I have a CRNN based model that I am training with audio data. My model works well with AudioSet, but now I am training it with a private dataset that has only 87 samples (about 10 samples per class) and variable input length. The audio files’ length varies between 30 seconds and 10 minutes. Since I got a collate_fn error during stacking of the input in the batch, I am training with batches of size 1. However I see that the training loss drops in the first epoch and then stays constant or slightly goes up and down in the next epochs. I trained it for 100 epochs, but the trend remains the same. The validation loss is also staying constant to 0.6. I am new in deep learning, so I am not quite clear why is this happening and how can I address this issue.