Ignore local dataset argument when running inside AML runs (#238)

This PR modifies mount_or_download_dataset such that we ignore the `local_dataset` argument inside AML runs (only used for local runs).

InnerEye/ML/deep_learning_config.py

@@ -190,8 +190,7 @@ class DeepLearningConfig(GenericConfig, CudaAwareConfig):
                                               "usually set from the class name.")
 
     random_seed: int = param.Integer(42, doc="The seed to use for all random number generators.")
-    azure_dataset_id: Optional[str] = param.String(None, allow_None=True,
+    azure_dataset_id: str = param.String(doc="If provided, the ID of the dataset to use. This dataset must exist as a "
-                                                   doc="The ID of the dataset to use. This dataset must exist as a "
                                                        "folder of the same name in the 'datasets' "
                                                        "container in the datasets storage account.")
     local_dataset: Optional[Path] = param.ClassSelector(class_=Path,

InnerEye/ML/run_ml.py

@@ -393,23 +393,27 @@ class MLRunner:
         mounted or downloaded.
         Returns the path of the dataset on the executing machine.
         """
+        azure_dataset_id = self.model_config.azure_dataset_id
+
+        if is_offline_run_context(RUN_CONTEXT):
+            # The present run is outside of AzureML: If local_dataset is set, use that as the path to the data.
+            # Otherwise, download the dataset specified by the azure_dataset_id
             local_dataset = self.model_config.local_dataset
+            if (not azure_dataset_id) and (local_dataset is None):
+                raise ValueError("The model must contain either local_dataset or azure_dataset_id.")
             if local_dataset:
                 expected_dir = Path(local_dataset)
                 if not expected_dir.is_dir():
                     raise FileNotFoundError(f"The model uses a dataset in {expected_dir}, but that does not exist.")
                 logging.info(f"Model training will use the local dataset provided in {expected_dir}")
                 return expected_dir
-        azure_dataset_id = self.model_config.azure_dataset_id
-        if not azure_dataset_id:
-            raise ValueError("The model must contain either local_dataset or azure_dataset_id.")
-        if is_offline_run_context(RUN_CONTEXT):
-            # The present run is outside of AzureML: If local_dataset is set, use that as the path to the data.
-            # Otherwise, download the dataset specified by the azure_dataset_id
             return download_dataset(azure_dataset_id=azure_dataset_id,
                                     target_folder=self.project_root / fixed_paths.DATASETS_DIR_NAME,
                                     azure_config=self.azure_config)
+
         # Inside of AzureML, datasets can be either mounted or downloaded.
+        if not azure_dataset_id:
+            raise ValueError("The model must contain azure_dataset_id for running on AML")
         mounted = try_to_mount_input_dataset(RUN_CONTEXT)
         if not mounted:
             raise ValueError("Unable to mount or download input dataset.")

Tests/ML/models/architectures/test_image_encoder_with_mlp.py

@@ -61,6 +61,7 @@ class ImageEncoder(ScalarModelBase):
             l_rate=1e-1,
             use_mixed_precision=True,
             aggregation_type=aggregation_type,
+            azure_dataset_id="test-dataset",
             **kwargs
         )
         self.encode_channels_jointly = encode_channels_jointly

azure-pipelines/build-pr.yml

@@ -2,7 +2,7 @@ name: PR-$(Date:yyyyMMdd)$(Rev:-r)
 variables:
   model: 'BasicModel2Epochs'
   train: 'True'
-  more_switches: '--log_level=DEBUG'
+  more_switches: '--log_level=DEBUG --local_dataset=F:\local'
   run_recovery_id: ''
   tags: 'PR'
   user_friendly_name: 'PR build'