Need to do the following things
Datasets to be use
1. ESC-50( Environmental Sound Classification)
Models to be use
The following are the references for Github code
1. [login to view URL]
2. kamalesh0406/Audio-Classification: Pytorch code for "Rethinking CNN Models for Audio Classification" ([login to view URL])
What I want to do..? is that I have to implement the methods [login to view URL] and [login to view URL] in the kemalesh0406 code. So in kamalesh0406 code I will do preprocessing to get spectrograms of all two datasets the code for making spectrograms is already given. Then in [login to view URL] file in MER code I will call these datasets and using above three mentioned model to train these datasets.
How I will train these datasets..?
First do normal training for ESC-50 dataset means train the model for 50 classes and save the results. Then divide the ESC-50 datasets into 25 and 25 classes then train the model for first 25 classes and save checkpoints code for checkpoints is already given in kamalesh0406 code. Then for incremental learning I have to train next 25 classes of ESC so for that I will use previous 25 classes checkpoints such that at testing time it give me the output of all 50 classes which is called incremental or continual learning. And same process for other datasets and models. Results of all methods should be plotted in graphs and also comparisons of normal training for 50 classes and incremental training for 25 and 25 be sure that in next incremental training of 25 classes do not use the labels of previous 25 trained classes you can use only checkpoints also called pre-trained weights.