Custom dataset generation #41
Description
Hello @sharathadavanne , i would like to report of an issue of using pyroomacoustics in order to generate new batch of dataset. It seems that whatever method i choose:
- reordering microphone placement
- increasing/decreasing number of detected class
- increasing/decreasing number of channels
- using reverberant or anechoic room
- increasing model error weights
- changing CNN, RNN, FNN size
- ect
I'm unable to reproduce the desired dataset without having the model to have a consistent ~0.95 DOA error rate. I would be very grateful if anyone can help me generating the correct dataset
Here are some pyroomacoustics wrapper i've made for this model
https://github.com/izzan-alfian/SELD-quadcopter/tree/main/dataset_generator
I also have made several attemps on reproducing this dataset on matlab as well without any success