Thanks for your ownsome work of Dynamic Low-Resolution Distillation for Cost-Efficient End-to-End Text Spotting.
I have a question about the implementation about beam search. It seems that the function beam_decode receive the softmax results on teacher logits, which means it is a probability not the log probability. However, in this and this line, it seems that you have incorrectly treat the probability as log probability (although the name is logp...). I think multiply operator should be used instead of the add operator.
Hi authors,
Thanks for your ownsome work of Dynamic Low-Resolution Distillation for Cost-Efficient End-to-End Text Spotting.
I have a question about the implementation about beam search. It seems that the function
beam_decodereceive the softmax results on teacher logits, which means it is a probability not the log probability. However, in this and this line, it seems that you have incorrectly treat the probability as log probability (although the name is logp...). I think multiply operator should be used instead of the add operator.Best.