【图神经网络&动作识别】【代码阅读】ICCV 2021 CTR-GCN(Channel-wise Topology Refinement Graph Convolution Network)

火鸡大师傅

已于 2023-11-29 16:33:10 修改

阅读量562

点赞数 2

文章标签：神经网络人工智能深度学习

于 2023-11-29 11:49:36 首次发布

本文链接：https://blog.csdn.net/weixin_51031772/article/details/134673601

版权

原文

代码

模型初始化

class Model(nn.Module):
    def __init__(self, num_class=60, num_point=25, num_person=2, graph=None, graph_args=dict(), in_channels=3,
                 drop_out=0, adaptive=True):
        super(Model, self).__init__()

        if graph is None:
            raise ValueError()
        else:
            Graph = import_class(graph)
            self.graph = Graph(**graph_args)

        A = self.graph.A # 3,25,25

        self.num_class = num_class
        self.num_point = num_point
        self.data_bn = nn.BatchNorm1d(num_person * in_channels * num_point)

        base_channel = 64
        self.l1 = TCN_GCN_unit(in_channels, base_channel, A, residual=False, adaptive=adaptive)
        self.l2 = TCN_GCN_unit(base_channel, base_channel, A, adaptive=adaptive)
        self.l3 = TCN_GCN_unit(base_channel, base_channel, A, adaptive=adaptive)
        self.l4 = TCN_GCN_unit(base_channel, base_channel, A, adaptive=adaptive)
        self.l5 = TCN_GCN_unit(base_channel, base_channel*2, A, stride=2, adaptive=adaptive)
        self.l6 = TCN_GCN_unit(base_channel*2, base_channel*2, A, adaptive=adaptive)
        self.l7 = TCN_GCN_unit(base_channel*2, base_channel*2, A, adaptive=adaptive)
        self.l8 = TCN_GCN_unit(base_channel*2, base_channel*4, A, stride=2, adaptive=adaptive)
        self.l9 = TCN_GCN_unit(base_channel*4, base_channel*4, A, adaptive=adaptive)
        self.l10 = TCN_GCN_unit(base_channel*4, base_channel*4, A, adaptive=adaptive)

        self.fc = nn.Linear(base_channel*4, num_class)
        nn.init.normal_(self.fc.weight, 0, math.sqrt(2. / num_class))
        bn_init(self.data_bn, 1)
        if drop_out:
            self.drop_out = nn.Dropout(drop_out)
        else:
            self.drop_out = lambda x: x

前向传播

整个模型的流程如下图所示

def forward(self, x):
    if len(x.shape) == 3:
        N, T, VC = x.shape
        x = x.view(N, T, self.num_point, -1).permute(0, 3, 1, 2).contiguous().unsqueeze(-1)
    N, C, T, V, M = x.size()

    x = x.permute(0, 4, 3, 1, 2).contiguous().view(N, M * V * C, T)
    x = self.data_bn(x)
    x = x.view(N, M, V, C, T).permute(0, 1, 3, 4, 2).contiguous().view(N * M, C, T, V)
    x = self.l1(x)
    x = self.l2(x)
    x = self.l3(x)
    x = self.l4(x)
    x = self.l5(x)
    x = self.l6(x)
    x = self.l7(x)
    x = self.l8(x)
    x = self.l9(x)
    x = self.l10(x)

    # N*M,C,T,V
    c_new = x.size(1)
    x = x.view(N, M, c_new, -1)
    x = x.mean(3).mean(1)
    x = self.drop_out(x)

    return self.fc(x)

核心部分是TCN_GCN_unit模块，论文里的图：

代码的实现

class TCN_GCN_unit(nn.Module):
    def __init__(self, in_channels, out_channels, A, stride=1, residual=True, adaptive=True, kernel_size=5, dilations=[1,2]):
        super(TCN_GCN_unit, self).__init__()
        self.gcn1 = unit_gcn(in_channels, out_channels, A, adaptive=adaptive)
        self.tcn1 = MultiScale_TemporalConv(out_channels, out_channels, kernel_size=kernel_size, stride=stride, dilations=dilations,
                                            residual=False)
        self.relu = nn.ReLU(inplace=True)
        if not residual:
            self.residual = lambda x: 0

        elif (in_channels == out_channels) and (stride == 1):
            self.residual = lambda x: x

        else:
            self.residual = unit_tcn(in_channels, out_channels, kernel_size=1, stride=stride)

    def forward(self, x):
        y = self.relu(self.tcn1(self.gcn1(x)) + self.residual(x))
        return y

