Check there is no overlap between split datetimes. Fixes #299

Author: JackKelly

nowcasting_dataset/config/model.py

@@ -19,12 +19,14 @@
 from pydantic import BaseModel, Field
 from pydantic import validator, root_validator
 
-from nowcasting_dataset.consts import NWP_VARIABLE_NAMES
+# nowcasting_dataset imports
 from nowcasting_dataset.consts import (
+    NWP_VARIABLE_NAMES,
     SAT_VARIABLE_NAMES,
     DEFAULT_N_GSP_PER_EXAMPLE,
     DEFAULT_N_PV_SYSTEMS_PER_EXAMPLE,
 )
+from nowcasting_dataset.dataset.split import split
 
 
 IMAGE_SIZE_PIXELS_FIELD = Field(64, description="The number of pixels of the region of interest.")
@@ -278,6 +280,12 @@ class Process(BaseModel):
             " return data at 11:30, 12:00, 12:30, and 13:00."
         ),
     )
+    split_method: split.SplitMethod = Field(
+        split.SplitMethod.DAY,
+        description=(
+            "The method used to split the t0 datetimes into train, validation and test sets."
+        ),
+    )
     upload_every_n_batches: int = Field(
         16,
         description=(

nowcasting_dataset/dataset/split/split.py

@@ -39,6 +39,7 @@ class SplitName(Enum):
     TEST = "test"
 
 
+# Create a namedtuple for storing split t0 datetimes.
 SplitData = namedtuple(
     typename="SplitData",
     field_names=[SplitName.TRAIN.value, SplitName.VALIDATION.value, SplitName.TEST.value],
@@ -186,4 +187,9 @@ def split_data(
         f"test has {len(test_datetimes):,d} t0 datetimes."
     )
 
+    # Check there's no overlap.
+    assert len(train_datetimes.intersection(validation_datetimes)) == 0
+    assert len(train_datetimes.intersection(test_datetimes)) == 0
+    assert len(test_datetimes.intersection(validation_datetimes)) == 0
+
     return SplitData(train=train_datetimes, validation=validation_datetimes, test=test_datetimes)

nowcasting_dataset/manager.py

@@ -11,6 +11,7 @@
 from nowcasting_dataset import config
 from nowcasting_dataset.filesystem import utils as nd_fs_utils
 from nowcasting_dataset.data_sources import MAP_DATA_SOURCE_NAME_TO_CLASS, ALL_DATA_SOURCE_NAMES
+from nowcasting_dataset.dataset.split import split
 
 logger = logging.getLogger(__name__)
 
@@ -116,6 +117,9 @@ def create_files_specifying_spatial_and_temporal_locations_of_each_example_if_ne
         t0_datetimes = self.get_t0_datetimes_across_all_data_sources(
             freq=self.config.process.t0_datetime_frequency
         )
+        split_t0_datetimes = split.split_data(
+            datetimes=t0_datetimes, method=self.config.process.split_method
+        )
 
     def _locations_csv_file_exists(self) -> bool:
         "Check if filepath/train/spatial_and_temporal_locations_of_each_example.csv exists"
@@ -202,20 +206,11 @@ def sample_spatial_and_temporal_locations_for_examples(
             Each row of each the DataFrame specifies the position of each example, using
             columns: 't0_datetime_UTC', 'x_center_OSGB', 'y_center_OSGB'.
         """
-        # This code is for backwards-compatibility with code which expects the first DataSource
-        # in the list to be used to define which DataSource defines the spatial location.
-        # TODO: Remove this try block after implementing issue #213.
-        try:
-            data_source_which_defines_geospatial_locations = (
-                self.data_source_which_defines_geospatial_locations
-            )
-        except AttributeError:
-            data_source_which_defines_geospatial_locations = self[0]
-
         shuffled_t0_datetimes = np.random.choice(t0_datetimes, size=n_examples)
-        x_locations, y_locations = data_source_which_defines_geospatial_locations.get_locations(
-            shuffled_t0_datetimes
-        )
+        (
+            x_locations,
+            y_locations,
+        ) = self.data_source_which_defines_geospatial_locations.get_locations(shuffled_t0_datetimes)
         return pd.DataFrame(
             {
                 "t0_datetime_UTC": shuffled_t0_datetimes,