Cost continuing reduce while accuracy is always ZERO

[wangke1003]

Hello, everyone:
I run this script with author's dataset well, but i get into into trouble like title when i train the model with my own dataset.

some pics of my dataset:




these pics are 30x500, 25 chars in each pic. i used about 260k of these to train, 65k to validate.
words in pics are randomly selected from some drug infos like this:

with open('thistxt', 'r', encoding='utf-8') as f:
# read each line into a list
all_lines = f.read().split('\n').strip()
# link each line to a string
data_str = ''.join(all_lines)
# generate word with random index
rand_word = data_str[a_rand_num, a_rand_num + word_length]
there are 196 unique chars in this txt, so my num_classes in the model is 196. is my dataset not large enough or what? i'd appreciate if anyone can help. 中文也可以