Hello and thanks again,
I however have two questions:
You set last_layer as default here, but using that would mean absolutely 0 interpretability since every class is connected to 2000 prototype similarities and this parameter "last_layer" is totally omitted in the paper? Or am I missing how that is interpretable?
Additionally, even with gumbel_scale = 10**3, for one trained model on CUB with >80% accuracy, there is no clear assignment, with only 1800/200 entries being "close" to 1.
Greetings
Hello and thanks again,
I however have two questions:
You set last_layer as default here, but using that would mean absolutely 0 interpretability since every class is connected to 2000 prototype similarities and this parameter "last_layer" is totally omitted in the paper? Or am I missing how that is interpretable?
Additionally, even with gumbel_scale = 10**3, for one trained model on CUB with >80% accuracy, there is no clear assignment, with only 1800/200 entries being "close" to 1.
Greetings