Normalization

Currently, you can choose from four normalization strategies:

• std: Note that the normalization is performed field-wise (indices are combined according to the field scheme). After normalization the standard deviation over all elements of a field array (including time dimension) will be one.
• mean-std: Note that the normalization is performed field index-wise (each field index is considered separately). After normalization the mean over all elements of a field index array (including time dimension) will be zero. Furthermore, the standard deviation over this array will be one.
• zero-to-one: Each field index array is scaled linearly so that the smallest element is zero and the largest element is one.
• minus-one-to-one: Each field index array is scaled linearly so that the smallest element is minus one and the largest element is one.

To specify the normalization strategy for data and constants with normalize_data and normalize_const respectively:

loader = Dataloader(
    "my-dataset",
    time_steps=5,
    normalize_data="std",
    normalize_const="mean-std"
)