Covariance #2
Covariance #2
Conversation
|
The compiler complained about the lifetime of type |
There was a problem hiding this comment.
The compiler complained about the lifetime of type
Aand to "calm it" I slapped'staticin front ofA. Can you help me understand what the issue was?
The issue was the A: LinalgScalar bound on .mean_axis(), since LinalgScalar requires 'static. I've submitted rust-ndarray/ndarray#491 to remove the unnecessary bounds.
Ideally, this should be implemented for any n-dimensional array with n>=2, treating two user-specified axes as
random_variableaxis andobservationaxis, while using the remaining ones as batch axes. To implement the generalized version we have to be able to do an efficient batched matrix multiplication - the implementation of this routine is currently missing in ndarray.
IMO, only supporting 2-D arrays is fine for an initial version. This type of thing occurs for a lot of other methods too (e.g. quantiles for n-D arrays or changing .mean_axis() to .mean_axes()).
Feel free to merge this when you think it's ready; it looks good to me. It would be nice to choose the "squash and merge" option to keep the history clean.
I agree - it was more of a remainder for the future :) I'll merge this and open issue to track the status of |
A basic implementation for the computation of the covariance matrix starting from a 2-dimensional array of observations.
Ideally, this should be implemented for any n-dimensional array with n>=2, treating two user-specified axes as
random_variableaxis andobservationaxis, while using the remaining ones as batch axes. To implement the generalized version we have to be able to do an efficient batched matrix multiplication - the implementation of this routine is currently missing in ndarray.