There are currently five tests that fail when running pytest on a GPU-enabled machine. These are all numerical comparison failures and are all (to some extent) understood.
In test_GaussianProcess.py, the tests
test_GaussianProcessGPU_theta[0.0-1.0]
test_GaussianProcessGPU_theta[adaptive-0.0]
test_GaussianProcessGPU_predict
test_GaussianProcessGPU_predict_nugget
fail because the way that the "expected" values are calculated directly from the python Kernel has not been updated to take account of the new ways that the parameters are transformed. The equivalent way of calculating the expected values for the CPU implementation now uses the design matrix, which is not easily available to the current GPU implementation.
In test_fitting.py, the test test_fit_GP_MAP_GPU fails because of a numerical comparison between the fitted theta values and some "expected" ones, which in the CPU equivalent, are just provided by a patched function. For the GPU/C++ implementation there is not an obvious python function to patch in this way.
It is expected that these tests will be updated/fixed in #211.
There are currently five tests that fail when running pytest on a GPU-enabled machine. These are all numerical comparison failures and are all (to some extent) understood.
In
test_GaussianProcess.py, the teststest_GaussianProcessGPU_theta[0.0-1.0]test_GaussianProcessGPU_theta[adaptive-0.0]test_GaussianProcessGPU_predicttest_GaussianProcessGPU_predict_nuggetfail because the way that the "expected" values are calculated directly from the python Kernel has not been updated to take account of the new ways that the parameters are transformed. The equivalent way of calculating the expected values for the CPU implementation now uses the design matrix, which is not easily available to the current GPU implementation.In
test_fitting.py, the testtest_fit_GP_MAP_GPUfails because of a numerical comparison between the fitted theta values and some "expected" ones, which in the CPU equivalent, are just provided by a patched function. For the GPU/C++ implementation there is not an obvious python function to patch in this way.It is expected that these tests will be updated/fixed in #211.