Getting <UKN><UKN><UKN><UKN> as prediction from custom trained model

[gastodler]

Hi,

I trained a model on my Arabic data set (using SAR).
When I tried to inference on some image using my checkpoint (after 10 epochs - just to test it) I got the UKN predicition (probably unknown).

Does anyone knows what could be the reason for that?

Thank you in advance

[gaotongxiao]

Probably you haven't modified the dictionary yet to incorporate Arabic characters, leading to all of them being recognized as unknown tokens UKN. Please refer to https://mmocr.readthedocs.io/en/dev-1.x/get_started/faq.html#text-recognition for instruction and try again

[gastodler]

@gaotongxiao Thanks for your reply.

1. I have changed the dict as you said.
2. I can also see it in the cfg file that is created during the training:

dictionary = dict( type='Dictionary', dict_file= '/home/ubuntu/clipper/lib/python3.10/site-packages/mmocr/.mim/dicts/arabic_digits_symbols.txt'

3. Here is a sample of my dict:

ﺍ ﺏ ﺕ ﺙ ﺝ ﺡ ﺥ ﺩ

Can you think of some other reason?

Thank you in advance.

[gastodler]

Hi, @gaotongxiao
I was able to get Arabic letters in my inference pipeline.
It seems that the dict needs to be created using the font I created my labels (in my case Amiri-Regular).

If you have something to add about it, I'll be happy to hear.
Thanks for your help :)