The current model setup defaults to 7 classes (the standard FER-2013 target set). However, our dataset only contains 4 emotion categories:
- sadness
- happiness
- neutral
- surprise
Leaving the default set to 7 causes a dimension mismatch in the final classification head (output layer mismatch), invalid loss calculations, and unexpected behavior during evaluation metrics computation.
The current model setup defaults to 7 classes (the standard FER-2013 target set). However, our dataset only contains 4 emotion categories:
Leaving the default set to 7 causes a dimension mismatch in the final classification head (output layer mismatch), invalid loss calculations, and unexpected behavior during evaluation metrics computation.