: original DeepWave implementation. Used for benchmarking against the PyTorch implementation. Data loaders are also re-used from the original implementation.

cd DeepWave
python3 setup.py develop

• Start DeepWaveTorch: new DeepWave PyTorch implementation.

cd DeepWaveTorch
pip install -r requirements.txt
python3 setup.py develop

Executing example notebooks

• Start your_new_venv

source /path/to/<your_new_venv>/bin/activate

• Create a Jupyter Kernel to contain your required packages (first time setup only)

pip install ipykernel
ipython kernel install --user --name=<your_new_kernel_name>

• Start Jupyter

jupyter notebook

• Select a given notebook you want to work with
• Select your_new_kernel_name under: Kernel > Change kernel > your_new_kernel_name

Qualitative and quantitative comparison against the original DeepWave implementation

Qualitatively, the implementation from this repository generates the same intensity fields as the original DeepWave implementation.

Inferred intensity field for a single frequency band: DeepWave original Vs. DeepWave PyTorch

YouTube presentation

License

Shield:

This work is licensed under a Creative Commons Attribution 4.0 International License.