Add function to perform partial download of dataset for tests

datasets/flwr_datasets/mock_utils_test.py

@@ -19,7 +19,7 @@
 import random
 import string
 from datetime import datetime, timedelta
-from typing import Any, Dict, List, Set, Tuple, Union
+from typing import Any, Dict, List, Optional, Set, Tuple, Union
 
 import numpy as np
 from PIL import Image
@@ -375,3 +375,47 @@ def _load_mocked_dataset(
     for params in zip(num_rows, split_names):
         dataset_dict[params[1]] = dataset_creation_fnc(params[0])
     return datasets.DatasetDict(dataset_dict)
+
+
+def _download_partial_dataset(
+    dataset_name: str,
+    split_name: str,
+    skip_take_list: List[Tuple[int, int]],
+    subset_name: Optional[str] = None,
+) -> Dataset:
+    """Download a partial dataset.
+
+    This functionality is not supported in the datasets library. This is an informal
+    way of achieving this by using the `streaming=True` and creating a dataset.Dataset
+    from in-memory objects.
+
+    Parameters
+    ----------
+    dataset_name: str
+        Name of the dataset (passed to load_dataset).
+    split_name: str
+        Name of the split (passed to load_dataset) e.g. "train".
+    skip_take_list: List[Tuple[int, int]]
+        The streaming mode has a specific type of accessing the data, the first tuple
+        value is how many samples to skip, the second is how many samples to take. Due
+        to this mechanism, diverse samples can be taken (especially if the dataset is
+        sorted by the natual_id for NaturalIdPartitioner).
+    subset_name: Optional[str]
+        Name of the subset (passed to load_dataset) e.g. "v0.01" for speech_commands.
+
+    Returns
+    -------
+    dataset: Dataset
+        The dataset with the requested samples.
+    """
+    dataset = datasets.load_dataset(
+        dataset_name, name=subset_name, split=split_name, streaming=True
+    )
+    dataset_list = []
+    # It's a list of dict such that each dict represent a single sample of the dataset
+    # The sample is exactly the same as if the full dataset was downloaded and indexed
+    for skip, take in skip_take_list:
+        # dataset.skip(n).take(m) in streaming mode is equivalent (in terms of return)
+        # to the fully downloaded dataset index: dataset[n+1: (n+1 + m)]
+        dataset_list.extend(list(dataset.skip(skip).take(take)))
+    return Dataset.from_list(dataset_list)