BGNet代码解读

理论准备

Bilateral Grid Learning for Stereo Matching Networks论文内容解读请参看“论文阅读笔记：Bilateral Grid Learning for Stereo Matching Networks”

特征提取模块

class feature_extraction(nn.Module):
    def __init__(self):
        super(feature_extraction, self).__init__()

        self.inplanes = 32
#         self.firstconv = convbn_relu(1, 32, 7, 2, 3, 1)
        self.firstconv = nn.Sequential(convbn_relu(1, 32, 3, 2, 1, 1), #1/2
                               convbn_relu(32, 32, 3, 1, 1, 1),
                               convbn_relu(32, 32, 3, 1, 1, 1))
        self.layer1 = self._make_layer(BasicBlock, 32, 1, 1, 1, 1)
        self.layer2 = self._make_layer(BasicBlock, 64, 1, 2, 1, 1) #1/4
        self.layer3 = self._make_layer(BasicBlock, 128, 1, 2, 1, 1) #1/8
        self.layer4 = self._make_layer(BasicBlock, 128, 1, 1, 1, 1)
        self.reduce = convbn_relu(128, 32, 3, 1, 1, 1)
        
        self.conv1a = BasicConv(32, 48, kernel_size=3, stride=2, padding=1)
        self.conv2a = BasicConv(48, 64, kernel_size=3, stride=2, padding=1)
        self.conv3a = BasicConv(64, 96, kernel_size=3, stride=2, padding=1)
#         self.conv4a = BasicConv(96, 128, kernel_size=3, stride=2, padding=1)

#         self.deconv4a = Conv2x(128, 96, deconv=True)
        self.deconv3a = Conv2x(96, 64, deconv=True)
        self.deconv2a = Conv2x(64, 48, deconv=True)
        self.deconv1a = Conv2x(48, 32, deconv=True)

        self.conv1b = Conv2x(32, 48)
        self.conv2b = Conv2x(48, 64)
        self.conv3b = Conv2x(64, 96)
#         self.conv4b = Conv2x(96, 128)

#         self.deconv4b = Conv2x(128, 96, deconv=True)
        self.deconv3b = Conv2x(96, 64, deconv=True)
        self.deconv2b = Conv2x(64, 48, deconv=True)
        self.deconv1b = Conv2x(48, 32, deconv=True)

    def _make_layer(self, block, planes, blocks, stride, pad, dilation):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes * block.expansion),)

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample, pad, dilation))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes, 1, None, pad, dilation))

        return nn.Sequential(*layers)
    def forward(self, x):
        #1/2
        x = self.firstconv(x)
        x = self.layer1(x)
        conv0a = x
        x = self.layer2(x) #1/4
        conv1a = x
        x = self.layer3(x) #1/8 * 128
        feat0 = x
        x = self.layer4(x) #1/8 * 128
        feat1 = x
        x = self.reduce(x) #1/8 * 32
        feat2 = x
        rem0 = x
        #1/2 * 1/2 * 48
        x = self.conv1a(x)
        rem1 = x
        #1/4 * 1/4 * 64
        x = self.conv2a(x)
        rem2 = x
        #1/8 * 1/8 * 96
        x = self.conv3a(x)
        rem3 = x
        #1/16 * 1/16 * 128
#         x = self.conv4a(x)
#         rem4 = x
#         x = self.deconv4a(x, rem3)
#         rem3 = x

        x = self.deconv3a(x, rem2)
        rem2 = x
        x = self.deconv2a(x, rem1)
        rem1 = x
        x = self.deconv1a(x, rem0)
        feat3 = x
        rem0 = x
        #1/2
        x = self.conv1b(x, rem1)
        rem1 = x
        x = self.conv2b(x, rem2)
        rem2 = x
        x = self.conv3b(x, rem3)
        rem3 = x
        #1/16
#         x = self.conv4b(x, rem4)

#         x = self.deconv4b(x, rem3)
        x = self.deconv3b(x, rem2)
        x = self.deconv2b(x, rem1)
        x = self.deconv1b(x, rem0)
        feat4 = x
        gwc_feature = torch.cat((feat0,feat1,feat2,feat3,feat4),dim = 1)
        return conv0a,gwc_feature

卷积过程示意图如下图所示：

输入	输出形状
input	torch.Size([1, 1, 384, 1280])
firstconv	torch.Size([1, 32, 192, 640])
layer1	torch.Size([1, 32, 192, 640])
layer2	torch.Size([1, 64, 96, 320])
layer3	torch.Size([1, 128, 48, 160])
layer4	torch.Size([1, 128, 48, 160])
reduce	torch.Size([1, 32, 48, 160])
conv1a	torch.Size([1, 48, 24, 80])
conv2a	torch.Size([1, 64, 12, 40])
conv3a	torch.Size([1, 96, 6, 20])
deconv3a	torch.Size([1, 64, 12, 40])
deconv2a	torch.Size([1, 48, 24, 80])
deconv1a	torch.Size([1, 32, 48, 160])
conv1b	torch.Size([1, 48, 24, 80])
conv2b	torch.Size([1, 64, 12, 40])
conv3b	torch.Size([1, 96, 6, 20])
deconv3b	torch.Size([1, 64, 12, 40])
deconv2b	torch.Size([1, 48, 24, 80])
deconv1b	torch.Size([1, 32, 48, 160])
gwc_feature	torch.Size([1, 352, 48, 160])

视差预测模块

class CoeffsPredictor(HourGlass):

    def __init__(self, hourglass_inplanes=16):
        super(CoeffsPredictor, self).__init__(hourglass_inplanes)

    def forward(self, input):
        output0 = self.conv1(input)
        output0_a = self.conv2(output0) + output0

        output0 = self.conv1_1(output0_a)
        output0_c = self.conv2_1(output0) + output0

        output0 = self.conv3(output0_c)
        output0 = self.conv4(output0) + output0

        output1 = self.conv5(output0) + output0_c
        output1 = self.conv6(output1) + output0_a
        output1 = self.conv7(output1)
        #[B,G,D,H,W] -> [B,D,G,H,W]
        coeffs = self.last_for_guidance(output1).permute(0,2,1,3,4).contiguous()
        return coeffs