空间卷积

作者在gcn_unit中，对应每一个group，分别采用CTR-GC模块去计算，对应论文所说的Topology-non-shared methods。

class unit_gcn(nn.Module):
    def __init__(self, in_channels, out_channels, A, coff_embedding=4, adaptive=True, residual=True):
        super(unit_gcn, self).__init__()
        inter_channels = out_channels // coff_embedding
        self.inter_c = inter_channels
        self.out_c = out_channels
        self.in_c = in_channels
        self.adaptive = adaptive
        self.num_subset = A.shape[0]
        self.convs = nn.ModuleList()
        for i in range(self.num_subset):
            self.convs.append(CTRGC(in_channels, out_channels))

        if residual:
            if in_channels != out_channels:
                self.down = nn.Sequential(
                    nn.Conv2d(in_channels, out_channels, 1),
                    nn.BatchNorm2d(out_channels)
                )
            else:
                self.down = lambda x: x
        else:
            self.down = lambda x: 0
        if self.adaptive:
            self.PA = nn.Parameter(torch.from_numpy(A.astype(np.float32)))
        else:
            self.A = Variable(torch.from_numpy(A.astype(np.float32)), requires_grad=False)
        self.alpha = nn.Parameter(torch.zeros(1))
        self.bn = nn.BatchNorm2d(out_channels)
        self.soft = nn.Softmax(-2)
        self.relu = nn.ReLU(inplace=True)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                conv_init(m)
            elif isinstance(m, nn.BatchNorm2d):
                bn_init(m, 1)
        bn_init(self.bn, 1e-6)

    def forward(self, x):
        y = None
        if self.adaptive:
            A = self.PA
        else:
            A = self.A.cuda(x.get_device())
        for i in range(self.num_subset):
            z = self.convs[i](x, A[i], self.alpha)
            y = z + y if y is not None else z
        y = self.bn(y)
        y += self.down(x)
        y = self.relu(y)


        return y

这个地方用参数化的自适应邻接矩阵PA

if self.adaptive:
    self.PA = nn.Parameter(torch.from_numpy(A.astype(np.float32)))

还添加了类似于ST-GCN的注意力权重mask

self.alpha = nn.Parameter(torch.zeros(1))

核心部分是CTR-GC，这部分在论文中的结构长这样

在代码中是这个结构

对应来看是这样的

class CTRGC(nn.Module):
    def __init__(self, in_channels, out_channels, rel_reduction=8, mid_reduction=1):
        super(CTRGC, self).__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        if in_channels == 3 or in_channels == 9:
            self.rel_channels = 8
            self.mid_channels = 16
        else:
            self.rel_channels = in_channels // rel_reduction
            self.mid_channels = in_channels // mid_reduction
        self.conv1 = nn.Conv2d(self.in_channels, self.rel_channels, kernel_size=1)
        self.conv2 = nn.Conv2d(self.in_channels, self.rel_channels, kernel_size=1)
        self.conv3 = nn.Conv2d(self.in_channels, self.out_channels, kernel_size=1)
        self.conv4 = nn.Conv2d(self.rel_channels, self.out_channels, kernel_size=1)
        self.tanh = nn.Tanh()
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                conv_init(m)
            elif isinstance(m, nn.BatchNorm2d):
                bn_init(m, 1)

    def forward(self, x, A=None, alpha=1):
        x1, x2, x3 = self.conv1(x).mean(-2), self.conv2(x).mean(-2), self.conv3(x)
        x1 = self.tanh(x1.unsqueeze(-1) - x2.unsqueeze(-2))
        x1 = self.conv4(x1) * alpha + (A.unsqueeze(0).unsqueeze(0) if A is not None else 0)  # N,C,V,V
        x1 = torch.einsum('ncuv,nctv->nctu', x1, x3)
        return x1

