gaozhihan

@XingyuXie i using the optimizer visualization for verification, it feels that the adan algorithm is less robust than other algorithms. and my tf implementation is [here](https://github.com/cpuimage/keras-optimizer/blob/main/optimizer/Adan.py), and [visualization](https://github.com/cpuimage/keras-optimizer/blob/main/visualization/visualization.py)

> I test in 32k sample, but it is wrong. Maybe the code only work in 8k samplerate. 好的，回头我看下这个问题， 最近忙得没什么时间~

> 可以在[此处](https://github.com/tensorflow/tpu/issues/890)找到 Tensorflow 的 TPU 存储库中提出的完整错误描述。 Perhaps, you can try replacing the "math_ops.euclidean_norm" with the following: ``` def euclidean_norm(x, axis, keepdims=False): return math_ops.reduce_sum(x ** 2, axis=axis, keepdims=keepdims) ** 0.5 ```