len(confidence vector) > len(pred_string) ?

[erikatbeof]

Hi and thanks for providing parseq!

I'm trying your example 'Pretrained Models via Torch Hub' on new unseen text images and it works very well. However I'm a bit puzzled by the confidence. The length of the confidence vector seems to be one element longer that the predicted text string. What is the meaning of the elements in the confidence vector?

[baudm]

The confidence tensor also contains the value for [E], the end-of-sequence token. I suggest you take a look at the HuggingFace demo. It shows the decoded output as well as the corresponding confidence values.

[erikatbeof]

Thanks!