Add rescaled version of harmonics dataset (#188)

deep_bottleneck/datasets/harmonics_zero_mean.py

@@ -0,0 +1,23 @@
+from deep_bottleneck.datasets.base_dataset import Dataset
+from deep_bottleneck.datasets.harmonics import load as load_original_harmonics
+
+
+def load(nb_dir='') -> Dataset:
+    """Load the Information Bottleneck harmonics dataset with mean 0, i.e.
+    the original harmonics dataset but 0.5 subtracted
+
+    Returns:
+        The zero centered harmonics dataset.
+    """
+
+    data = load_original_harmonics()
+
+    X_train = data.train.examples - 0.5
+    y_train = data.train.labels
+    X_test = data.test.examples - 0.5
+    y_test = data.test.labels
+    n_classes = data.n_classes
+
+    shifted_dataset = Dataset.from_labelled_subset(X_train, y_train, X_test, y_test, n_classes)
+
+    return shifted_dataset