cvxpy and cvxpylayers solutions disagree for cvxpy==1.3.0 and cvxpylayers==0.1.4

[Chris-Sorini]

In an anaconda environment with cvxpy==1.3.0 and cvxpylayers==0.1.4 I noticed that cvxpylayers and cvxpy result in different solutions for the following problem:

x = cp.Variable(2) xc = cp.Parameter(2) xc.value = np.array([1.0, 2.5]) xc_torch = torch.tensor([1.0, 2.5]) obj = cp.Minimize((x[0] - xc[0])2 + (x[1] - xc[1])2) cons = [(x[0] - 2 x[1] + 2) >= 0 , (-x[0] - 2 x[1] + 6 >= 0), (-x[0] + 2 * x[1] + 2) >=0] prob = cp.Problem(obj, cons) cvxpylayer = CvxpyLayer(problem=prob, parameters=[xc], variables=[x])

prob.solve(solver_args={"solve_method": "ECOS"}) outputs 0.8 for i in prob.variables(): print(i.value) outputs the optimal variable values: [1.4, 1.7]

cvxpylayer(xc_torch, solver_args={"solve_method": "ECOS"}) outputs:

• (tensor([2.9926, 1.4963]),) with cvxpy==1.3.0 (w/ cvxpylayers==0.1.4)
• (tensor([1.4000, 1.7000]),) with cvxpy==1.2.3 (w/ cvxpylayers==0.1.4)

The solutions agree when I use cvxpy==1.2.3, but not with cvxpy==1.3.0 (in both cases using cvxpylayers==0.1.4). Also note that the example on the github readme for cvxpylayers did not show the same issue for either cvxpy==1.2.3 or 1.3.0. I am going to stick with cvxpy==1.2.3 for now, but does anyone know what is causing this issue for cvxpy==1.3.0? Thanks!

[jtorde]

I am having the same issue (with a different optimization problem):

• cvxpylayers==0.1.5 and cvxpy==1.2.3 works (both results agree)
• cvxpylayers==0.1.5 and cvxpy==1.3.0 generate different results

For now I switched to cvxpy==1.2.3

[bamos]

Hi, this may be due to cvxpylayers/cvxpy defaulting to using SCS with a quadratic objective without having full support for it (which I believe may be the difference between 1.2.3 and 1.3.0 here). We just merged https://github.com/cvxgrp/cvxpylayers/pull/142 into the main branch of cvxpylayers to disable this behavior. Can you please try running this latest version of cvxpylayers (in the main branch here) with the latest version of cvxpy (or 1.3.0) and let us know if this is still an issue?

[jtorde]

With the latest versions of both cvxpy (cvxpy==1.3.1) and cvxpylayers (cvxpylayers=0.1.6) this issue seems to be fixed:

import cvxpy as cp
import numpy as np
import torch
from cvxpylayers.torch import CvxpyLayer

x = cp.Variable(2)
xc = cp.Parameter(2)
xc.value = np.array([1.0, 2.5])
xc_torch = torch.tensor([1.0, 2.5])
obj = cp.Minimize((x[0] - xc[0])**2 + (x[1] - xc[1])**2)
cons = [(x[0] - 2 * x[1] + 2) >= 0 ,
(-x[0] - 2 * x[1] + 6 >= 0),
(-x[0] + 2 * x[1] + 2) >=0]
prob = cp.Problem(obj, cons)
cvxpylayer = CvxpyLayer(problem=prob, parameters=[xc], variables=[x])

prob.solve(solver='ECOS')

for i in prob.variables(): 
        print(i.value) 

z=cvxpylayer(xc_torch, solver_args={"solve_method": "ECOS"})
print(z)

Both return [1.4, 1.7]