网络结构

class BGNet(SubModule):
    def __init__(self):
        super(BGNet, self).__init__()
        self.softmax = nn.Softmax(dim = 1)
        # self.refinement_net = HourglassRefinement()
      
        self.feature_extraction = feature_extraction()
        self.coeffs_disparity_predictor = CoeffsPredictor()


        self.dres0 = nn.Sequential(convbn_3d_lrelu(44, 32, 3, 1, 1),
                                   convbn_3d_lrelu(32, 16, 3, 1, 1))
        self.guide = GuideNN()
        self.slice = Slice()
        self.weight_init()

    def forward(self, left_input, right_input):         
        left_low_level_features_1,  left_gwc_feature  = self.feature_extraction(left_input)
        _,                          right_gwc_feature = self.feature_extraction(right_input)
        
        guide = self.guide(left_low_level_features_1) #[B,1,H,W]
        # torch.cuda.synchronize()
        # start = time.time()
        cost_volume = build_gwc_volume(left_gwc_feature,right_gwc_feature,25,44)
        cost_volume = self.dres0(cost_volume)
        #coeffs:[B,D,G,H,W]
        coeffs = self.coeffs_disparity_predictor(cost_volume)

        list_coeffs = torch.split(coeffs,1,dim = 1)
        index = torch.arange(0,97)
        index_float = index/4.0
        index_a = torch.floor(index_float)
        index_b = index_a + 1
        
        index_a = torch.clamp(index_a, min=0, max= 24)
        index_b = torch.clamp(index_b, min=0, max= 24)
        
        wa = index_b - index_float
        wb = index_float   - index_a

        list_float = []  
        device = list_coeffs[0].get_device()
        wa  = wa.view(1,-1,1,1)
        wb  = wb.view(1,-1,1,1)
        wa = wa.to(device)
        wb = wb.to(device)
        wa = wa.float()
        wb = wb.float()

        N, _, H, W = guide.shape
        #[H,W]
        hg, wg = torch.meshgrid([torch.arange(0, H), torch.arange(0, W)]) # [0,511] HxW
        if device >= 0:
            hg = hg.to(device)
            wg = wg.to(device)
        #[B,H,W,1]
        hg = hg.float().repeat(N, 1, 1).unsqueeze(3) / (H-1) * 2 - 1 # norm to [-1,1] NxHxWx1
        wg = wg.float().repeat(N, 1, 1).unsqueeze(3) / (W-1) * 2 - 1 # norm to [-1,1] NxHxWx1
        slice_dict = []
        # torch.cuda.synchronize()
        # start = time.time()
        for i in range(25):
            slice_dict.append(self.slice(list_coeffs[i], wg, hg, guide)) #[B,1,H,W]
        slice_dict_a = []
        slice_dict_b = []
        for i in range(97):
            inx_a = i//4
            inx_b = inx_a + 1
            inx_b  = min(inx_b,24)
            slice_dict_a.append(slice_dict[inx_a])
            slice_dict_b.append(slice_dict[inx_b])
            
        final_cost_volume = wa * torch.cat(slice_dict_a,dim = 1) + wb * torch.cat(slice_dict_b,dim = 1)
        slice = self.softmax(final_cost_volume)
        disparity_samples = torch.arange(0, 97, dtype=slice.dtype, device=slice.device).view(1, 97, 1, 1)
        
        disparity_samples = disparity_samples.repeat(slice.size()[0],1,slice.size()[2],slice.size()[3])
        half_disp = torch.sum(disparity_samples * slice,dim = 1).unsqueeze(1)
        out2 = F.interpolate(half_disp * 2.0, scale_factor=(2.0, 2.0),
                                      mode='bilinear',align_corners =False).squeeze(1)
                                            
        return out2,out2

feature_extraction()返回conv0a，gwc_feature两个值，其中conv0a是仅经过firstconv和layer1的结果，gwc_feature是经过整个feature_extraction()的结果。
left_low_level_features_1会接着进入到GuideNN()里变成guide ，guide 就是论文里提到的Guidance Map，形状为[B,1,H,W]。
cost_volume的构建分两步，build_gwc_volume和dres0，build_gwc_volume就是论文里提到的为成本聚合构建了一个分组相关成本量，通过将特征通道划分为 Ng 组，为视差d 的每个像素位置 (x,y) 计算分组相关，这边具体的操作很细节，需要单独细讲。
coeffs_disparity_predictor是继承的HourGlass，有点像PSMNET里面的3D卷积结构。

