Instructions for re-creating the results for EUSIPCO2023 submission
"Novel Generative Model for Acoustic EventClassification".


1. Data 
Feature data for the four data partitions are provided:
ESC8a.mat, ESC8b.mat, ESC8c.mat, ESC8d.mat. 
The data format for these files (Default data format) is 
described in the PBN Toolkit documentation.  
   https://class-specific.com/pbntk/pbntk.pdf

2. Getting an initial parameter set for each partition (all classes).
First read in PBN Toolkit documentation how to run in batch mode on a script.
Run the PBN Toolkit in batch mode using the script file run_ESC8_init0.txt
You will need to run it 4 times (once on each partition),
by editing  the file to change from ESC8a to ESC8b, ESC8c, ESC8d

3. Initialize each class-dependent PBN by running script run_ESC8_init.txt
You will need to run it 4 times (once on each partition),
by editing  the file to change from ESC8a to ESC8b, ESC8c, ESC8d

4. Final training. Train the  class-dependent PBNs by running script 
run_ESC8.txt. Each time the script is run, it does 200 epochs
on each class.  You will need to run it maybe 10 or 20 times
for each partition. To change partition, edit  the file to change from 
ESC8a to ESC8b, ESC8c, ESC8d

5. Create the output files by running script run_esc_fwd.txt.
You will need to run it 4 times (once on each partition),
by editing  the file to change from ESC8a to ESC8b, ESC8c, ESC8d

6. Get the final results by running MATLAB script combine_cnn.m   
This scipt uses the pre-existing cnn scores matrix cnn_scores.mat, 
but you can use your own CNN and create this file.  This script also 
runs MATLAB function esc_class.m to create the HMM scores.  Note, 
some required functions use HMM software that is partially
written in CMEX.