Representation functions calculation in fit are necessary in QUANTTranformer?

[danielnowakassis]

In QUANTTranformer

These representation functions are calculated for training data :

representation_functions = (
          lambda X: X,
          lambda X: F.avg_pool1d(F.pad(X.diff(), (2, 2), "replicate"), 5, 1),
          lambda X: X.diff(n=2),
          lambda X: torch.fft.rfft(X).abs(),
)

And only the shape of the results are extracted :

for function in representation_functions:
            Z = function(X)
            self.intervals_.append(self._make_intervals(input_length=Z.shape[-1]))

Can the shape of the representation functions be specified instead of unnecessary calculations for training data?

[MatthewMiddlehurst]

Hi, converted this to an issue as it seems more relevant there. Seems like a reasonable improvement if it can be replaced. Likely not a big difference in efficiency for most cases.