BGNet_plus网络具体结构与每层权重大小

层名	权重大小
refinement_net.conv1.0.weight	torch.Size([16, 2, 3, 3])
refinement_net.conv1.1.weight	torch.Size([16])
refinement_net.conv1.1.bias	torch.Size([16])
refinement_net.conv1.1.running_mean	torch.Size([16])
refinement_net.conv1.1.running_var	torch.Size([16])
refinement_net.conv1.1.num_batches_tracked	torch.Size([])
refinement_net.conv2.0.weight	torch.Size([16, 1, 3, 3])
refinement_net.conv2.1.weight	torch.Size([16])
refinement_net.conv2.1.bias	torch.Size([16])
refinement_net.conv2.1.running_mean	torch.Size([16])
refinement_net.conv2.1.running_var	torch.Size([16])
refinement_net.conv2.1.num_batches_tracked	torch.Size([])
refinement_net.conv_start.0.weight	torch.Size([32, 32, 3, 3])
refinement_net.conv_start.1.weight	torch.Size([32])
refinement_net.conv_start.1.bias	torch.Size([32])
refinement_net.conv_start.1.running_mean	torch.Size([32])
refinement_net.conv_start.1.running_var	torch.Size([32])
refinement_net.conv_start.1.num_batches_tracked	torch.Size([])
refinement_net.conv1a.conv.weight	torch.Size([48, 32, 3, 3])
refinement_net.conv1a.bn.weight	torch.Size([48])
refinement_net.conv1a.bn.bias	torch.Size([48])
refinement_net.conv1a.bn.running_mean	torch.Size([48])
refinement_net.conv1a.bn.running_var	torch.Size([48])
refinement_net.conv1a.bn.num_batches_tracked	torch.Size([])
refinement_net.conv2a.conv.weight	torch.Size([64, 48, 3, 3])
refinement_net.conv2a.bn.weight	torch.Size([64])
refinement_net.conv2a.bn.bias	torch.Size([64])
refinement_net.conv2a.bn.running_mean	torch.Size([64])
refinement_net.conv2a.bn.running_var	torch.Size([64])
refinement_net.conv2a.bn.num_batches_tracked	torch.Size([])
refinement_net.conv3a.conv.weight	torch.Size([96, 64, 3, 3])
refinement_net.conv3a.bn.weight	torch.Size([96])
refinement_net.conv3a.bn.bias	torch.Size([96])
refinement_net.conv3a.bn.running_mean	torch.Size([96])
refinement_net.conv3a.bn.running_var	torch.Size([96])
refinement_net.conv3a.bn.num_batches_tracked	torch.Size([])
refinement_net.conv4a.conv.weight	torch.Size([128, 96, 3, 3])
refinement_net.conv4a.bn.weight	torch.Size([128])
refinement_net.conv4a.bn.bias	torch.Size([128])
refinement_net.conv4a.bn.running_mean	torch.Size([128])
refinement_net.conv4a.bn.running_var	torch.Size([128])
refinement_net.conv4a.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv4a.conv1.conv.weight	torch.Size([128, 96, 4, 4])
refinement_net.deconv4a.conv1.bn.weight	torch.Size([96])
refinement_net.deconv4a.conv1.bn.bias	torch.Size([96])
refinement_net.deconv4a.conv1.bn.running_mean	torch.Size([96])
refinement_net.deconv4a.conv1.bn.running_var	torch.Size([96])
refinement_net.deconv4a.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv4a.conv2.conv.weight	torch.Size([96, 192, 3, 3])
refinement_net.deconv4a.conv2.bn.weight	torch.Size([96])
refinement_net.deconv4a.conv2.bn.bias	torch.Size([96])
refinement_net.deconv4a.conv2.bn.running_mean	torch.Size([96])
refinement_net.deconv4a.conv2.bn.running_var	torch.Size([96])
refinement_net.deconv4a.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv3a.conv1.conv.weight	torch.Size([96, 64, 4, 4])
refinement_net.deconv3a.conv1.bn.weight	torch.Size([64])
refinement_net.deconv3a.conv1.bn.bias	torch.Size([64])
refinement_net.deconv3a.conv1.bn.running_mean	torch.Size([64])
refinement_net.deconv3a.conv1.bn.running_var	torch.Size([64])
refinement_net.deconv3a.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv3a.conv2.conv.weight	torch.Size([64, 128, 3, 3])
refinement_net.deconv3a.conv2.bn.weight	torch.Size([64])
refinement_net.deconv3a.conv2.bn.bias	torch.Size([64])
refinement_net.deconv3a.conv2.bn.running_mean	torch.Size([64])
refinement_net.deconv3a.conv2.bn.running_var	torch.Size([64])
refinement_net.deconv3a.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv2a.conv1.conv.weight	torch.Size([64, 48, 4, 4])
refinement_net.deconv2a.conv1.bn.weight	torch.Size([48])
refinement_net.deconv2a.conv1.bn.bias	torch.Size([48])
refinement_net.deconv2a.conv1.bn.running_mean	torch.Size([48])
refinement_net.deconv2a.conv1.bn.running_var	torch.Size([48])
refinement_net.deconv2a.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv2a.conv2.conv.weight	torch.Size([48, 96, 3, 3])
refinement_net.deconv2a.conv2.bn.weight	torch.Size([48])
refinement_net.deconv2a.conv2.bn.bias	torch.Size([48])
refinement_net.deconv2a.conv2.bn.running_mean	torch.Size([48])
refinement_net.deconv2a.conv2.bn.running_var	torch.Size([48])
refinement_net.deconv2a.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv1a.conv1.conv.weight	torch.Size([48, 32, 4, 4])
refinement_net.deconv1a.conv1.bn.weight	torch.Size([32])
refinement_net.deconv1a.conv1.bn.bias	torch.Size([32])
refinement_net.deconv1a.conv1.bn.running_mean	torch.Size([32])
refinement_net.deconv1a.conv1.bn.running_var	torch.Size([32])
refinement_net.deconv1a.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv1a.conv2.conv.weight	torch.Size([32, 64, 3, 3])
