Hello. I am a university student who’s started learning Machine Learning. I am doing ASL sign Language Recognition project by using Transfer Learning with pertained Resnet50 model (Fine-tuning without freezing any layers). There are 24 classes(since J and Z are motions) and for the datasets, I am collecting my own data - 700 images per class. I trained the model around 10 epochs and turns out it’s overfitting. When I classify a newly-taken image with the model, the output is always biased towards one class(classify nearly every image as “C” in my case). I did the data preprocessing the same the official Pytorch Transfer learning tutorial.
What should I do to my model for better generalization(to data or model)?
Data samples - https://drive.google.com/file/d/1h9T1bNzlgpEJsDtiTcDJYRN0xprbEQOB/view