There's often a need to fit a single model to several datasets, where only e.g. the normalisation is changing.
At the moment, the model is evaluated once for each dataset, but there's much room for improvement here: we need only evaluate the model once and can use different normalisations for each one. The caveat is; if the normalisation is being used in a convolution model, the effect is potentially non-linear, but in theory we have all the information to know that at compile time.
For now:
A solution would be to introduce a new data wrapper that detects if the domains are all the same / overlapping, and then uses the transformer function to effectively copy the model n times and apply the normalisations.
Abuse or implement an alternative AutoCache
Just concatenate all the datasets into a new dataset wrapper that implements the dataset API so the model is just evaluated once (but then can't adjust normalisations).
There's often a need to fit a single model to several datasets, where only e.g. the normalisation is changing.
At the moment, the model is evaluated once for each dataset, but there's much room for improvement here: we need only evaluate the model once and can use different normalisations for each one. The caveat is; if the normalisation is being used in a convolution model, the effect is potentially non-linear, but in theory we have all the information to know that at compile time.
For now:
ntimes and apply the normalisations.AutoCache