refinement_net.deconv1a.conv2.bn.weight	torch.Size([32])
refinement_net.deconv1a.conv2.bn.bias	torch.Size([32])
refinement_net.deconv1a.conv2.bn.running_mean	torch.Size([32])
refinement_net.deconv1a.conv2.bn.running_var	torch.Size([32])
refinement_net.deconv1a.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.conv1b.conv1.conv.weight	torch.Size([48, 32, 3, 3])
refinement_net.conv1b.conv1.bn.weight	torch.Size([48])
refinement_net.conv1b.conv1.bn.bias	torch.Size([48])
refinement_net.conv1b.conv1.bn.running_mean	torch.Size([48])
refinement_net.conv1b.conv1.bn.running_var	torch.Size([48])
refinement_net.conv1b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.conv1b.conv2.conv.weight	torch.Size([48, 96, 3, 3])
refinement_net.conv1b.conv2.bn.weight	torch.Size([48])
refinement_net.conv1b.conv2.bn.bias	torch.Size([48])
refinement_net.conv1b.conv2.bn.running_mean	torch.Size([48])
refinement_net.conv1b.conv2.bn.running_var	torch.Size([48])
refinement_net.conv1b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.conv2b.conv1.conv.weight	torch.Size([64, 48, 3, 3])
refinement_net.conv2b.conv1.bn.weight	torch.Size([64])
refinement_net.conv2b.conv1.bn.bias	torch.Size([64])
refinement_net.conv2b.conv1.bn.running_mean	torch.Size([64])
refinement_net.conv2b.conv1.bn.running_var	torch.Size([64])
refinement_net.conv2b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.conv2b.conv2.conv.weight	torch.Size([64, 128, 3, 3])
refinement_net.conv2b.conv2.bn.weight	torch.Size([64])
refinement_net.conv2b.conv2.bn.bias	torch.Size([64])
refinement_net.conv2b.conv2.bn.running_mean	torch.Size([64])
refinement_net.conv2b.conv2.bn.running_var	torch.Size([64])
refinement_net.conv2b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.conv3b.conv1.conv.weight	torch.Size([96, 64, 3, 3])
refinement_net.conv3b.conv1.bn.weight	torch.Size([96])
refinement_net.conv3b.conv1.bn.bias	torch.Size([96])
refinement_net.conv3b.conv1.bn.running_mean	torch.Size([96])
refinement_net.conv3b.conv1.bn.running_var	torch.Size([96])
refinement_net.conv3b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.conv3b.conv2.conv.weight	torch.Size([96, 192, 3, 3])
refinement_net.conv3b.conv2.bn.weight	torch.Size([96])
refinement_net.conv3b.conv2.bn.bias	torch.Size([96])
refinement_net.conv3b.conv2.bn.running_mean	torch.Size([96])
refinement_net.conv3b.conv2.bn.running_var	torch.Size([96])
refinement_net.conv3b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.conv4b.conv1.conv.weight	torch.Size([128, 96, 3, 3])
refinement_net.conv4b.conv1.bn.weight	torch.Size([128])
refinement_net.conv4b.conv1.bn.bias	torch.Size([128])
refinement_net.conv4b.conv1.bn.running_mean	torch.Size([128])
refinement_net.conv4b.conv1.bn.running_var	torch.Size([128])
refinement_net.conv4b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.conv4b.conv2.conv.weight	torch.Size([128, 256, 3, 3])
refinement_net.conv4b.conv2.bn.weight	torch.Size([128])
refinement_net.conv4b.conv2.bn.bias	torch.Size([128])
refinement_net.conv4b.conv2.bn.running_mean	torch.Size([128])
refinement_net.conv4b.conv2.bn.running_var	torch.Size([128])
refinement_net.conv4b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv4b.conv1.conv.weight	torch.Size([128, 96, 4, 4])
refinement_net.deconv4b.conv1.bn.weight	torch.Size([96])
refinement_net.deconv4b.conv1.bn.bias	torch.Size([96])
refinement_net.deconv4b.conv1.bn.running_mean	torch.Size([96])
refinement_net.deconv4b.conv1.bn.running_var	torch.Size([96])
refinement_net.deconv4b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv4b.conv2.conv.weight	torch.Size([96, 192, 3, 3])
refinement_net.deconv4b.conv2.bn.weight	torch.Size([96])
refinement_net.deconv4b.conv2.bn.bias	torch.Size([96])
refinement_net.deconv4b.conv2.bn.running_mean	torch.Size([96])
refinement_net.deconv4b.conv2.bn.running_var	torch.Size([96])
refinement_net.deconv4b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv3b.conv1.conv.weight	torch.Size([96, 64, 4, 4])
refinement_net.deconv3b.conv1.bn.weight	torch.Size([64])
refinement_net.deconv3b.conv1.bn.bias	torch.Size([64])
refinement_net.deconv3b.conv1.bn.running_mean	torch.Size([64])
refinement_net.deconv3b.conv1.bn.running_var	torch.Size([64])
refinement_net.deconv3b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv3b.conv2.conv.weight	torch.Size([64, 128, 3, 3])
refinement_net.deconv3b.conv2.bn.weight	torch.Size([64])
refinement_net.deconv3b.conv2.bn.bias	torch.Size([64])
refinement_net.deconv3b.conv2.bn.running_mean	torch.Size([64])
refinement_net.deconv3b.conv2.bn.running_var	torch.Size([64])
refinement_net.deconv3b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv2b.conv1.conv.weight	torch.Size([64, 48, 4, 4])
refinement_net.deconv2b.conv1.bn.weight	torch.Size([48])
refinement_net.deconv2b.conv1.bn.bias	torch.Size([48])
refinement_net.deconv2b.conv1.bn.running_mean	torch.Size([48])
refinement_net.deconv2b.conv1.bn.running_var	torch.Size([48])
