exporl / shift_detection_icassp2023

Code for the models and pre-trained versions of the models of the ICASSP 2023 submission "Self-supervised enhancement of stimulus-evoked brain response data"

SELF-SUPERVISED ENHANCEMENT OF STIMULUS-EVOKED BRAIN RESPONSE DATA

This repository contains code and pre-trained model versions of the models used in the ICASSP 2023 submission "Self-Supervised Enhancement of Stimulus-Evoked Brain Response Data" by Bernd Accou, Hugo Van hamme, and Tom Francart.

Requirements

The code/models in this repository use Tensorflow version >= 2.3.0 and Python >= 3.6.0. which can be installed via conda or pip (see also the Tensorflow installation guide).

Example installation using pip

pip install tensorflow

Code

Code for all models is stored in models.py.

Pre-trained models

Pre-trained models can be found in the pretrained_models folder:

• The shift detection model contains the weights for the full shift detection model, including the enhancement module (multi-view CNN based architecture) and the comparison model.
• The subject-independent linear decoder contains the weights for the subject-independent linear decoder, used in the paper for the downstream speech envelope decoding task.

Both models are saved in tensorflow SavedModel format.

Example code for loading the models:

import tensorflow as tf

# Load the model
shift_detection_model = tf.keras.models.load_model('pretrained_models/shift_detection_model')
# Extract the multi-view CNN based enhancement module
enhancement_module = shift_detection_model.get_layer('multiview_cnn')
# Extract the simple comparison model
simple_comparison_model = shift_detection_model.get_layer('simple_comparison_model')

# Load the subject-independent linear decoder
subject_independent_decoder = tf.keras.models.load_model('pretrained_models/subject_independent_decoder')