Another language

[ChaoRuk]

if i would like to train model that no english (another language). what do i have to do ?

[gaotongxiao]

The answer depends on MMOCR's version. If you are new to MMOCR, I strongly recommend you start with MMOCR 1.0 since it has significant changes, improvements and more detailed documentation, compared to 0.x.

MMOCR 0.x

In MMOCR 0.x, you need to prepare a dictionary file where characters are listed line by line (example dict). Then update lable_convertor's config as something similar to

mmocr/configs/textrecog/sar/sar_r31_parallel_decoder_chinese.py

Lines 6 to 11 in 5fc9204

	dict_file = 'data/chineseocr/labels/dict_printed_chinese_english_digits.txt'
	label_convertor = dict(
	type='AttnConvertor',
	dict_file=dict_file,
	with_unknown=True,
	max_seq_len=max_seq_len)

Sometimes you may also need to revise the model architecture to ensure the output channel of the final projection layer matches the dictionary size. This variable is usually named num_chars.

MMOCR 1.0

Prepare a dictionary file and replace the Dictionary's dict_file field with its path (example). The model architecture will be automatically updated without any manual effort.