I'm working on audio classification task. I have dataset with around 20k samples and I'm using some model, let's call it x from torchvision models hub(trained on Imagenet)
I'm pretty fast (around 10-15) epochs reaching high performance on my metrics for the training set and the loss starting to saturate.
When the train loss stop decreasing the validation metrics also stop improving.
- Should I apply augmentation on this scenario? Even with augmentation the training stats reaching saturation pretty fast. Thus it mean I should build my own small nn network ?
Thanks in advance.