Tensorflow2 implementation of BinauralLocalizationCNN

This repository contains a minimal, more recent tensorflow-2.x implementation of the Francl and McDermott (2022) sound localization CNN model. The original tensorflow-1.x implementation can be found in the BinauralLocalizationCNN repository.

Setup

• Clone this repository
• Download the trained model weights (checkpoint files for both implementations) and test stimuli from our Google Drive.

Python packages

Package                      Version
---------------------------- ---------
h5py                         3.9.0
keras                        2.13.1
matplotlib                   3.8.2
numpy                        1.24.3
pandas                       1.5.3
scipy                        1.10.1
tensorflow                   2.13.0
tensorflow-io                0.33.0
tqdm                         4.65.0

Contents

• DEMO_tf2_model.ipynb :
  • Minimal code example of building the model, loading trained weights, and evaluating on example sound stimuli
• compare_tf1_and_tf2_model_weights_and_outputs.ipynb :
  • Code tests comparing model weights and outputs between the tensorflow-2.x and tensorflow-1.x model implementations
• run_model_optimization.sh and run_model.py :
  • Example code for how to train or evaluate the model at scale on tfrecords files
• models/tensorflow2 and models/tensorflow1 :
  • Model directories containing config files for each of the 10 convolutional neural network architectures
  • Model directories for the tensorflow-2.x implementation also include config files with cochlear model parameters and optimization hyperparameters
  • Trained model weights can be downloaded from our Google Drive

Contact

Mark R. Saddler (msaddler@mit.edu)