why decoding starts from 3rd position? about crnn-with-stn HOT 3 OPEN

Just noticed the same comment in your code :)

from crnn-with-stn.

If I start decoding with zero I indeed receive "garbage" sometimes (usually a duplicate of first character), but if the same slicing is is in the cost function then it's not suprising

from crnn-with-stn.

Hi,

I wonder what is the theoretical basis for starting decoding from 3rd position. I'm referring to this line:
ctc_decode = bknd.ctc_decode(y_pred[:, 2:, :], input_length=np.ones(shape[0])*shape[1])[0][0]

In image_ocr.py example on keras github there's a comment:

# the 2 is critical here since the first couple outputs of the RNN
# tend to be garbage:

But why? And why everyone is using 2 regardless of dataset, image width and text length?

To be honest, I don't know the specific reason for this, either.

from crnn-with-stn.