为什么我的具有连续输出的 CNN 对于具有正标签的测试示例仅预测为零？

Question

我正在解决一个问题，其中我的训练示例是表示神经尖峰时间序列的二进制矩阵，而真实标签是表示 -40 到 +40 范围内的电压的连续值。 当我训练一个相当浅的 CNN 时，预测值和 grouth 真实值之间的相关性非常好，尽管显然存在异方差问题：

class CNN_shallow(torch.nn.Module):
    def __init__(self):
        super(CNN_shallow, self).__init__()
        self.conv1 = torch.nn.Conv2d(1,16,(3,1),stride=(1,1))
        self.conv2 = torch.nn.Conv2d(16,16,(3,1),stride=(2,1))
        self.conv3 = torch.nn.Conv2d(16,16,(5,1),stride=(1,1))
        self.conv4 = torch.nn.Conv2d(16,16,(5,1),stride=(3,1))
        self.conv5 = torch.nn.Conv2d(16,16,(7,1),stride=(1,1))
        self.dropout = torch.nn.Dropout(p=0.25)
        self.fc1 = torch.nn.Linear(4224,1)
        
    def forward(self, x):
        x = x.float()
        x = F.relu(self.conv1(x))
        x = F.relu(self.conv2(x))
        x = F.relu(self.conv3(x))
        x = F.relu(self.conv4(x))
        x = F.relu(self.conv5(x))
        x = x.view(-1,4224)
        x = self.dropout(x)
        u_pred = self.fc1(x)
        return(u_pred)

然而，当我训练一个更深的模型时，过滤器的数量在更深的层中逐渐增加，在模型的末尾添加几个完全连接的层，添加更多的 dropout，并添加 BatchNorm，我得到的模型无法预测任何对于具有真实正电压的测试用例，电压高于零：

class CNN_Model_v10(torch.nn.Module):
    def __init__(self):
        super(CNN_Model_v10, self).__init__()
        self.conv1 = torch.nn.Conv2d(1,16,3,stride=(1,1),padding=2)
        self.conv2 = torch.nn.Conv2d(16,32,3,stride=(2,1),padding=2)
        self.conv3 = torch.nn.Conv2d(32,32,3,stride=(1,1),padding=2)
        self.conv4 = torch.nn.Conv2d(32,64,3,stride=(2,1),padding=2)
        self.conv5 = torch.nn.Conv2d(64,64,3,stride=(1,1),padding=2)
        self.conv6 = torch.nn.Conv2d(64,64,3,stride=(2,1),padding=2)
        self.conv7 = torch.nn.Conv2d(64,128,3,stride=(1,1),padding=2)
        self.conv8 = torch.nn.Conv2d(128,128,3,stride=(2,1),padding=2)
        self.conv9 = torch.nn.Conv2d(128,128,3,stride=(1,1),padding=2)
        self.conv10 = torch.nn.Conv2d(128,256,3,stride=(2,1),padding=2)
        self.conv11 = torch.nn.Conv2d(256,256,3,stride=(1,1),padding=2)
        self.conv12 = torch.nn.Conv2d(256,256,3,stride=(2,1),padding=2)
        self.conv13 = torch.nn.Conv2d(256,256,3,stride=(1,1),padding=2)
        self.dropout = torch.nn.Dropout(p=0.5)
        self.fc1 = torch.nn.Linear(87040,128)
        self.fc2 = torch.nn.Linear(128,256)
        self.fc3 = torch.nn.Linear(256,256)
        self.fc4 = torch.nn.Linear(256,1)
        self.bn1 = nn.BatchNorm2d(16)
        self.bn2 = nn.BatchNorm2d(32)
        self.bn3 = nn.BatchNorm2d(32)
        self.bn4 = nn.BatchNorm2d(64)
        self.bn5 = nn.BatchNorm2d(64)
        self.bn6 = nn.BatchNorm2d(64)
        self.bn7 = nn.BatchNorm2d(128)
        self.bn8 = nn.BatchNorm2d(128)
        self.bn9 = nn.BatchNorm2d(128)
        self.bn10 = nn.BatchNorm2d(256)
        self.bn11 = nn.BatchNorm2d(256)
        self.bn12 = nn.BatchNorm2d(256)
        self.bn13 = nn.BatchNorm2d(256)
        self.bn14 = nn.BatchNorm1d(128)
        self.bn15 = nn.BatchNorm1d(256)
        self.bn16 = nn.BatchNorm1d(256)
        
        
    def forward(self, x):
        x = x.float()
        x = self.conv1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv2(x)
        x = self.bn2(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv3(x)
        x = self.bn3(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv4(x)
        x = self.bn4(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv5(x)
        x = self.bn5(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv6(x)
        x = self.bn6(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv7(x)
        x = self.bn7(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv8(x)
        x = self.bn8(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv9(x)
        x = self.bn9(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv10(x)
        x = self.bn10(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv11(x)
        x = self.bn11(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv12(x)
        x = self.bn12(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.conv13(x)
        x = self.bn13(x)
        x = F.relu(x)
        x = x.view(x.size(0),87040)
        x = self.dropout(x)
        x = self.fc1(x)
        x = self.bn14(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.fc2(x)
        x = self.bn15(x)
        x = F.relu(x)
        x = self.dropout(x)
        x = self.fc3(x)
        x = self.bn16(x)
        x = F.relu(x)
        x = self.dropout(x)
        u_pred = self.fc4(x)
        return(u_pred)

我只包含了前向传递的代码，因为两个模型的所有其余代码都是相同的。

对于更深的模型，似乎应该相对于更浅的模型过度拟合，似乎第二个模型中的偏差要严重得多，尤其是当基本事实为正时。 有没有人见过类似的结果？