Multitask KISS-GP IndexError

In Multitask_GP_Regression_Scalable_With_KISSGP.ipynb replacing test_x = torch.linspace(0, 1, 51) with test_x = torch.zeros(1), i.e., predicting on a single test point gives the following error:

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-46-17c4b6da319c> in <module>()
     11 with torch.no_grad():
     12     test_x = torch.zeros(1)
---> 13     observed_pred = likelihood(model(test_x))
     14     # Get mean
     15     mean = observed_pred.mean()

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/models/exact_gp.py in __call__(self, *args, **kwargs)
    142                 n_train=n_train,
    143                 likelihood=self.likelihood,
--> 144                 precomputed_cache=self.mean_cache,
    145             )
    146             predictive_covar, covar_cache = exact_predictive_covar(

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/functions/__init__.py in exact_predictive_mean(full_covar, full_mean, train_labels, n_train, likelihood, precomputed_cache)
     89 
     90         full_covar = NonLazyVariable(full_covar)
---> 91     return full_covar.exact_predictive_mean(full_mean, train_labels, n_train, likelihood, precomputed_cache)
     92 
     93 

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/lazy/lazy_evaluated_kernel_variable.py in exact_predictive_mean(self, full_mean, train_labels, n_train, likelihood, precomputed_cache)
    124         else:
    125             return super(LazyEvaluatedKernelVariable, self).exact_predictive_mean(
--> 126                 full_mean, train_labels, n_train, likelihood, precomputed_cache
    127             )
    128 

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/lazy/lazy_variable.py in exact_predictive_mean(self, full_mean, train_labels, n_train, likelihood, precomputed_cache)
    427         else:
    428             test_train_covar = self[n_train:, :n_train]
--> 429         res = test_train_covar.matmul(precomputed_cache)
    430         if res.ndimension() == 3:
    431             res = res.squeeze(-1)

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/lazy/lazy_variable.py in matmul(self, tensor)
    632             raise RuntimeError
    633 
--> 634         func = Matmul(self.representation_tree())
    635         return func(tensor, *self.representation())
    636 

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/lazy/lazy_evaluated_kernel_variable.py in representation_tree(self)
    112 
    113     def representation_tree(self):
--> 114         return LazyVariableRepresentationTree(self.evaluate_kernel())
    115 
    116     def evaluate(self):

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/lazy/lazy_evaluated_kernel_variable.py in evaluate_kernel(self)
     96                 x2 = self.x2
     97 
---> 98             self._cached_kernel_eval = super(Kernel, self.kernel).__call__(x1, x2, **self.params)
     99             if self.squeeze_row:
    100                 self._cached_kernel_eval.squeeze_(-2)

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/module.py in __call__(self, *inputs, **kwargs)
    160 
    161     def __call__(self, *inputs, **kwargs):
--> 162         outputs = self.forward(*inputs, **kwargs)
    163         if torch.is_tensor(outputs) or isinstance(outputs, RandomVariable) or isinstance(outputs, LazyVariable):
    164             return outputs

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/kernels/multitask_kernel.py in forward(self, x1, x2)
     62         task_indices = torch.arange(self.n_tasks, device=x1.device).long()
     63         covar_i = self.task_covar_module(task_indices).evaluate_kernel()
---> 64         covar_x = self.data_covar_module.forward(x1, x2)
     65         if covar_x.size(0) == 1:
     66             covar_x = covar_x[0]

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/kernels/grid_interpolation_kernel.py in forward(self, x1, x2, **kwargs)
     56             base_lazy_var = base_lazy_var.repeat(x1.size(0), 1, 1)
     57 
---> 58         left_interp_indices, left_interp_values = self._compute_grid(x1)
     59         if torch.equal(x1.data, x2.data):
     60             right_interp_indices = left_interp_indices

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/kernels/grid_interpolation_kernel.py in _compute_grid(self, inputs)
     39         batch_size, n_data, n_dimensions = inputs.size()
     40         inputs = inputs.view(batch_size * n_data, n_dimensions)
---> 41         interp_indices, interp_values = Interpolation().interpolate(Variable(self.grid), inputs)
     42         interp_indices = interp_indices.view(batch_size, n_data, -1)
     43         interp_values = interp_values.view(batch_size, n_data, -1)

~/anaconda/envs/py3/lib/python3.6/site-packages/gpytorch/utils/interpolation.py in interpolate(self, x_grid, x_target, interp_points)
    114                     dim_interp_values[left_boundary_pts[j], :] = 0
    115                     dim_interp_values[left_boundary_pts[j], closest_from_first[j]] = 1
--> 116                     lower_grid_pt_idxs[left_boundary_pts[j]] = 0
    117 
    118             right_boundary_pts = torch.nonzero(lower_grid_pt_idxs > num_grid_points - num_coefficients)

IndexError: too many indices for tensor of dimension 0

[gpleiss]

Okay - the code in 1b809b3 should fix the problem you were seeing with 2D inputs. As far as test predictions for a single point - this commit might also fix that. If not, would you mind opening a separate issue for that?

[tzoiker]

Seems working, thank you!