Skip to content


Repository files navigation


  • 1.6.2 Narrowed learning rate to either 1e-04 or 1e-05.
  • 1.6.1 Remove default LeakyReLU after last layer because regression should not have nonlinear output activation. Also added batch norm.


Step 0. Set up

There are three directories excluded from Git, so you should set them up. The first is the DNNs folder. You should just mkdir DNNs under the project root. Then you need to download the data folder containing all training data folders, and download scan_batteries containing all the evaluation data folders. All three directories should be under project root.

Step 1. Create Models

To use constraint satisfaction to find any number of models, you must install SWI-Prolog first in order to use swipl. After you have done this, call this command:

swipl 50 # Create 50 FCN models

2. Training instructions

Requires package h5py.

Recommended concurrency is 4 because above 4 the GPU IO is bottlenecked.

python "*" 4 # Train all created models, 4 at a time (concurrency)
python "fcn_v1.6*" 3 # Train all created models with names starting with 'fcn_v1.6', 3 at a time (concurrency)

3. Evaluation instructions

Packages needed include h5py and a monkey-patched torchaudio.

To evaluate trained models, do

python "*" # Evaluate all trained models that can be found
python "*" 1 # Evaluate 1 trained model, whichever one is found
python "fcn_v1.6.2_blah" # Evaluate model named 'fcn_v1.6.2_blah_trained', if it can be found


Convolutional Neural Networks for Ultrasound Imaging. A pipeline to create, train, and deploy models to denoise ultrasound data






No releases published


No packages published