Shift in the results

[lelapin123]

I tried to use this code:
https://github.com/tulasiram58827/ocr_tflite/blob/main/colabs/captcha_ocr_tflite.ipynb

I had to convert it to py to make it work.
Now it is able to guess letters, but they get scrambled.

To fix this i change this line:
characters = set(char for label in labels for char in label)
with this line:
characters = (sorted(set(char for label in labels for char in label)))

Now, when there is a 3 on the picture, 2 is detected,
4 is 3
5 is 4
8 is 7
b is 8,

There is always a shift of one, but beside converting the string from one to the other, i am not able to make the code work.

What worked was to change: num_oov_indices=0
to num_oov_indices=1

But then, i am not able to train the model anymore.
I wonder is there is not a problem in the background.

It seems like these lines:
preds = prediction_model.predict(batch_images)
pred_texts = decode_batch_predictions(preds)

Always return a result that is shifted from the real answer.