from pytorch_forecasting.data.encoders import EncoderNormalizer
import torch

# generate a random sequence with mean -0.5 and unit std
x = torch.randn(60) - 0.5
print('x mean and std:', x.mean(), x.std())

# create a standard encoder with centering
encoder = EncoderNormalizer(
    method='standard',
    center=True,
    max_length=None,
    transformation=None,
    method_kwargs={}
)

# fit transform
y = encoder.fit_transform(x)

# center_ and scale_ match mean and std of x
print('center_ and scale_:', encoder.center_, encoder.scale_)

# the transformed is what is expected
print('error:', (y - (x - x.mean()) / x.std()).abs().max())


# create a standard encoder with *NO* centering
encoder = EncoderNormalizer(
    method='standard',
    center=False,
    max_length=None,
    transformation=None,
    method_kwargs={}
)

# fit transform
y = encoder.fit_transform(x)

# center_ is zero as expected, but scale_ *is not* the std of x, it is the *mean*
print('center_ and scale_:', encoder.center_, encoder.scale_)

# the transformed is *not* what is expected
print('error:', (y - x / x.std()).abs().max())