mismatch dimension between HWDB checkpoint and charset

[zjl123001]

i am confused that the dimension in finetuned checkpoint HWDB is 2704 but the charset in /data/HWDB_v1/charset.pkl is 7356 could you pls help me about this problem? thanks

[raphael-baena]

Hello,

Thanks for bringing this up! Here are a few clarifications regarding the pretrained and finetuned checkpoints:

• The pretrained checkpoint has a charset of 7,356 characters and was trained using synthetic examples generated from the CASIA v1 dataset.
• The finetuned checkpoint has a reduced charset of 2,704 characters.

I noticed that I initially forgot to include the charset from the HWDB dataset (CASIA v2). I've now added the missing charset in the corresponding folder. HWDB.py (CASIA v2) were already designed to look for this charset but i was missing:

self.data = pickle.load(open(os.path.join(datasets_path, 'HWDB', 'data.pkl'), "rb"))
self.charset = self.data['charset']

Please let me know if you need further assistance.