search

da03 / Attention-OCR

Visual Attention based OCR
MIT License
1.11k stars 362 forks source link

a question #19

Closed Yang507 closed 7 years ago

Yang507 commented 7 years ago

when I train my data, the accuracy is one substract the loss value ?

da03 commented 7 years ago

Sorry that I might not understand, if we train the model, the log file will output the perplexity values. Accuray will be evaluated when we are testing the model.

Yang507 commented 7 years ago

In fact I'm not understand the perplexity value, i don't know the meaning of it.

da03 commented 7 years ago

I see. Perplexity is a common measurement for evaluating language model. It represents the average uncertainty of predicting a symbol. For example, if we have a perplexity of 2, that means the model needs to guess on average 2 times for each character to get the correct one. Therefore, the ideal perplexity shall be 1, which means the model guess once, and gets the correct prediction. I would recommend reading wikipedia for a reference: https://en.wikipedia.org/wiki/Perplexity.

Yang507 commented 7 years ago

Thank you for answering my doubts!

da03 commented 7 years ago

You are welcome!