Allow to force to a single GPU

dan/ocr/manager/training.py

@@ -101,7 +101,7 @@ class GenericTrainingManager:
 
     def init_hardware_config(self):
         # Debug mode
-        if self.device_params["force"] == "cpu":
+        if self.device_params["force"] not in [None, "cuda"] or not torch.cuda.is_available():
             self.device_params["use_ddp"] = False
             self.device_params["use_amp"] = False
 

dan/ocr/train.py

@@ -146,7 +146,8 @@ def serialize_config(config):
 def start_training(config, mlflow_logging: bool) -> None:
     if (
         config["training"]["device"]["use_ddp"]
-        and config["training"]["device"]["force"] != "cpu"
+        and config["training"]["device"]["force"] in [None, "cuda"]
+        and torch.cuda.is_available()
     ):
         mp.spawn(
             train_and_test,