Why the code for Learner is so complicated?

[fikry102]

init and forward methds for Learner is so complicated

class Learner(nn.Module):
    """

    """

    def __init__(self, config, imgc, imgsz):
        """

        :param config: network config file, type:list of (string, list)
        :param imgc: 1 or 3
        :param imgsz:  28 or 84
        """
        super(Learner, self).__init__()


        self.config = config

        # this dict contains all tensors needed to be optimized
        self.vars = nn.ParameterList()
        # running_mean and running_var
        self.vars_bn = nn.ParameterList()

        for i, (name, param) in enumerate(self.config):
            if name is 'conv2d':
                # [ch_out, ch_in, kernelsz, kernelsz]
                w = nn.Parameter(torch.ones(*param[:4]))
                # gain=1 according to cbfin's implementation
                torch.nn.init.kaiming_normal_(w)
                self.vars.append(w)
                # [ch_out]
                self.vars.append(nn.Parameter(torch.zeros(param[0])))

            elif name is 'convt2d':
                # [ch_in, ch_out, kernelsz, kernelsz, stride, padding]
                w = nn.Parameter(torch.ones(*param[:4]))
                # gain=1 according to cbfin's implementation
                torch.nn.init.kaiming_normal_(w)
                self.vars.append(w)
                # [ch_in, ch_out]
                self.vars.append(nn.Parameter(torch.zeros(param[1])))

            elif name is 'linear':
                # [ch_out, ch_in]
                w = nn.Parameter(torch.ones(*param))
                # gain=1 according to cbfinn's implementation
                torch.nn.init.kaiming_normal_(w)
                self.vars.append(w)
                # [ch_out]
                self.vars.append(nn.Parameter(torch.zeros(param[0])))

            elif name is 'bn':
                # [ch_out]
                w = nn.Parameter(torch.ones(param[0]))
                self.vars.append(w)
                # [ch_out]
                self.vars.append(nn.Parameter(torch.zeros(param[0])))

                # must set requires_grad=False
                running_mean = nn.Parameter(torch.zeros(param[0]), requires_grad=False)
                running_var = nn.Parameter(torch.ones(param[0]), requires_grad=False)
                self.vars_bn.extend([running_mean, running_var])


            elif name in ['tanh', 'relu', 'upsample', 'avg_pool2d', 'max_pool2d',
                          'flatten', 'reshape', 'leakyrelu', 'sigmoid']:
                continue
            else:
                raise NotImplementedError

Why can't we construct a learner using the common way?
Are there any extra considerations?

def conv_block(in_channels, out_channels):
    return nn.Sequential(
        nn.Conv2d(in_channels, out_channels, 3, padding=1),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(),
        nn.MaxPool2d(2)
    )


@register('convnet4')
class ConvNet4(nn.Module):

    def __init__(self, x_dim=3, hid_dim=64, z_dim=64):
        super().__init__()
        self.encoder = nn.Sequential(
            conv_block(x_dim, hid_dim),
            conv_block(hid_dim, hid_dim),
            conv_block(hid_dim, hid_dim),
            conv_block(hid_dim, z_dim),
        )
        self.out_dim = 1600

    def forward(self, x):
        x = self.encoder(x)
        return x.view(x.shape[0], -1)

[d3ac]

Because we need to find the second-order derivative. Use nn.sequential can't achieve fast_weight and slow_weight, you can't get second-order derivative due to computation graph