refinement_net.deconv2b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv2b.conv2.conv.weight	torch.Size([48, 96, 3, 3])
refinement_net.deconv2b.conv2.bn.weight	torch.Size([48])
refinement_net.deconv2b.conv2.bn.bias	torch.Size([48])
refinement_net.deconv2b.conv2.bn.running_mean	torch.Size([48])
refinement_net.deconv2b.conv2.bn.running_var	torch.Size([48])
refinement_net.deconv2b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv1b.conv1.conv.weight	torch.Size([48, 32, 4, 4])
refinement_net.deconv1b.conv1.bn.weight	torch.Size([32])
refinement_net.deconv1b.conv1.bn.bias	torch.Size([32])
refinement_net.deconv1b.conv1.bn.running_mean	torch.Size([32])
refinement_net.deconv1b.conv1.bn.running_var	torch.Size([32])
refinement_net.deconv1b.conv1.bn.num_batches_tracked	torch.Size([])
refinement_net.deconv1b.conv2.conv.weight	torch.Size([32, 64, 3, 3])
refinement_net.deconv1b.conv2.bn.weight	torch.Size([32])
refinement_net.deconv1b.conv2.bn.bias	torch.Size([32])
refinement_net.deconv1b.conv2.bn.running_mean	torch.Size([32])
refinement_net.deconv1b.conv2.bn.running_var	torch.Size([32])
refinement_net.deconv1b.conv2.bn.num_batches_tracked	torch.Size([])
refinement_net.final_conv.weight	torch.Size([1, 32, 3, 3])
refinement_net.final_conv.bias	torch.Size([1])
feature_extraction.firstconv.0.0.weight	torch.Size([32, 1, 3, 3])
feature_extraction.firstconv.0.1.weight	torch.Size([32])
feature_extraction.firstconv.0.1.bias	torch.Size([32])
feature_extraction.firstconv.0.1.running_mean	torch.Size([32])
feature_extraction.firstconv.0.1.running_var	torch.Size([32])
feature_extraction.firstconv.0.1.num_batches_tracked	torch.Size([])
feature_extraction.firstconv.1.0.weight	torch.Size([32, 32, 3, 3])
feature_extraction.firstconv.1.1.weight	torch.Size([32])
feature_extraction.firstconv.1.1.bias	torch.Size([32])
feature_extraction.firstconv.1.1.running_mean	torch.Size([32])
feature_extraction.firstconv.1.1.running_var	torch.Size([32])
feature_extraction.firstconv.1.1.num_batches_tracked	torch.Size([])
feature_extraction.firstconv.2.0.weight	torch.Size([32, 32, 3, 3])
feature_extraction.firstconv.2.1.weight	torch.Size([32])
feature_extraction.firstconv.2.1.bias	torch.Size([32])
feature_extraction.firstconv.2.1.running_mean	torch.Size([32])
feature_extraction.firstconv.2.1.running_var	torch.Size([32])
feature_extraction.firstconv.2.1.num_batches_tracked	torch.Size([])
feature_extraction.layer1.0.conv1.0.weight	torch.Size([32, 32, 3, 3])
feature_extraction.layer1.0.conv1.1.weight	torch.Size([32])
feature_extraction.layer1.0.conv1.1.bias	torch.Size([32])
feature_extraction.layer1.0.conv1.1.running_mean	torch.Size([32])
feature_extraction.layer1.0.conv1.1.running_var	torch.Size([32])
feature_extraction.layer1.0.conv1.1.num_batches_tracked	torch.Size([])
feature_extraction.layer1.0.conv2.0.weight	torch.Size([32, 32, 3, 3])
feature_extraction.layer1.0.conv2.1.weight	torch.Size([32])
feature_extraction.layer1.0.conv2.1.bias	torch.Size([32])
feature_extraction.layer1.0.conv2.1.running_mean	torch.Size([32])
feature_extraction.layer1.0.conv2.1.running_var	torch.Size([32])
feature_extraction.layer1.0.conv2.1.num_batches_tracked	torch.Size([])
feature_extraction.layer2.0.conv1.0.weight	torch.Size([64, 32, 3, 3])
feature_extraction.layer2.0.conv1.1.weight	torch.Size([64])
feature_extraction.layer2.0.conv1.1.bias	torch.Size([64])
feature_extraction.layer2.0.conv1.1.running_mean	torch.Size([64])
feature_extraction.layer2.0.conv1.1.running_var	torch.Size([64])
feature_extraction.layer2.0.conv1.1.num_batches_tracked	torch.Size([])
feature_extraction.layer2.0.conv2.0.weight	torch.Size([64, 64, 3, 3])
feature_extraction.layer2.0.conv2.1.weight	torch.Size([64])
feature_extraction.layer2.0.conv2.1.bias	torch.Size([64])
feature_extraction.layer2.0.conv2.1.running_mean	torch.Size([64])
feature_extraction.layer2.0.conv2.1.running_var	torch.Size([64])
feature_extraction.layer2.0.conv2.1.num_batches_tracked	torch.Size([])
feature_extraction.layer2.0.downsample.0.weight	torch.Size([64, 32, 1, 1])
feature_extraction.layer2.0.downsample.1.weight	torch.Size([64])
feature_extraction.layer2.0.downsample.1.bias	torch.Size([64])
feature_extraction.layer2.0.downsample.1.running_mean	torch.Size([64])
feature_extraction.layer2.0.downsample.1.running_var	torch.Size([64])
feature_extraction.layer2.0.downsample.1.num_batches_tracked	torch.Size([])
feature_extraction.layer3.0.conv1.0.weight	torch.Size([128, 64, 3, 3])
feature_extraction.layer3.0.conv1.1.weight	torch.Size([128])
feature_extraction.layer3.0.conv1.1.bias	torch.Size([128])
feature_extraction.layer3.0.conv1.1.running_mean	torch.Size([128])
feature_extraction.layer3.0.conv1.1.running_var	torch.Size([128])
feature_extraction.layer3.0.conv1.1.num_batches_tracked	torch.Size([])
feature_extraction.layer3.0.conv2.0.weight	torch.Size([128, 128, 3, 3])
feature_extraction.layer3.0.conv2.1.weight	torch.Size([128])
