There is too much worse of inter than the effect of trained

[hbrt]

I trained one type of bold font to another type of bold font(the dataset has 500 words), and through tests I found that batch_size = 32 and epoch = 200 is the best effect, the training samples is very similarity,but when I use this model to infer outside of the 500 words, the effect is really poor, I do not know what the problem is.