Custom Mean error with MaternKernel

[ToennisStef]

Hello,
I encountered errors with the gpytorch.kernels.MaternKernel() as well as the gpytorch.kernels.ScaleKernel.MaternKernel(). I implemented a custom mean function like in #1669. (other useful discussions for implementing a custom mean are #1777 and #674). But i always got this RuntimError when making predictions with the custom GP class using the custom kernel:

File [c:\ProgramData\Anaconda3\envs\BO_init\Lib\site-packages\gpytorch\functions\matern_covariance.py:16], in MaternCovariance.forward(ctx, x1, x2, lengthscale, nu, dist_func)
     [13] # Subtract mean for numerical stability. Won't affect computations
     [14] # because covariance matrix is stationary.
     [15] needs_grad = any(ctx.needs_input_grad)
---> [16] mean = x1.reshape(-1, x1.size(-1)).mean(0)[(None,) * (x1.dim() - 1)]
     [17] x1_ = (x1 - mean).div(lengthscale)
     [18] x2_ = (x2 - mean).div(lengthscale)

RuntimeError: mean(): could not infer output dtype. Input dtype must be either a floating point or complex dtype. Got: Long

Apparently the mean function got the dtype Long passed but i don't get how. Using the gpytorch.kernels.RBFKernel(), gpytorch.kernels.ScaleKernel.RBFKernel() or gpytorch.kernels.CosineKernel() this error does not orrcur and using the custom kernel works perfectly fine.

Code for reproducing the error:
Definition of Custom Kernel:

class QuadraticMean(gpytorch.means.Mean):
    def __init__(self): #input_size, batch_shape=torch.Size()        
        super(QuadraticMean, self).__init__()
        self.a = torch.tensor(-2.0,dtype=torch.double) 

    def forward(self, x):
        return self.a*torch.sin(x**2).squeeze(-1)

Definition of the custom GPClass & initialization of the model:

# We will use the simplest form of GP model, exact inference
class ExactGPModel(gpytorch.models.ExactGP):
    def __init__(self, train_x, train_y, likelihood):
        super(ExactGPModel, self).__init__(train_x, train_y, likelihood)
        self.mean_module = QuadraticMean()
        self.covar_module = gpytorch.kernels.ScaleKernel.MaternKernel()

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

# initialize likelihood and model
likelihood = gpytorch.likelihoods.GaussianLikelihood()
model = ExactGPModel(train_X_init, train_Y_init, likelihood)

The error ocurred when executing the next cell (making predictions with the model):

res = 100
x1_eval = torch.linspace(-5, 5, res, dtype=torch.double).reshape([-1,1])
y1_eval = torchForresterFunction(x1_eval)
model.likelihood.noise_covar.noise = 1
model.covar_module.outputscale = 1
model.covar_module.base_kernel.lengthscale = 0.1

with torch.no_grad(), gpytorch.settings.fast_pred_var():
    model.eval()
    pred = model(x1_eval)

Does anybody know why this error ocurrs and why only with the MaternKernel?
Also what are the gpytorch.kernels.ScaleKernels? Is the only difference to the gpytorch.kernels. that they were given the additional Lengthscale parameter?
I am thankfull for any help!

Also i am fairly new to scripting in python and writing in VS-Code are there any recomendations for debugging such problems with these large dependencies? The long list of dependencies, (the different function calls and class calls) are always a bit intimidating. For example in this case i would not know how to go about finding the problematic line of code that produces the Long type that is eventually passed to the MaternCovariance.forward function, apart from maybe adding a print(x.dtype) at some point in the code. Again i am thankfull for any help!