[Dy2Stat]Support convert sublayers in Sequential Container

[Aurelius84]

PR types

New features

PR changes

Others

Describe

1. PR功能

支持 Paddle.nn.Sequential中的sublayer的代码转换。

之前对于paddle.nn.Sequential容器内的layers，由于在convert_call_func时，识别为Paddle内部API，因此会跳过转换。此Layer为一个容器，内部常包含很多用户定义的Layer，因此需要进行转换。

当Sequential包含了Paddle自身的Layer，由于convert_call_func会递归判断，因此不会对Paddle内部的Layer误处理。

2. 代码样例

class SequentialNet(paddle.nn.Layer):
    def __init__(self, sub_layer, in_channel, out_channel):
        super(SequentialNet, self).__init__()
        self.layer = paddle.nn.Sequential(
            ('l1', paddle.nn.Linear(in_channel, in_channel)),  
            ('l2', paddle.nn.Linear(in_channel, out_channel)),
            ('l3', sub_layer(out_channel)))  # 仅会对此sub layer进行动转静

    def forward(self, x):
        out = self.layer(x)
        return out

class BufferLayers(paddle.nn.Layer):
    def __init__(self, out_channel):
        super(BufferLayers, self).__init__()
        self.out_channel = out_channel

    def forward(self, x):
        mean = paddle.mean(x)
        if mean < 0.:
            x = x * self._mask()

        out = x - mean
        return out

    def _mask(self):
        return paddle.to_tensor(np.zeros([self.out_channel], 'float32'))

[paddle-bot-old]

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

[liym27]

LGTM

[liym27]

LGTM