Issue with higher-dimensional constraints

[paddy-r]

I'm interested in using IPFN to solve 3D problems with 2D constraints. I simplified and adapted the example given in the readme, but for three sets of 2D constraints - see below. I've added a third constraint for the zx dimension. However, the algorithm does not give the correct solution.

# Seed data
m = np.random.rand(2, 4, 3)

# Constraints
xijp = np.array([[9, 17, 19, 7], [11, 13, 16, 8]])  # xy, 2x4
xpjk = np.array([[7, 9, 4], [8, 12, 10], [15, 12, 8], [5, 7, 3]])  # yz, 4x3
xipk = np.array([[10, 20], [20, 20], [10, 20]])  # zx, 3x2

# Create arguments
aggregates = [xijp, xpjk, xipk]
dimensions = [[0, 1], [1, 2], [2, 0]]

# Run IPF algorithm and compare slice
IPF = ipfn.ipfn(m, aggregates, dimensions, verbose=2)
m = IPF.iteration()

print('Slice should be:', xijp[0,0])
print('IPF gives:', m[0][0, 0, :].sum())

print('Target sum (xy):', xijp.sum())
print('IPF gives:', m[0].sum())

print('Convergence sequence:\n', m[-1].values.tolist())

This gives the following output and a conv value of 1:

ipfn converged: convergence_rate not updating or below rate_tolerance
Slice should be: 9
IPF gives: 5.20361253053332
Target sum (xy): 100
IPF gives: 60.0
Convergence sequence:
 [[1.0], [1.0]]

Happy to contribute to coding the solution, although I couldn't pinpoint the source of the issue.

EDIT: Confirmed working with same set of constraints using R package MIPFP from here.