add mvbart to _interface.py

[miaoqingyu2]

I've got the multivariate logic running and passing the linter (for now I’m using # noqa to call a few internal helpers).

At the moment, MvBart reuses some private methods from Bart. I am thinking about either 1.factoring shared pieces into a public BaseBart (so Bart and MvBart inherit from it), or 2. extending Bart directly to support both univariate and multivariate cases. Which way do you prefer?

I still need to add some unit tests on convergence and sanity checks. will do it soon.

[Gattocrucco]

Do everything in Bart, I guess it should be possible by adding a few helper methods.

P.S.: If you find it easier, make a completely separate version work like you are doing right now, then we can merge them together later.

[Gattocrucco]

Btw I'm about to merge #62 which, amongst other things, adds multivariate support throughout bartz.grove and bartz.mcmcloop, please keep my version when you merge.

[miaoqingyu2]

Do everything in Bart, I guess it should be possible by adding a few helper methods.

P.S.: If you find it easier, make a completely separate version work like you are doing right now, then we can merge them together later.

I've move everything to Bart. btw just curious why in your mcmcloop. evaluate_trace function the outputed y is of dimension n*k? i thought you would prefer k first.

[Gattocrucco]

Ach right, I flipped them, I'm going to fix that.

[Gattocrucco]

I quickly read your code and just noted down the easy things I could notice immediately, I still haven't checked the thing in detail.

[Gattocrucco]

For offset: also support shape (k,)

[Gattocrucco]

For lambda and tau_num: try to work out a good multivariate generalization of this thing.

I guess tau_num can be a scalar, with leaf_prior_cov_inv just being 1/sigma_mu^2 on the diagonal, that part is fine as is.

For error_cov_scale you do multivariate linear regression and get the estimate of the error cov matrix, but then what to do with the chisquared quantile thing used for univariate? I guess there should be something with the maximum eigenvalue of the error covariance matrix, i.e., set the quantile on the weighted outcome average with the largest residual variance.

I haven't read in detail your code so maybe you've already done the above, but I need to go to bed now.

[miaoqingyu2]

For lambda and tau_num: try to work out a good multivariate generalization of this thing.

I guess tau_num can be a scalar, with leaf_prior_cov_inv just being 1/sigma_mu^2 on the diagonal, that part is fine as is.

For error_cov_scale you do multivariate linear regression and get the estimate of the error cov matrix, but then what to do with the chisquared quantile thing used for univariate? I guess there should be something with the maximum eigenvalue of the error covariance matrix, i.e., set the quantile on the weighted outcome average with the largest residual variance.

I haven't read in detail your code so maybe you've already done the above, but I need to go to bed now.

For tau_num: Currently, I allow it to be a vector (length $k$) calculated based on the range of each response dimension individually. This allows the leaf node variance to scale differently for each output if needed.

For error_cov_scale: I currently mimic the univariate logic for each dimension independently. I run the standard quantile calibration on each row of $y$ to get specific $\lambda$ values and place them on the diagonal of the prior matrix. It's just because I wanted to ensure that if $k=1$, the multivariate initialization path yields the exact same priors as the existing univariate path.

However, I was also thinking about using the info from eigenvalues and make the scale matrix reveal some info about correlation of y's. Any preference or comments?

[Gattocrucco]

I merged #64 which created some conflicts with you, sorry! Most of the conflicts look easy to solve, the only tricky one is _predict(). The new version from main is simpler (it does not use the auxiliary _blahblah_chain methods), but you'll have to adapt it again. collapse(out, 0, -1) merges the chain with the sample axes, it means "merge all axes from 0 (included) to -1 (excluded)", i.e., "merge all axes but the last", so it's not already ok for the multivariate version.

[miaoqingyu2]

I merged #64 which created some conflicts with you, sorry! Most of the conflicts look easy to solve, the only tricky one is _predict(). The new version from main is simpler (it does not use the auxiliary _blahblah_chain methods), but you'll have to adapt it again. collapse(out, 0, -1) merges the chain with the sample axes, it means "merge all axes from 0 (included) to -1 (excluded)", i.e., "merge all axes but the last", so it's not already ok for the multivariate version.

i changed it to collapse(out, 0, 2) so that it works for both cases

[Gattocrucco]

Hi, I merged #72 which created a small conflict! Please drop self.ndpost = ... as ndpost is now a property instead of an attribute.