时间卷积

时间卷积其实就是1*n的二维卷积，作者在做时间卷积的时候，分了四个层次，分别是dilation为1,2,3,4，这样可以扩大感受野，四种感受野的结果被拼到一起，然后再用了1x1的卷积和maxpool模块，同时还添加了residual模块。

class MultiScale_TemporalConv(nn.Module):
    def __init__(self,
                 in_channels,
                 out_channels,
                 kernel_size=3,
                 stride=1,
                 dilations=[1,2,3,4],
                 residual=True,
                 residual_kernel_size=1):

        super().__init__()
        assert out_channels % (len(dilations) + 2) == 0, '# out channels should be multiples of # branches'

        # Multiple branches of temporal convolution
        self.num_branches = len(dilations) + 2
        branch_channels = out_channels // self.num_branches
        if type(kernel_size) == list:
            assert len(kernel_size) == len(dilations)
        else:
            kernel_size = [kernel_size]*len(dilations)
        # Temporal Convolution branches
        self.branches = nn.ModuleList([
            nn.Sequential(
                nn.Conv2d(
                    in_channels,
                    branch_channels,
                    kernel_size=1,
                    padding=0),
                nn.BatchNorm2d(branch_channels),
                nn.ReLU(inplace=True),
                TemporalConv(
                    branch_channels,
                    branch_channels,
                    kernel_size=ks,
                    stride=stride,
                    dilation=dilation),
            )
            for ks, dilation in zip(kernel_size, dilations)
        ])

        # Additional Max & 1x1 branch
        self.branches.append(nn.Sequential(
            nn.Conv2d(in_channels, branch_channels, kernel_size=1, padding=0),
            nn.BatchNorm2d(branch_channels),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=(3,1), stride=(stride,1), padding=(1,0)),
            nn.BatchNorm2d(branch_channels)  # 为什么还要加bn
        ))

        self.branches.append(nn.Sequential(
            nn.Conv2d(in_channels, branch_channels, kernel_size=1, padding=0, stride=(stride,1)),
            nn.BatchNorm2d(branch_channels)
        ))

        # Residual connection
        if not residual:
            self.residual = lambda x: 0
        elif (in_channels == out_channels) and (stride == 1):
            self.residual = lambda x: x
        else:
            self.residual = TemporalConv(in_channels, out_channels, kernel_size=residual_kernel_size, stride=stride)

        # initialize
        self.apply(weights_init)

    def forward(self, x):
        # Input dim: (N,C,T,V)
        res = self.residual(x)
        branch_outs = []
        for tempconv in self.branches:
            out = tempconv(x)
            branch_outs.append(out)

        out = torch.cat(branch_outs, dim=1)
        out += res
        return out

具体的时间卷积实现在 TemporalConv 这个类里

class TemporalConv(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1):
        super(TemporalConv, self).__init__()
        pad = (kernel_size + (kernel_size-1) * (dilation-1) - 1) // 2
        self.conv = nn.Conv2d(
            in_channels,
            out_channels,
            kernel_size=(kernel_size, 1),
            padding=(pad, 0),
            stride=(stride, 1),
            dilation=(dilation, 1))

        self.bn = nn.BatchNorm2d(out_channels)

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        return x

在 TCN_GCN_unit 这个类里，还有一个residual模块，用了一个 unit_tcn，相当于不通过时空图卷积，直接用一个temporal的卷积来处理了

class unit_tcn(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size=9, stride=1):
        super(unit_tcn, self).__init__()
        pad = int((kernel_size - 1) / 2)
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=(kernel_size, 1), padding=(pad, 0),
                              stride=(stride, 1))

        self.bn = nn.BatchNorm2d(out_channels)
        self.relu = nn.ReLU(inplace=True)
        conv_init(self.conv)
        bn_init(self.bn, 1)

    def forward(self, x):
        x = self.bn(self.conv(x))
        return x

以上就是CTR-GCN这篇文章模型部分的全部代码，感觉这个代码看起来还是有点乱，而且里面还有一些小的trick论文里并没有提到。