IndexError with sum_smallest

[nicklar]

I am having trouble with an index error when using sum smallest. I am on an m1 Mac using cvxpy 1.2.1 and dccp 1.0.4.

If I just use following constraint my problem converges

x = cp.Variable((3, 3), nonneg = True)
constr = [cp.sum_smallest(x, 2) <= 1]
obj = cp.sum_squares(x)
prob = cp.Problem(cp.Minimize(obj), constr)
prob.solve(method = 'dccp')

While if swap the constraint to grab ndexes:
constr = [cp.sum_smallest(x[0], 2) <= 1]

I get this error.

IndexError Traceback (most recent call last)
/Users/nicholaslarson/Documents/code/auc/dls_mix.ipynb Cell 3' in <cell line: 8>()
6 print(prob.is_dpp())
7 print(dccp.is_dccp(prob))
----> 8 prob.solve(method = 'dccp')

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/cvxpy/problems/problem.py:481, in Problem.solve(self, *args, **kwargs)
479 else:
480 solve_func = Problem._solve
--> 481 return solve_func(self, *args, **kwargs)

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/dccp/problem.py:54, in dccp(self, max_iter, tau, mu, tau_max, solver, ccp_times, max_slack, ep, **kwargs)
50 dccp_ini(
51 self, random=(ccp_times > 1), solver=solver, **kwargs
52 ) # initialization; random initial value is mandatory if ccp_times>1
53 # iterations
---> 54 result_temp = iter_dccp(
55 self, max_iter, tau, mu, tau_max, solver, ep, max_slack, **kwargs
56 )
57 # first iteration
58 if t == 0:

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/dccp/problem.py:241, in iter_dccp(self, max_iter, tau, mu, tau_max, solver, ep, max_slack_tol, **kwargs)
239 count_slack = 0
240 for arg in self.constraints:
--> 241 temp = convexify_constr(arg)
242 if not arg.is_dcp():
243 while temp is None:
244 # damping
File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/dccp/constraint.py:66, in convexify_constr(constr)
64 # left hand concave
65 if constr.args[0].curvature == "CONCAVE":
---> 66 left = linearize(constr.args[0])
67 if left is None:
68 return None

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/dccp/linearize.py:58, in linearize(expr)
54 if tangent is None:
55 raise ValueError(
56 "Cannot linearize non-affine expression with missing variable values."
57 )
---> 58 grad_map = expr.grad
59 for var in expr.variables():
60 if grad_map[var] is None:

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/cvxpy/atoms/atom.py:408, in Atom.grad(self)
404 result = {}
405 for idx, arg in enumerate(self.args):
406 # A dictionary of gradients w.r.t. variables
407 # Partial argument / Partial x.
--> 408 grad_arg = arg.grad
409 for key in grad_arg:
410 # None indicates gradient is not defined.
411 if grad_arg[key] is None or grad_self[idx] is None:

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/cvxpy/atoms/atom.py:402, in Atom.grad(self)
399 arg_values.append(arg.value)
401 # A list of gradients w.r.t. arguments
--> 402 grad_self = self._grad(arg_values)
403 # The Chain rule.
404 result = {}

File ~/miniconda3/envs/dccp_env/lib/python3.9/site-packages/cvxpy/atoms/sum_largest.py:62, in sum_largest._grad(self, values)
60 value = intf.from_2D_to_1D(values[0].flatten().T)
61 indices = np.argsort(-value)[:int(self.k)]
---> 62 D = np.zeros((self.args[0].shape[0]*self.args[0].shape[1], 1))
63 D[indices] = 1
64 return [sp.csc_matrix(D)]

IndexError: tuple index out of range

Any ideas? Thanks for the help!

Nick

[xinyueshen]

Hi @nicklar, thanks for reporting the error. It is from the gradient computation, and while we're fixing it, a quick workaround for your case is to write the constraint as follows.

constr = [cp.sum_smallest(cp.reshape(x[0], (3, 1)), 2) <= 1]

[nicklar]

Thanks for the quick reply and workaround! The workaround has fixed my problem for now!

Really appreciate the help.