feature_extraction.layer3.0.conv2.1.bias	torch.Size([128])
feature_extraction.layer3.0.conv2.1.running_mean	torch.Size([128])
feature_extraction.layer3.0.conv2.1.running_var	torch.Size([128])
feature_extraction.layer3.0.conv2.1.num_batches_tracked	torch.Size([])
feature_extraction.layer3.0.downsample.0.weight	torch.Size([128, 64, 1, 1])
feature_extraction.layer3.0.downsample.1.weight	torch.Size([128])
feature_extraction.layer3.0.downsample.1.bias	torch.Size([128])
feature_extraction.layer3.0.downsample.1.running_mean	torch.Size([128])
feature_extraction.layer3.0.downsample.1.running_var	torch.Size([128])
feature_extraction.layer3.0.downsample.1.num_batches_tracked	torch.Size([])
feature_extraction.layer4.0.conv1.0.weight	torch.Size([128, 128, 3, 3])
feature_extraction.layer4.0.conv1.1.weight	torch.Size([128])
feature_extraction.layer4.0.conv1.1.bias	torch.Size([128])
feature_extraction.layer4.0.conv1.1.running_mean	torch.Size([128])
feature_extraction.layer4.0.conv1.1.running_var	torch.Size([128])
feature_extraction.layer4.0.conv1.1.num_batches_tracked	torch.Size([])
feature_extraction.layer4.0.conv2.0.weight	torch.Size([128, 128, 3, 3])
feature_extraction.layer4.0.conv2.1.weight	torch.Size([128])
feature_extraction.layer4.0.conv2.1.bias	torch.Size([128])
feature_extraction.layer4.0.conv2.1.running_mean	torch.Size([128])
feature_extraction.layer4.0.conv2.1.running_var	torch.Size([128])
feature_extraction.layer4.0.conv2.1.num_batches_tracked	torch.Size([])
feature_extraction.reduce.0.weight	torch.Size([32, 128, 3, 3])
feature_extraction.reduce.1.weight	torch.Size([32])
feature_extraction.reduce.1.bias	torch.Size([32])
feature_extraction.reduce.1.running_mean	torch.Size([32])
feature_extraction.reduce.1.running_var	torch.Size([32])
feature_extraction.reduce.1.num_batches_tracked	torch.Size([])
feature_extraction.conv1a.conv.weight	torch.Size([48, 32, 3, 3])
feature_extraction.conv1a.bn.weight	torch.Size([48])
feature_extraction.conv1a.bn.bias	torch.Size([48])
feature_extraction.conv1a.bn.running_mean	torch.Size([48])
feature_extraction.conv1a.bn.running_var	torch.Size([48])
feature_extraction.conv1a.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv2a.conv.weight	torch.Size([64, 48, 3, 3])
feature_extraction.conv2a.bn.weight	torch.Size([64])
feature_extraction.conv2a.bn.bias	torch.Size([64])
feature_extraction.conv2a.bn.running_mean	torch.Size([64])
feature_extraction.conv2a.bn.running_var	torch.Size([64])
feature_extraction.conv2a.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv3a.conv.weight	torch.Size([96, 64, 3, 3])
feature_extraction.conv3a.bn.weight	torch.Size([96])
feature_extraction.conv3a.bn.bias	torch.Size([96])
feature_extraction.conv3a.bn.running_mean	torch.Size([96])
feature_extraction.conv3a.bn.running_var	torch.Size([96])
feature_extraction.conv3a.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv3a.conv1.conv.weight	torch.Size([96, 64, 4, 4])
feature_extraction.deconv3a.conv1.bn.weight	torch.Size([64])
feature_extraction.deconv3a.conv1.bn.bias	torch.Size([64])
feature_extraction.deconv3a.conv1.bn.running_mean	torch.Size([64])
feature_extraction.deconv3a.conv1.bn.running_var	torch.Size([64])
feature_extraction.deconv3a.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv3a.conv2.conv.weight	torch.Size([64, 128, 3, 3])
feature_extraction.deconv3a.conv2.bn.weight	torch.Size([64])
feature_extraction.deconv3a.conv2.bn.bias	torch.Size([64])
feature_extraction.deconv3a.conv2.bn.running_mean	torch.Size([64])
feature_extraction.deconv3a.conv2.bn.running_var	torch.Size([64])
feature_extraction.deconv3a.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv2a.conv1.conv.weight	torch.Size([64, 48, 4, 4])
feature_extraction.deconv2a.conv1.bn.weight	torch.Size([48])
feature_extraction.deconv2a.conv1.bn.bias	torch.Size([48])
feature_extraction.deconv2a.conv1.bn.running_mean	torch.Size([48])
feature_extraction.deconv2a.conv1.bn.running_var	torch.Size([48])
feature_extraction.deconv2a.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv2a.conv2.conv.weight	torch.Size([48, 96, 3, 3])
feature_extraction.deconv2a.conv2.bn.weight	torch.Size([48])
feature_extraction.deconv2a.conv2.bn.bias	torch.Size([48])
feature_extraction.deconv2a.conv2.bn.running_mean	torch.Size([48])
feature_extraction.deconv2a.conv2.bn.running_var	torch.Size([48])
feature_extraction.deconv2a.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv1a.conv1.conv.weight	torch.Size([48, 32, 4, 4])
feature_extraction.deconv1a.conv1.bn.weight	torch.Size([32])
feature_extraction.deconv1a.conv1.bn.bias	torch.Size([32])
feature_extraction.deconv1a.conv1.bn.running_mean	torch.Size([32])
feature_extraction.deconv1a.conv1.bn.running_var	torch.Size([32])
feature_extraction.deconv1a.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv1a.conv2.conv.weight	torch.Size([32, 64, 3, 3])
feature_extraction.deconv1a.conv2.bn.weight	torch.Size([32])
