search

cornellius-gp / gpytorch

A highly efficient implementation of Gaussian Processes in PyTorch
MIT License
3.46k stars 546 forks source link

[Bug] test_t_matmul_matrix fails #2396

Open veprbl opened 10 months ago

veprbl commented 10 months ago

🐛 Bug

The test fails with:

RuntimeError: Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. This is probably a bug in GPyTorch.

To reproduce

Run tests

Stack trace/error message 

___________ TestLazyEvaluatedKernelTensorBatch.test_t_matmul_matrix ____________

self = <test.lazy.test_lazy_evaluated_kernel_tensor.TestLazyEvaluatedKernelTensorBatch testMethod=test_t_matmul_matrix>

    def test_t_matmul_matrix(self):
        with torch.no_grad():
            linear_op = self.create_linear_op()
            rhs = torch.randn(*linear_op.batch_shape, linear_op.size(-2), 4)
            linear_op_copy = torch.clone(linear_op)
            evaluated = self.evaluate_linear_op(linear_op_copy)
            rhs_evaluated = to_dense(rhs)

            # Test operator
>           res = linear_op._t_matmul(rhs)

/nix/store/ghl89pdl4wyqifz498kzz9n20pph3jsh-python3.10-linear_operator-0.5.1/lib/python3.10/site-packages/linear_operator/test/linear_operator_test_case.py:403: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/nix/store/ghl89pdl4wyqifz498kzz9n20pph3jsh-python3.10-linear_operator-0.5.1/lib/python3.10/site-packages/linear_operator/operators/_linear_operator.py:914: in _t_matmul
    return self.mT._matmul(rhs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <gpytorch.lazy.lazy_evaluated_kernel_tensor.LazyEvaluatedKernelTensor object at 0x19da25180>
rhs = tensor([[[-1.1256e+00, -3.1700e-01, -1.0925e+00, -8.5194e-02],
         [-9.3348e-02,  6.8705e-01, -8.3832e-01,  8.918... [ 1.2390e-01,  1.1648e+00,  9.2337e-01,  1.3873e+00],
         [ 1.3750e+00,  6.5963e-01,  4.7656e-01, -1.0163e+00]]])

    def _matmul(self, rhs):
        # This _matmul is defined computes the kernel in chunks
        # It is only used when we are using kernel checkpointing
        # It won't be called if checkpointing is off
        x1 = self.x1
        x2 = self.x2

        split_size = beta_features.checkpoint_kernel.value()
        if not split_size:
>           raise RuntimeError(
                "Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. "
                "This is probably a bug in GPyTorch."
            )
E           RuntimeError: Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. This is probably a bug in GPyTorch.

gpytorch/lazy/lazy_evaluated_kernel_tensor.py:254: RuntimeError
_______ TestLazyEvaluatedKernelTensorMultitaskBatch.test_t_matmul_matrix _______

self = <test.lazy.test_lazy_evaluated_kernel_tensor.TestLazyEvaluatedKernelTensorMultitaskBatch testMethod=test_t_matmul_matrix>

    def test_t_matmul_matrix(self):
        with torch.no_grad():
            linear_op = self.create_linear_op()
            rhs = torch.randn(*linear_op.batch_shape, linear_op.size(-2), 4)
            linear_op_copy = torch.clone(linear_op)
            evaluated = self.evaluate_linear_op(linear_op_copy)
            rhs_evaluated = to_dense(rhs)

            # Test operator
>           res = linear_op._t_matmul(rhs)

/nix/store/ghl89pdl4wyqifz498kzz9n20pph3jsh-python3.10-linear_operator-0.5.1/lib/python3.10/site-packages/linear_operator/test/linear_operator_test_case.py:403: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/nix/store/ghl89pdl4wyqifz498kzz9n20pph3jsh-python3.10-linear_operator-0.5.1/lib/python3.10/site-packages/linear_operator/operators/_linear_operator.py:914: in _t_matmul
    return self.mT._matmul(rhs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <gpytorch.lazy.lazy_evaluated_kernel_tensor.LazyEvaluatedKernelTensor object at 0x19de19870>
rhs = tensor([[[ 1.6459, -1.3602,  0.3446,  0.5199],
         [-2.6133, -1.6965, -0.2282,  0.2800],
         [ 0.2469,  0.07...891, -1.5100,  1.0241],
         [ 0.1954, -0.7371,  1.7001,  0.3462],
         [ 0.9711,  1.4503, -0.0519, -0.6284]]])

    def _matmul(self, rhs):
        # This _matmul is defined computes the kernel in chunks
        # It is only used when we are using kernel checkpointing
        # It won't be called if checkpointing is off
        x1 = self.x1
        x2 = self.x2

        split_size = beta_features.checkpoint_kernel.value()
        if not split_size:
>           raise RuntimeError(
                "Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. "
                "This is probably a bug in GPyTorch."
            )
E           RuntimeError: Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. This is probably a bug in GPyTorch.

gpytorch/lazy/lazy_evaluated_kernel_tensor.py:254: RuntimeError
__________ TestLazyEvaluatedKernelTensorAdditive.test_t_matmul_matrix __________

self = <test.lazy.test_lazy_evaluated_kernel_tensor.TestLazyEvaluatedKernelTensorAdditive testMethod=test_t_matmul_matrix>

    def test_t_matmul_matrix(self):
        with torch.no_grad():
            linear_op = self.create_linear_op()
            rhs = torch.randn(*linear_op.batch_shape, linear_op.size(-2), 4)
            linear_op_copy = torch.clone(linear_op)
            evaluated = self.evaluate_linear_op(linear_op_copy)
            rhs_evaluated = to_dense(rhs)

            # Test operator
>           res = linear_op._t_matmul(rhs)

/nix/store/ghl89pdl4wyqifz498kzz9n20pph3jsh-python3.10-linear_operator-0.5.1/lib/python3.10/site-packages/linear_operator/test/linear_operator_test_case.py:403: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
/nix/store/ghl89pdl4wyqifz498kzz9n20pph3jsh-python3.10-linear_operator-0.5.1/lib/python3.10/site-packages/linear_operator/operators/_linear_operator.py:914: in _t_matmul
    return self.mT._matmul(rhs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <gpytorch.lazy.lazy_evaluated_kernel_tensor.LazyEvaluatedKernelTensor object at 0x19d962410>
rhs = tensor([[ 0.9383,  0.4889, -0.6731,  0.8728],
        [ 1.1108,  1.2899, -1.4782,  2.5672],
        [-0.4731,  0.3356,  1.5091,  2.0820],
        [-0.4798, -0.4997, -1.0670,  1.1149],
        [-0.1407,  0.8058,  0.3276, -0.7607]])

    def _matmul(self, rhs):
        # This _matmul is defined computes the kernel in chunks
        # It is only used when we are using kernel checkpointing
        # It won't be called if checkpointing is off
        x1 = self.x1
        x2 = self.x2

        split_size = beta_features.checkpoint_kernel.value()
        if not split_size:
>           raise RuntimeError(
                "Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. "
                "This is probably a bug in GPyTorch."
            )
E           RuntimeError: Should not have ended up in LazyEvaluatedKernelTensor._matmul without kernel checkpointing. This is probably a bug in GPyTorch.

gpytorch/lazy/lazy_evaluated_kernel_tensor.py:254: RuntimeError

Expected Behavior

System information

Please complete the following information:

Additional context

None