why use cls not use avg

I saw the pooler_ type in the training parameters often uses cls instead of avg. 
Attempting to set avg in infocse will result in an error.
```python
pooler_output = pooler_output,view((batch_size, num_sent, pooler_output.size(-1)))
RuntimeError: shape '[16,2,768]' is invalid for input of size 29884416
```
May I ask why CLS is used instead of AVG?