why donot use optimizer.zero_grad in each training step?

[meshiguge]

https://github.com/wabyking/TextClassificationBenchmark/blob/b5ac116c74493c28f2af2541f21385df7c73ef93/main.py#L71