feature_extraction.deconv1a.conv2.bn.bias	torch.Size([32])
feature_extraction.deconv1a.conv2.bn.running_mean	torch.Size([32])
feature_extraction.deconv1a.conv2.bn.running_var	torch.Size([32])
feature_extraction.deconv1a.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv1b.conv1.conv.weight	torch.Size([48, 32, 3, 3])
feature_extraction.conv1b.conv1.bn.weight	torch.Size([48])
feature_extraction.conv1b.conv1.bn.bias	torch.Size([48])
feature_extraction.conv1b.conv1.bn.running_mean	torch.Size([48])
feature_extraction.conv1b.conv1.bn.running_var	torch.Size([48])
feature_extraction.conv1b.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv1b.conv2.conv.weight	torch.Size([48, 96, 3, 3])
feature_extraction.conv1b.conv2.bn.weight	torch.Size([48])
feature_extraction.conv1b.conv2.bn.bias	torch.Size([48])
feature_extraction.conv1b.conv2.bn.running_mean	torch.Size([48])
feature_extraction.conv1b.conv2.bn.running_var	torch.Size([48])
feature_extraction.conv1b.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv2b.conv1.conv.weight	torch.Size([64, 48, 3, 3])
feature_extraction.conv2b.conv1.bn.weight	torch.Size([64])
feature_extraction.conv2b.conv1.bn.bias	torch.Size([64])
feature_extraction.conv2b.conv1.bn.running_mean	torch.Size([64])
feature_extraction.conv2b.conv1.bn.running_var	torch.Size([64])
feature_extraction.conv2b.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv2b.conv2.conv.weight	torch.Size([64, 128, 3, 3])
feature_extraction.conv2b.conv2.bn.weight	torch.Size([64])
feature_extraction.conv2b.conv2.bn.bias	torch.Size([64])
feature_extraction.conv2b.conv2.bn.running_mean	torch.Size([64])
feature_extraction.conv2b.conv2.bn.running_var	torch.Size([64])
feature_extraction.conv2b.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv3b.conv1.conv.weight	torch.Size([96, 64, 3, 3])
feature_extraction.conv3b.conv1.bn.weight	torch.Size([96])
feature_extraction.conv3b.conv1.bn.bias	torch.Size([96])
feature_extraction.conv3b.conv1.bn.running_mean	torch.Size([96])
feature_extraction.conv3b.conv1.bn.running_var	torch.Size([96])
feature_extraction.conv3b.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.conv3b.conv2.conv.weight	torch.Size([96, 192, 3, 3])
feature_extraction.conv3b.conv2.bn.weight	torch.Size([96])
feature_extraction.conv3b.conv2.bn.bias	torch.Size([96])
feature_extraction.conv3b.conv2.bn.running_mean	torch.Size([96])
feature_extraction.conv3b.conv2.bn.running_var	torch.Size([96])
feature_extraction.conv3b.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv3b.conv1.conv.weight	torch.Size([96, 64, 4, 4])
feature_extraction.deconv3b.conv1.bn.weight	torch.Size([64])
feature_extraction.deconv3b.conv1.bn.bias	torch.Size([64])
feature_extraction.deconv3b.conv1.bn.running_mean	torch.Size([64])
feature_extraction.deconv3b.conv1.bn.running_var	torch.Size([64])
feature_extraction.deconv3b.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv3b.conv2.conv.weight	torch.Size([64, 128, 3, 3])
feature_extraction.deconv3b.conv2.bn.weight	torch.Size([64])
feature_extraction.deconv3b.conv2.bn.bias	torch.Size([64])
feature_extraction.deconv3b.conv2.bn.running_mean	torch.Size([64])
feature_extraction.deconv3b.conv2.bn.running_var	torch.Size([64])
feature_extraction.deconv3b.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv2b.conv1.conv.weight	torch.Size([64, 48, 4, 4])
feature_extraction.deconv2b.conv1.bn.weight	torch.Size([48])
feature_extraction.deconv2b.conv1.bn.bias	torch.Size([48])
feature_extraction.deconv2b.conv1.bn.running_mean	torch.Size([48])
feature_extraction.deconv2b.conv1.bn.running_var	torch.Size([48])
feature_extraction.deconv2b.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv2b.conv2.conv.weight	torch.Size([48, 96, 3, 3])
feature_extraction.deconv2b.conv2.bn.weight	torch.Size([48])
feature_extraction.deconv2b.conv2.bn.bias	torch.Size([48])
feature_extraction.deconv2b.conv2.bn.running_mean	torch.Size([48])
feature_extraction.deconv2b.conv2.bn.running_var	torch.Size([48])
feature_extraction.deconv2b.conv2.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv1b.conv1.conv.weight	torch.Size([48, 32, 4, 4])
feature_extraction.deconv1b.conv1.bn.weight	torch.Size([32])
feature_extraction.deconv1b.conv1.bn.bias	torch.Size([32])
feature_extraction.deconv1b.conv1.bn.running_mean	torch.Size([32])
feature_extraction.deconv1b.conv1.bn.running_var	torch.Size([32])
feature_extraction.deconv1b.conv1.bn.num_batches_tracked	torch.Size([])
feature_extraction.deconv1b.conv2.conv.weight	torch.Size([32, 64, 3, 3])
feature_extraction.deconv1b.conv2.bn.weight	torch.Size([32])
feature_extraction.deconv1b.conv2.bn.bias	torch.Size([32])
feature_extraction.deconv1b.conv2.bn.running_mean	torch.Size([32])
feature_extraction.deconv1b.conv2.bn.running_var	torch.Size([32])
feature_extraction.deconv1b.conv2.bn.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv1.0.weight	torch.Size([32, 16, 3, 3, 3])
