How can I use this model to train on other language such as Chinese?

[Lzc6996]

[dajiangxiaoyan]

Do you fix this?

[zoe218]

@Lzc6996 Do you fix this? :)

[lulupanPPY]

I guess only collect pics and texts will do

[yhxuyou]

if you have yourself dataset,you should deal with your data-gen problems,and modify your img size.

[kspook]

I made a successful training with Korean.

my edit is as follows; 0. labels / ord(c) value : Korean's 44032 ~ (10 ~ 251 characters) , 44 . 46
0~9 : 48 - 58 A-Z : 65 - 90 a-z :97~122

1. data_gen.py

    for c in lex:
        #assert 64 < ord(c) < num_char or 47 < ord(c) < 58
        #assert 96 < ord(c) < 123 or 47 < ord(c) < 58
        #print('c', ord(c))
        fp.write(img_path+'\t'+lex+'\t')
        fp.write(str(ord(c)))
        word.append(
            ord(c) - 44032- 43 +43 + 150 +3   if ord(c) > 44031   
            else ord(c) - 43 + 3)
        fp.write('\n')
    fp.close()
    '''        
    for c in lex:
        #assert 96 < ord(c) < num_char or 47 < ord(c) < 58
        #0assert 96 < ord(c) < 123 or 47 < ord(c) < 58
        print('c', ord(c))
        word.append(
            ord(c) - 97 + 13 if ord(c) > 96 else ord(c) - 48 + 3)
    '''
    word.append(self.EOS)
    word = np.array(word, dtype=np.int32)
    word = np.array( [self.GO] +
    [ord(c) -44032- 43 + 43 + 150 +3   if ord(c) > 44031   # Korean Character starts
    else ord(c) - 43 + 3
   

2. model.py def visualize_attention(self, filename, attentions, output_valid, ground_valid, flag_incorrect, real_len): # omit......

    with open(os.path.join(output_dir, 'word.txt'), 'w', encoding='utf8') as fword: 
        aaa=[chr(c -3-150-43+44032+43).encode('utf8','ignore')  if c-3-150-43+44032+43>44031 			
                               else chr(c-3+43).encode('utf8','ignore') for c in ground_valid]
        fword.write( ' '.join(map(bytes.decode,aaa)))
        fword.write("\n")								   
        aaa=[chr(c -3-150-43+44032+43).encode('utf8','ignore')  if c-3-150-43+44032+43>44031 			
                               else chr(c-3+43).encode('utf8','ignore') for c in output_valid]
        fword.write( ' '.join(map(bytes.decode,aaa)))	
        print('fword-o aaa', aaa)
   

3. exp_config.py LOAD_MODEL = True OLD_MODEL_VERSION = False TARGET_VOCAB_SIZE = 13500 + 256+26+3+26+10+3 # 0: PADDING, 1: GO, 2: EOS, >2: 0-9, a-z

[kspook]

The method previous produces low accuracy because I couldn't give index for the number of training characters and assigned more target_vocal_size. I give better one by assigning index for each training letters and the same number of TARGET_VOCAB_SIZE.

Please refer to https://www.github.com/kspook/Attention-OCR-1/

Detail change logs as follows

1.data_gen.py #line 10 SCRIPT_PATH = os.path.dirname(os.path.abspath(file)) DEFAULT_LABEL_FILE = os.path.join(SCRIPT_PATH, '../labels/bank_labelsS.txt')

#line 118
# 'a':97, '0':48 word = [self.GO] # , 44 . 46 0~9 :48-58 A-Z : 65-90 a-z :97~122

    try:
        fp=open('outputs.txt', 'w+')
    except:
        print('could not open file'+outputs.txt)
        quit()

    label_file = DEFAULT_LABEL_FILE
    with io.open(label_file, 'r', encoding='utf-8') as f:
       labels = f.read().splitlines()

    for c in lex:
        print('c ord(c)', c, ord(c))
        for i, l in enumerate(labels):
            #print('i  l', i , l)
            if c== l:
               n=i+3
               #print('i n : ', i, n)
               word.append(n)
            '''
            else:
               print('exit')
               exit()
            '''

    word.append(self.EOS)

    #print('ord(c), c: ', ord(c), c)
    #print('word:', word)

    return img_bw, word

2.bucketdata.py #line 24

    label_file = DEFAULT_LABEL_FILE
    with io.open(label_file, 'r', encoding='utf-8') as f:
       labels = f.read().splitlines()

    #print('self,  label, filename')
    #print(self,  label,filename)

    for i,c in enumerate(label):
        if c> 2:
            for j, l in enumerate(labels):
               if c== ord(l):
                  n=j+3
                  label[i]=n

3.model.py line 426, ~ 428 similar way such as 2.data_gen.py

    label_file = DEFAULT_LABEL_FILE
    with io.open(label_file, 'r', encoding='utf-8') as f:
       labels = f.read().splitlines()
    labels_list = enumerate(labels)
    #print(labels_list(1))
    #print(labels[1])
    with open(os.path.join(output_dir, 'word.txt'), 'w', encoding='utf8') as fword:
        gv=ground_valid
        for i, c in enumerate(ground_valid):
            #print('c ord(c)', c, ord(c))
            for j, l in enumerate(labels):
               #print('j  l', j , l)
               if (c-3)== j:
                   ##print('j n : ', j, n)
                   gv[i]=l

        print('aaa ground valid', gv, ground_valid)
        fword.write(' '.join(gv))

        ov=output_valid
        for i, c in enumerate(output_valid):
            #print('c ord(c)', c, ord(c))
            for j, l in enumerate(labels):
               #print('j  l', j , l)
               if (c-3)== j:
                   ##print('j n : ', j, n)
                   ov[i]=l

        print('aaa output valid', ov,output_valid)
        fword.write(' '.join(ov))

        #fword.write(' '.join([chr(c-13+97) if c-13+97>96 else chr(c-3+48) for c in ground_valid])+'\n')
        #fword.write(' '.join([chr(c-13+97) if c-13+97>96 else chr(c-3+48) for c in output_valid]))

[phj1012]

@kspook
how to recognize Korean? I tried your method.. but output is as below

ground valid ['두'] ['두'] output valid [] []

ground valid ['5', '7', '3', '8'] ['5', '7', '3', '8'] output valid [] []

what is problem?

[kspook]

@kspook how to recognize Korean? I tried your method.. but output is as below

ground valid ['두'] ['두'] output valid [] []

ground valid ['5', '7', '3', '8'] ['5', '7', '3', '8'] output valid [] []

what is problem?

did you tried my first suggestion? try my second suggestion.

my first suggestion tries to learn more than the number of characters that I want to learn. so result become wrong.

in my second case, the number of training is the same as I trained. So the result is the same.

In other case, you may fix if you print out or make a log the variables in the process of testing.