为多类分类给出相等的输出值

Question

我正在尝试使用 PyTorch 构建一个用于预测图像中手指数量的 CNN。网络：

class Net(nn.Module):

    def __init__(self):
        super(Net, self).__init__()

        self.Layer1 = nn.Sequential(
        nn.Conv2d(in_channels=3, out_channels=16, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=(2, 2)),
        nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=128, out_channels=256, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=(2, 2)),
        nn.Conv2d(in_channels=256, out_channels=16, kernel_size=(1, 1)),
        nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=128, out_channels=256, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=(2, 2)),
        nn.Conv2d(in_channels=256, out_channels=16, kernel_size=(1, 1)),
        nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=(2, 2)),
        nn.Conv2d(in_channels=128, out_channels=16, kernel_size=(1, 1)),
        nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.MaxPool2d(kernel_size=(2, 2)),
        nn.Conv2d(in_channels=128, out_channels=16, kernel_size=(1, 1)),
        nn.Conv2d(in_channels=16, out_channels=32, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=64, out_channels=128, kernel_size=(3, 3)),
        nn.ReLU(),
        nn.Conv2d(in_channels=128, out_channels=256, kernel_size=(3, 3)),
        nn.ReLU(),
        )

        self.Layer2 = nn.Sequential(
        nn.Linear(1536, 100),
        nn.Tanh(),
        nn.Linear(100, 6),
        nn.Softmax()
        )
        self.optimizer = optimizers.Adadelta(self.parameters())

    def forward(self, X):
        X = self.Layer1(X)
        print(X.shape)
        X = self.Layer2(X.reshape(1, 1536))
        X = X.squeeze()

        return X

    def calc_loss(self, X, num):
        out = self.forward(X).unsqueeze(dim=0)
        print("Output: "+str(out))
        target = torch.tensor([num], dtype=torch.int64).cuda()
        criterion = nn.CrossEntropyLoss()
        loss = criterion(out, target)
        return loss

    def train_step(self, X, Y):
        loss = self.calc_loss(X, Y)
        print(loss)
        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.step()

但是，训练完成后，所有预测的值几乎相同（大约 0.15 ~ 0.18）。

看起来好像网络平均了输出概率以最小化损失，而不是学习实际值。

无论我将 Softmax 用于具有交叉熵损失的最后一层，还是使用具有二元交叉熵、MSE 或 SmoothL1Loss 的 Sigmoid，我都会得到相同的结果。

如果使用 Adam 优化器，我会得到类似的结果，只是在 1e-12 ~ 1e-14 的范围内。

我错过了什么？

Answer 1

如果您使用CrossEntropyLoss，则无需在forward 中使用Softmax。它已经包含在CrossEntropyLoss 中，因此您需要“原始”输出。但是，如果您在推理期间需要 Softmax，请改用 NLLLoss + 'Softmax'。

您可以找到更多信息here