Created by: txyugood

在复现论文《Representation Flow for Action Recognition》的光流层的过程中，在动态图中迭代循环20次去计算光流，在循环中会调用如下函数，会导致显存占用暴涨10G，导致 batch_size 十分小。请问这是由于什么原因导致的，怎样才能解决，已经卡在这好几天了，望回复谢谢！

    self.f_div_conv = Conv2D(
            num_channels=channels,
            num_filters=channels,
            filter_size=(1, 2),
            groups=self.channels,
            param_attr=ParamAttr(initializer=NumpyArrayInitializer(self.div)),
            bias_attr=False
        )
    
    def divergence(self, x, y):
        tx = fluid.layers.pad(x[:, :, :, :-1], [0, 0, 0, 0, 0, 0, 1, 0])
        ty = fluid.layers.pad(y[:, :, :-1, :], [0, 0, 0, 0, 1, 0, 0, 0])

        tx = fluid.layers.pad(tx, [0, 0, 0, 0, 0, 0, 0, 1])
        ty = fluid.layers.pad(ty, [0, 0, 0, 0, 0, 1, 0, 0])
        grad_x = self.f_div_conv(tx)
        grad_y = self.f_div2_conv(ty)

        #pytorch 实现方法
        # grad_x = F.conv2d(F.pad(tx, (0, 0, 0, 1)), self.div)  # , groups=self.channels)
        # grad_y = F.conv2d(F.pad(ty, (0, 0, 0, 1)), self.div2)  # , groups=self.channels)

        return grad_x + grad_y



self.n_iter = 20
for i in range(self.n_iter):
    some code ....
    u1 = v1 + self.t * self.divergence(p11, p12)
    u2 = v2 + self.t * self.divergence(p21, p22)

    some code ..
    p11 = (p11 + taut * u1x) / (1. + taut * fluid.layers.sqrt(u1x ** 2 + u1y ** 2 + 1e-12))
    p12 = (p12 + taut * u1y) / (1. + taut * fluid.layers.sqrt(u1x ** 2 + u1y ** 2 + 1e-12))
    p21 = (p21 + taut * u2x) / (1. + taut * fluid.layers.sqrt(u2x ** 2 + u2y ** 2 + 1e-12))
    p22 = (p22 + taut * u2y) / (1. + taut * fluid.layers.sqrt(u2x ** 2 + u2y ** 2 + 1e-12))
return u1,u2