在本例中，是否可以使用我创建的所有类通用的超类？

Question

我有所有这些类都非常相似，除了它们的covar_module。我想知道我是否可以创建一个通用的超类，它拥有所有的属性和函数，而子类只有covar_module。我是python的新手，所以我不知道synthax会是什么样子。

class RbfGP(ExactGP, GPyTorchModel):
    _num_outputs = 1  # to inform GPyTorchModel API

    def __init__(self, train_X, train_Y):
        # squeeze output dim before passing train_Y to ExactGP
        super().__init__(train_X, train_Y.squeeze(-1), GaussianLikelihood())
        self.mean_module = ConstantMean()
        self.covar_module = gpytorch.kernels.ScaleKernel(
            base_kernel= gpytorch.kernels.RBFKernel(ard_num_dims=train_X.shape[-1]),
        )
        self.to(train_X)  # make sure we're on the right device/dtype

    def forward(self, x):
        mean_x = self.mean_module(x)
        covar_x = self.covar_module(x)
        return MultivariateNormal(mean_x, covar_x)


class Matern12GP(ExactGP, GPyTorchModel):
    _num_outputs = 1  # to inform GPyTorchModel API

    def __init__(self, train_X, train_Y):
        # squeeze output dim before passing train_Y to ExactGP
        super().__init__(train_X, train_Y.squeeze(-1), GaussianLikelihood())
        self.mean_module = ConstantMean()
        self.covar_module = gpytorch.kernels.ScaleKernel(
            base_kernel=gpytorch.kernels.MaternKernel(nu=0.5, ard_num_dims=train_X.shape[-1]),
        )
        self.to(train_X)  # make sure we're on the right device/dtype

    def forward(self, x):
        mean_x = self.mean_module(x)
        covar_x = self.covar_module(x)
        return MultivariateNormal(mean_x, covar_x)


class Matern32GP(ExactGP, GPyTorchModel):
    _num_outputs = 1  # to inform GPyTorchModel API

    def __init__(self, train_X, train_Y):
        # squeeze output dim before passing train_Y to ExactGP
        super().__init__(train_X, train_Y.squeeze(-1), GaussianLikelihood())
        self.mean_module = ConstantMean()
        self.covar_module = gpytorch.kernels.ScaleKernel(
            base_kernel=gpytorch.kernels.MaternKernel(nu=1.5, ard_num_dims=train_X.shape[-1]),
        )
        self.to(train_X)  # make sure we're on the right device/dtype

    def forward(self, x):
        mean_x = self.mean_module(x)
        covar_x = self.covar_module(x)
        return MultivariateNormal(mean_x, covar_x)

Answer 1

为什么不这样做呢：

class GP(ExactGP, GPyTorchModel):
    _num_outputs = 1  # to inform GPyTorchModel API

    def __init__(self, train_X, train_Y, base_kernel, **kwargs):
        # squeeze output dim before passing train_Y to ExactGP
        super().__init__(train_X, train_Y.squeeze(-1), GaussianLikelihood())
        self.mean_module = ConstantMean()
        self.covar_module = gpytorch.kernels.ScaleKernel(
            base_kernel=base_kernel(ard_num_dims=train_X.shape[-1], **kwargs),
        )
        self.to(train_X)  # make sure we're on the right device/dtype

    def forward(self, x):
        mean_x = self.mean_module(x)
        covar_x = self.covar_module(x)
        return MultivariateNormal(mean_x, covar_x)

然后

rbf_gp = GP(train_x, train_y, base_kernel=gpytorch.kernels.RBFKernel)

matern_12_gp = GP(train_x, train_y, base_kernel=gpytorch.kernels.MaternKernel, nu=0.5)