卷积模块重参数化

最新推荐文章于 2025-12-31 15:46:29 发布

原创

最新推荐文章于 2025-12-31 15:46:29 发布 · 314 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#深度学习 #pytorch #神经网络

卷积模块的重参数化实现。
class SeqConv3x3(nn.Module):
def init(self, seq_type, inp_planes, out_planes, depth_multiplier):
super(SeqConv3x3, self).init()

    self.type = seq_type
    self.inp_planes = inp_planes
    self.out_planes = out_planes

    if self.type == 'conv1x1-conv3x3':
        self.mid_planes = int(out_planes * depth_multiplier)
        conv0 = torch.nn.Conv2d(self.inp_planes, self.mid_planes, kernel_size=1, padding=0)
        self.k0 = conv0.weight
        self.b0 = conv0.bias

        conv1 = torch.nn.Conv2d(self.mid_planes, self.out_planes, kernel_size=3)
        self.k1 = conv1.weight
        self.b1 = conv1.bias

    elif self.type == 'conv1x1-sobelx':
        conv0 = torch.nn.Conv2d(self.inp_planes, self.out_planes, kernel_size=1, padding=0)
        self.k0 = conv0.weight
        self.b0 = conv0.bias

        # init scale & bias
        scale = torch.randn(size=(self.out_planes, 1, 1, 1)) * 1e-3
        self.scale = nn.Parameter(scale)
        # bias = 0.0
        # bias = [bias for c in range(self.out_planes)]
        # bias = torch.FloatTensor(bias)
        bias = torch.randn(self.out_planes) * 1e-3
        bias = torch.reshape(bias, (self.out_planes,))
        self.bias = nn.Parameter(bias)
        # init mask
        self.mask = torch.zeros((self.out_planes, 1, 3, 3), dtype=torch.float32)
        for i in range(self.out_planes):
            self.mask[i, 0, 0, 0] = 1.0
            self.mask[i, 0, 1, 0] = 2.0
            self.mask[i, 0, 2, 0] = 1.0
            self.mask[i, 0, 0, 2] = -1.0
            self.mask[i, 0, 1, 2] = -2.0
            self.mask[i, 0, 2, 2] = -1.0
        self.mask = nn.Parameter(data=self.mask, requires_grad=False)

    elif self.type == 'conv1x1-sobely':
        conv0 = torch.nn.Conv2d(self.inp_planes, self.out_planes, kernel_size=1, padding=0)
        sel