[PyTorch] Deformable Convolution示例

需要注意Deform Conv的参数初始化和学习率设置

import torch
import torch.nn as nn
import torchvision.ops as ops


class Net(nn.Module):
    def __init__(self, in_c, out_c, k=3):
        super().__init__()
        p = (k - 1) // 2
        self.split_size  = (2 * k * k, k * k)
        self.conv_offset = nn.Conv2d(in_c, 3 * k * k, k, padding=p)
        self.conv_deform = ops.DeformConv2d(in_c, out_c, k, padding=p)

        # initialize
        nn.init.constant_(self.conv_offset.weight, 0)
        nn.init.constant_(self.conv_offset.bias, 0)
        nn.init.kaiming_normal_(self.conv_deform.weight, mode='fan_out', nonlinearity='relu')

    def forward(self, x):
        offset, mask = torch.split(self.conv_offset(x), self.split_size, dim=1)
        mask = torch.sigmoid(mask)
        y = self.conv_deform(x, offset, mask)
        return y


if __name__ == '__main__':
    input = torch.rand(4, 3, 240, 320)
    net = Net(3, 7, 3)  # deform conv
    output = net(input)
    print(output.shape)

    # optimize
    lr = 0.01
    ids = list(map(id, net.conv_offset.parameters()))
    base_param = filter(lambda p: id(p) not in ids, net.parameters())
    optimizer = torch.optim.SGD(
                    [ {'params': base_param}, {'params': net.conv_offset.parameters(), 'lr': 0.1 * lr} ],
                    lr=lr,
                    momentum=0.9
                )