coeffs_disparity_predictor.conv1.1.weight	torch.Size([32])
coeffs_disparity_predictor.conv1.1.bias	torch.Size([32])
coeffs_disparity_predictor.conv1.1.running_mean	torch.Size([32])
coeffs_disparity_predictor.conv1.1.running_var	torch.Size([32])
coeffs_disparity_predictor.conv1.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv2.0.weight	torch.Size([32, 32, 3, 3, 3])
coeffs_disparity_predictor.conv2.1.weight	torch.Size([32])
coeffs_disparity_predictor.conv2.1.bias	torch.Size([32])
coeffs_disparity_predictor.conv2.1.running_mean	torch.Size([32])
coeffs_disparity_predictor.conv2.1.running_var	torch.Size([32])
coeffs_disparity_predictor.conv2.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv1_1.0.weight	torch.Size([64, 32, 3, 3, 3])
coeffs_disparity_predictor.conv1_1.1.weight	torch.Size([64])
coeffs_disparity_predictor.conv1_1.1.bias	torch.Size([64])
coeffs_disparity_predictor.conv1_1.1.running_mean	torch.Size([64])
coeffs_disparity_predictor.conv1_1.1.running_var	torch.Size([64])
coeffs_disparity_predictor.conv1_1.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv2_1.0.weight	torch.Size([64, 64, 3, 3, 3])
coeffs_disparity_predictor.conv2_1.1.weight	torch.Size([64])
coeffs_disparity_predictor.conv2_1.1.bias	torch.Size([64])
coeffs_disparity_predictor.conv2_1.1.running_mean	torch.Size([64])
coeffs_disparity_predictor.conv2_1.1.running_var	torch.Size([64])
coeffs_disparity_predictor.conv2_1.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv3.0.weight	torch.Size([128, 64, 3, 3, 3])
coeffs_disparity_predictor.conv3.1.weight	torch.Size([128])
coeffs_disparity_predictor.conv3.1.bias	torch.Size([128])
coeffs_disparity_predictor.conv3.1.running_mean	torch.Size([128])
coeffs_disparity_predictor.conv3.1.running_var	torch.Size([128])
coeffs_disparity_predictor.conv3.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv4.0.weight	torch.Size([128, 128, 3, 3, 3])
coeffs_disparity_predictor.conv4.1.weight	torch.Size([128])
coeffs_disparity_predictor.conv4.1.bias	torch.Size([128])
coeffs_disparity_predictor.conv4.1.running_mean	torch.Size([128])
coeffs_disparity_predictor.conv4.1.running_var	torch.Size([128])
coeffs_disparity_predictor.conv4.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv5.0.weight	torch.Size([128, 64, 3, 3, 3])
coeffs_disparity_predictor.conv5.1.weight	torch.Size([64])
coeffs_disparity_predictor.conv5.1.bias	torch.Size([64])
coeffs_disparity_predictor.conv5.1.running_mean	torch.Size([64])
coeffs_disparity_predictor.conv5.1.running_var	torch.Size([64])
coeffs_disparity_predictor.conv5.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv6.0.weight	torch.Size([64, 32, 3, 3, 3])
coeffs_disparity_predictor.conv6.1.weight	torch.Size([32])
coeffs_disparity_predictor.conv6.1.bias	torch.Size([32])
coeffs_disparity_predictor.conv6.1.running_mean	torch.Size([32])
coeffs_disparity_predictor.conv6.1.running_var	torch.Size([32])
coeffs_disparity_predictor.conv6.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.conv7.0.weight	torch.Size([32, 16, 3, 3, 3])
coeffs_disparity_predictor.conv7.1.weight	torch.Size([16])
coeffs_disparity_predictor.conv7.1.bias	torch.Size([16])
coeffs_disparity_predictor.conv7.1.running_mean	torch.Size([16])
coeffs_disparity_predictor.conv7.1.running_var	torch.Size([16])
coeffs_disparity_predictor.conv7.1.num_batches_tracked	torch.Size([])
coeffs_disparity_predictor.last_for_guidance.0.weight	torch.Size([32, 16, 3, 3, 3])
coeffs_disparity_predictor.last_for_guidance.1.weight	torch.Size([32])
coeffs_disparity_predictor.last_for_guidance.1.bias	torch.Size([32])
coeffs_disparity_predictor.last_for_guidance.1.running_mean	torch.Size([32])
coeffs_disparity_predictor.last_for_guidance.1.running_var	torch.Size([32])
coeffs_disparity_predictor.last_for_guidance.1.num_batches_tracked	torch.Size([])
dres0.0.0.weight	torch.Size([32, 44, 3, 3, 3])
dres0.0.1.weight	torch.Size([32])
dres0.0.1.bias	torch.Size([32])
dres0.0.1.running_mean	torch.Size([32])
dres0.0.1.running_var	torch.Size([32])
dres0.0.1.num_batches_tracked	torch.Size([])
dres0.1.0.weight	torch.Size([16, 32, 3, 3, 3])
dres0.1.1.weight	torch.Size([16])
dres0.1.1.bias	torch.Size([16])
dres0.1.1.running_mean	torch.Size([16])
dres0.1.1.running_var	torch.Size([16])
dres0.1.1.num_batches_tracked	torch.Size([])
guide.conv1.0.weight	torch.Size([16, 32, 1, 1])
guide.conv1.1.weight	torch.Size([16])
guide.conv1.1.bias	torch.Size([16])
guide.conv1.1.running_mean	torch.Size([16])
guide.conv1.1.running_var	torch.Size([16])
guide.conv1.1.num_batches_tracked	torch.Size([])
guide.conv2.0.weight	torch.Size([1, 16, 1, 1])
guide.conv2.1.weight	torch.Size([1])
guide.conv2.1.bias	torch.Size([1])
guide.conv2.1.running_mean	torch.Size([1])
guide.conv2.1.running_var	torch.Size([1])
guide.conv2.1.num_batches_tracked	torch.Size([])

Number of model parameters: 5315811
参数总量为5.31M，参数主要集中在coeffs_disparity_predictor，因为是一个5维的匹配代价卷，所以参数量比较多。
（本文仅代表个人见解，如有不妥，欢迎大家共同探讨！如需转载，请联系本人。）