Fix optimizer dtype

[zhiqiu]

PR types

Bug fixes

PR changes

APIs

Describe

The dtype of Optimizer should be inferred by given parameters, which will be used in learning rate and other vars in optimizer.

For example,

class MyLayer(paddle.nn.Layer):
    def __init__(self, dtype):
        super(MyLayer, self).__init__()
        self._w = self.create_parameter([2, 3], dtype='float64')
        self._b = self.create_parameter([2, 3], dtype='float64')

    def forward(self, x):
        return x * self._w + self._b
            model = MyLayer(dtype)

x = paddle.rand([10, 2, 3], dtype=dtype)
loss = model(x)
adam = paddle.optimizer.Adam(parameters=model.parameters())
loss.backward()
adam.step()

[paddle-bot-old]

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

[chenwhql]

LGTM