Setting Verbose=True in r2HPDHG initialization causes error in mpax.solver_log.display_problem_details

[biltang]

When setting Verbose=True for r2HPDHG, the function will call mpax.solver_log.display_problem_details, which throws an error on line 195, since qp.constraint_matrix.data will throw an AttributeError: 'jaxlib.xla_extension.ArrayImpl' object has no attribute 'data'. I'm not familiar with jax, but wanted to raise this as an issue! Setting Verbose=False gets around this issue of course.

[ZedongPeng]

Hi @biltang, thank you for submitting this issue. I wasn't able to reproduce the error using the "gen-ip054.mps" example. Could you please share your model or a minimal working example so I can investigate further?

[thisiscam]

I am running into the same problem. Here's a minimal repro:

import jax.numpy as np
from mpax import create_lp, r2HPDHG

# Create a simple LP problem:
# Maximize: 3x + 4y
# Subject to:
#    x + 2y <= 14
#    3x - y >= 0
#    x - y <= 2
#    x, y >= 0


def main():
    # Objective coefficients (we negate because create_lp assumes minimization)
    c = np.array([-3.0, -4.0])

    # Equality constraints (none in this example)
    A = np.zeros((0, 2))  # Empty matrix for no equality constraints
    b = np.zeros(0)  # Empty vector for no equality constraints

    # Inequality constraints (Gx <= h format)
    G = np.array([
        [1.0, 2.0],  # x + 2y <= 14
        [-3.0, 1.0],  # 3x - y >= 0 converted to -3x + y <= 0
        [1.0, -1.0],  # x - y <= 2
    ])

    h = np.array([14.0, 0.0, 2.0])

    # Lower bounds for variables
    l = np.array([0.0, 0.0])

    # Upper bounds (none specified)
    u = np.ones(2) * 1e10  # Large value instead of infinity

    # Create the LP problem
    lp = create_lp(c, A, b, G, h, l, u, use_sparse_matrix=False)

    # Solve using r2HPDHG
    solver = r2HPDHG(eps_abs=1e-6, eps_rel=1e-6, verbose=True)
    result = solver.optimize(lp)

    print("Optimization finished!")
    print(
        f"Solution: x = {result.primal_solution[0]:.6f}, y = {result.primal_solution[1]:.6f}"
    )
    print(f"Objective value: {-np.dot(c, result.primal_solution):.6f}"
          )  # Negate back for maximization


if __name__ == "__main__":
    main()

I believe the ArrayImpl is already deprecated in jax in any case.

[ZedongPeng]

Hi @thisiscam. Thank you for the example. I will look into it.

[fromseto]

The error is from the difference between a standard jax array and a sparse array in jax.experimental.sparse. The former does not have data attribute, while the latter has. You can pass use_sparse_matrix=True to fix it but it is an bug in the implementation.

[ZedongPeng]

Thank you @fromseto for pointing it out. Indeed, it results from the difference between the standard JAX array and the sparse array. I will fix it hopefully tomorrow.

[ZedongPeng]

Hi @fromseto @thisiscam @biltang . Thank you all for identifying and helping to fix this bug. It has been resolved in version v0.2.4.