Torch optimizer small parsing

I press torch 1.11 Version written ：

With sgd For example ：

Inside maximize = group[‘maximize’] Illustrates the maxmize Parameters can be divided group Appoint . Of course, it can also be used globally , also maxmize yes sgd Unique （ By 1.11 edition ）
Be careful , Other versions of optimizer such as adam, I didn't see it maxmize Parameters . Want to perform gradient rise , My idea is to rewrite the optimizer .

Be careful ,self.param_groups It's a list , Is defined in the base class ,sgd Can't find in the . As can be seen from the picture above ,sgd Is in accordance with the self.param_groups To each params Perform gradient optimization （ Up or down ）

Be careful sgd init Methods defaults Variable , This is for the base class init Methodical . of super, You can see b Stand a video , It's very detailed ：

Back to base class , Pay attention to defaults notes ：

The dictionary is used in the back setdefault Method to set the global default defaults For each group：