[feature request] Exponential moving average (EMA) of a tensor across a dimension

[vadimkantorov]

🚀 The feature, motivation and pitch

If the dimension size is high, Python loop may be inefficient and maybe inplace computation graph (for efficiency) could confuse TorchScript.

In pandas this exists as ewm() function: https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.ewm.html

In NumPy various reimpl methods exist (recurrent; truncated based on convolution): https://stackoverflow.com/questions/42869495/numpy-version-of-exponential-weighted-moving-average-equivalent-to-pandas-ewm

Alternatives

Special APIs for returning convolution weights, e.g. https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.gaussian_filter1d.html, and the same for exponential weights. Then have a recipe/example in docs of using these weights to create an approximate exponential moving average via convolution.

Although maybe given that gaussian / ewm filtering is very popular, both APIs would be useful.

Additional context

No response

cc @mruberry @rgommers

[vadimkantorov]

probably some API designs for such function already exist in PyTorch time series packages...