Sh/ep device

setup.py

@@ -34,9 +34,7 @@ def read(fname):
         "tqdm",
         "pyyaml",
     ],
-    extras_require={
-        "test": ["pytest", "pytest-console-scripts", "pytest-datadir"]
-    },
+    extras_require={"test": ["pytest", "pytest-console-scripts", "pytest-datadir"]},
     license="MIT",
     description="SchNetPack - Deep Neural Networks for Atomistic Systems",
     long_description="""

src/schnetpack/utils/script_utils/parsing.py

@@ -143,6 +143,12 @@ def get_mode_parsers():
         help="Number of checkpoints that will be stored (default: %(default)s)",
         default=3,
     )
+    train_parser.add_argument(
+        "--environment_provider_device",
+        help="Choose device for environment providers. It is recommended to keep CPU. (default: %(default)s)",
+        choices=["cpu", "cuda"],
+        default="cpu",
+    )
 
     # evaluation parser
     eval_parser = ArgumentParser(add_help=False)
@@ -173,6 +179,12 @@ def get_mode_parsers():
     eval_parser.add_argument(
         "--overwrite", help="Remove previous evaluation files", action="store_true"
     )
+    eval_parser.add_argument(
+        "--environment_provider_device",
+        help="Choose device for environment providers. It is recommended to keep CPU. (default: %(default)s)",
+        choices=["cpu", "cuda"],
+        default="cpu",
+    )
 
     return json_parser, train_parser, eval_parser
 

src/schnetpack/utils/script_utils/settings.py

@@ -135,7 +135,7 @@ def get_environment_provider(args, device):
         return spk.environment.AseEnvironmentProvider(cutoff=args.cutoff)
     elif args.environment_provider == "torch":
         return spk.environment.TorchEnvironmentProvider(
-            cutoff=args.cutoff, device="cpu"
+            cutoff=args.cutoff, device=device
         )
     else:
         raise NotImplementedError

src/scripts/spk_run.py

@@ -26,9 +26,10 @@ def main(args):
     train_args = setup_run(args)
 
     device = torch.device("cuda" if args.cuda else "cpu")
+    ep_device = torch.device(args.environment_provider_device)
 
     # get dataset
-    environment_provider = get_environment_provider(train_args, device=device)
+    environment_provider = get_environment_provider(train_args, device=ep_device)
     dataset = get_dataset(train_args, environment_provider=environment_provider)
 
     # get dataloaders