is variable feature from NodeBipartiteObs sorted by unique index?

[th-yoon]

Hi, I have a question about variable features extracted from NodeBipartiteObs module.
I see that it gets columns as in this line , and it depends on SCIPgetLPCols() function in SCIP C code.
Does it ensure that the variable features are sorted row-wise based on unique index of each variable?
Thanks.

[gasse]

Hi @th-yoon ,

The variable features in NodeBipartiteObs indeed are extracted for all variables in the current node's LP (obtained from SCIPgetLPCols()). They are also registered in the same order, and thus can be indexed using the column position (var.getCol().getLPPos()) of the variables. So, if you have a list of variables and you want to get their associated features in NodeBipartiteObs, you can use something like this:

import numpy as np

obs = ...  # NodeBipartiteObs

m = model.as_pyscipopt()
vars = m.getVars(transformed=True)
vars_in_lp = [var for var in vars if var.isInLP()]
lppos_vars_in_lp = np.asarray([var.getCol().getLPPos() for var in vars_in_lp])
feats_vars_in_lp = obs.column_features[lppos_vars_in_lp]

See also this post which is very related:
ds4dm/ml4co-competition#24 (comment)

Hope that helps,

Best,
Maxime

[th-yoon]

Thanks @gasse , it helped a lot!