Change trainable to stop_gradient in optimizer

[chenwhql]

PR types

Function optimization

PR changes

Others

Describe

Change trainable to stop_gradient in optimizer.

Now users need to optimize a input Tensor but not Parameter, case like:

# 输入
img = paddle.to_tensor(img, dtype='float32', place=paddle.CUDAPlace(0), stop_gradient=False)

#设置为不保存梯度值 自然也无法修改
for param in model.parameters():
    param.stop_gradient = True
    
# 设置优化器
optimizer = paddle.optimizer.Adam(learning_rate=0.001, parameters=img)

But our optimizer now only support to optimize Parameter. Because now our ParamBase.trainable is equal to not ParamBase.stop_gradient, so we can use stop_gradient directly in optimzier to support optimizer Tensor.

the source code are:

    @property
    def trainable(self):
        return not self.stop_gradient

    @trainable.setter
    def trainable(self, trainable):
        if isinstance(trainable, bool):
            self.stop_gradient = not trainable
        else:
            raise ValueError(
                "The type of trainable MUST be bool, but the type is ",
                type(trainable))

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[JiabinYang]

LGTM