EXA use himalaya.backend.set_backend(..., on_error='warn')

Author: TomDLT

tutorials/movies_3T/02_plot_wordnet_model.py

@@ -134,12 +134,19 @@
 # target case, ``GridSearchCV`` can only optimize e.g. the mean score over
 # targets. Here, we want to find a different optimal hyperparameter per
 # target/voxel, so we use ``himalaya``'s ``KernelRidgeCV`` instead.
-# Moreover, ``himalaya`` implements different computational backends; we will
-# use the "torch_cuda" backend to use fast computations on GPU.
 from himalaya.kernel_ridge import KernelRidgeCV
 
+###############################################################################
+# Moreover, ``himalaya`` implements different computational backends, including
+# GPU backends. The available GPU backends are "torch_cuda" and "cupy". (These
+# backends are only available if you installed the corresponding package with
+# CUDA enabled. Check the pytorch/cupy documentation for install instructions.)
+#
+# Here we use the "torch_cuda" backend, but if the import fails we continue
+# with the default "numpy" backend. The "numpy" backend is expected to be
+# slower since it only uses the CPU.
 from himalaya.backend import set_backend
-backend = set_backend("torch_cuda")
+backend = set_backend("torch_cuda", on_error="warn")
 
 ###############################################################################
 # The scale of the regularization hyperparameter ``alpha`` is unknown, so we

tutorials/movies_3T/03_plot_motion_energy_model.py

@@ -106,7 +106,7 @@
 from voxelwise_tutorials.delayer import Delayer
 from himalaya.kernel_ridge import KernelRidgeCV
 from himalaya.backend import set_backend
-backend = set_backend("torch_cuda")
+backend = set_backend("torch_cuda", on_error="warn")
 
 alphas = np.logspace(1, 20, 20)
 

tutorials/movies_3T/04_plot_banded_ridge_model.py

@@ -101,7 +101,7 @@
 from sklearn.preprocessing import StandardScaler
 from voxelwise_tutorials.delayer import Delayer
 from himalaya.backend import set_backend
-backend = set_backend("torch_cuda")
+backend = set_backend("torch_cuda", on_error="warn")
 
 ###############################################################################
 # To fit the banded ridge model, we use ``himalaya``'s

tutorials/movies_4T/02_plot_ridge_model.py

@@ -150,15 +150,17 @@
 from voxelwise_tutorials.delayer import Delayer
 
 ###############################################################################
-# We set himalaya's backend to "torch_cuda" to fit the model using GPU.
-# The available backends are:
+# The package``himalaya`` implements different computational backends,
+# including GPU backends. The available GPU backends are "torch_cuda" and
+# "cupy". (These backends are only available if you installed the corresponding
+# package with CUDA enabled. Check the pytorch/cupy documentation for install
+# instructions.)
 #
-# - "numpy" (CPU) (default)
-# - "torch" (CPU)
-# - "torch_cuda" (GPU)
-# - "cupy" (GPU)
+# Here we use the "torch_cuda" backend, but if the import fails we continue
+# with the default "numpy" backend. The "numpy" backend is expected to be
+# slower since it only uses the CPU.
 from himalaya.backend import set_backend
-backend = set_backend("torch_cuda")
+backend = set_backend("torch_cuda", on_error="warn")
 
 ###############################################################################
 # The scale of the regularization hyperparameter alpha is unknown, so we use