Always missing one character in the output

[manhcuogntin4]

I used pyclstm for trainning my dataset on date but the output always missing the character 9. All the rest character is Ok. Hope for help !

[kba]

Can you provide sample images, sample output, what data you trained on, parameters used for training / recognition?

[manhcuogntin4]

@kba thank for your answer !

the output of test is 24/07/16

the output of test is 0/0/166

And here is my code:

all_imgs = [Image.open(p) for p in sorted(glob.glob("./book/*/*.png"))]
all_texts = [open(p).read().strip() for p in sorted(glob.glob("./book/*/*.gt.txt"))]
if sys.version_info <= (3,): 
 all_texts = [t.decode('utf8', "replace") for t in all_texts]

all_data = list(zip(all_imgs, all_texts))
train_data = all_data[:5500]
test_data = all_data[5500:]
len(all_data)
ocr = pyclstm.ClstmOcr()
graphemes = set(chain.from_iterable(all_texts))
#graphemes=('0','1','2','3','4','5','6','7','8','9', '/','*')
ocr.prepare_training(graphemes)

for i in range(100000):
    best_error = 1.
    img, txt = random.choice(train_data)
    out = ocr.train(img, txt)
    if not i % 10:
        aligned = ocr.aligned()
        print("Truth:   {}".format(txt))
        print("Aligned: {}".format(aligned))
        print("Output:  {}".format(out))
    if not i % 1000:
        errors = 0
        chars = 0
        for img, txt in test_data:
            out = ocr.recognize(img)
            errors += pyclstm.levenshtein(txt, out)
            chars += len(txt)
        error = errors / chars
        print ("=== Test set error after {} iterations: {:.2f}"
               .format(i, error))
        if error < best_error:
            print("=== New best error, saving model to model_date_permis.clstm")
            ocr.save("./model_date_permis.clstm")

[lomograb]

Train more = Better result