Why not the training accuracy and test accuracy rise?

[LouZhiX]

我是直接使用的此代码，其中只是更改了数据集的大小，设定的数据集为30000张训练集，6000张测试集，每幅验证码由数字和大小写字符构成。但是实际在训练时发现train accuracy 和test accuracy都上不去，而且loss达到4.12左右就不再下降，请问这是什么原因？

[OSchwark]

interesting, label_size=62 means you have A-Za-z0-9. I also tried that (with different captcha images than in example) and also have only ~2-3% accuracy. Removing all letters leads to good values.
Maybe script needs to be adjusted to reflect letters?