nnrg / contextskilldrift

A deep neural network architecture for gas classification using metal oxide-based sensors designed to be robust to sensor drift

A context-based gas sensor classifier

Please see the paper here: https://arxiv.org/abs/2003.07292

The context-based gas classifier processes a sequence of labeled odor samples in order to form a representation of the current sensor context, containing information pertaining to sensor drift. This context vector is used to inform the prediction of the odor label.

Usage

This repository includes three useful files:

• The main file that starts the training of the network with the GA (main_odor_GA.py)
• The neural network module that contains the class for the modified Context-Odor Model (mCOM, neural_net.py)
• The module that reads the batches of data (read_batches.py)

Quick set-up for the GA run

data_available = False by default. If the training data was loaded before, you can change it to True.

trainingTASKs = [1, 2, 3, 4, 5]. This is the list of tasks (batches) that the mCOM will be trained on. It has to be always more than 1 due to the Context module. Context module's, i.e., LSTM cell, memory is reset after each batch. In order to compare mCOM with the Baseline Model, Line-78 (net.resetContextMemory()) can be commented out.