cvxpylayers
cvxpylayers copied to clipboard
cvxgrp
Calling `jax.jacobian` results in error
Hi,
Thanks for the cool package!
I've just started playing around with it on a QP I need to differentiate through. I'm getting the following error when calling jax.jacobian on the layer. The forward pass works fine, and finds a nontrivial solution as indicated by the verbose output of the solver.
Any help here would be appreciated!
Solver output:
----------------------------------------------------------------------------
SCS v2.1.4 - Splitting Conic Solver
(c) Brendan O'Donoghue, Stanford University, 2012
----------------------------------------------------------------------------
Lin-sys: sparse-direct, nnz in A = 20101
eps = 1.00e-05, alpha = 1.50, max_iters = 5000, normalize = 1, scale = 1.00
acceleration_lookback = 10, rho_x = 1.00e-03
Variables n = 102, constraints m = 202
Cones: primal zero / dual free vars: 100
soc vars: 102, soc blks: 1
Setup time: 4.14e-02s
----------------------------------------------------------------------------
Iter | pri res | dua res | rel gap | pri obj | dua obj | kap/tau | time (s)
----------------------------------------------------------------------------
0| 2.91e+19 2.29e+19 1.00e+00 -6.29e+19 1.45e+18 5.11e+19 2.14e-02
100| 3.48e-05 1.08e-05 5.48e-05 5.81e-02 5.82e-02 3.46e-18 4.10e-02
200| 3.62e-03 1.26e-03 1.62e-03 4.83e-02 4.65e-02 2.10e-17 5.60e-02
300| 1.37e-01 1.28e-02 3.03e-02 1.30e-02 4.51e-02 2.48e-17 6.72e-02
400| 5.11e-02 1.60e-02 2.01e-02 7.96e-02 5.67e-02 2.13e-17 7.94e-02
500| 1.51e-02 5.96e-03 5.18e-03 4.17e-02 3.61e-02 7.22e-18 9.34e-02
600| 4.03e-04 1.67e-04 1.32e-04 3.91e-02 3.92e-02 3.66e-17 1.05e-01
700| 1.25e-03 2.41e-04 1.02e-03 3.82e-02 3.93e-02 1.07e-18 1.18e-01
800| 8.28e-03 2.12e-03 2.41e-03 3.25e-02 3.51e-02 2.42e-17 1.30e-01
900| 9.51e-04 4.79e-04 1.53e-04 3.87e-02 3.88e-02 3.10e-17 1.42e-01
1000| 1.73e-03 1.01e-03 1.14e-03 3.82e-02 3.95e-02 7.56e-18 1.58e-01
1100| 2.17e-03 1.14e-03 5.81e-04 3.79e-02 3.86e-02 6.94e-18 1.76e-01
1200| 2.57e-02 9.40e-03 5.71e-03 3.29e-02 3.91e-02 4.66e-17 1.93e-01
1300| 5.11e-03 2.51e-03 7.59e-03 4.72e-02 3.90e-02 5.41e-18 2.07e-01
1400| 1.46e-03 3.23e-04 5.90e-06 3.68e-02 3.68e-02 3.53e-18 2.19e-01
1500| 2.58e-03 2.64e-04 1.59e-03 3.64e-02 3.81e-02 5.11e-18 2.36e-01
1600| 1.07e-03 5.72e-04 2.59e-04 3.77e-02 3.80e-02 1.17e-17 2.49e-01
1700| 1.21e-03 5.51e-04 9.79e-04 3.82e-02 3.72e-02 2.62e-17 2.63e-01
1800| 6.70e-04 1.30e-04 2.71e-04 3.71e-02 3.74e-02 3.90e-17 2.76e-01
1900| 2.35e-02 1.15e-02 4.02e-03 4.06e-02 3.62e-02 3.50e-17 2.88e-01
2000| 1.21e-03 1.82e-04 1.99e-03 3.83e-02 3.61e-02 1.90e-17 3.01e-01
2100| 4.19e-02 1.53e-02 1.78e-02 5.94e-02 3.98e-02 4.11e-19 3.12e-01
2200| 9.61e-04 3.36e-04 5.88e-04 3.86e-02 3.79e-02 1.78e-17 3.27e-01
2300| 1.10e-03 5.49e-04 7.74e-04 3.86e-02 3.78e-02 2.92e-17 3.40e-01
2400| 5.45e-03 1.52e-03 2.36e-03 4.30e-02 4.04e-02 7.82e-17 3.57e-01
2500| 1.66e-02 9.82e-03 1.03e-02 5.18e-02 4.06e-02 3.10e-17 3.69e-01
2600| 1.31e-04 7.45e-05 9.50e-05 3.75e-02 3.76e-02 1.97e-17 3.86e-01
2700| 4.12e-04 2.23e-04 4.29e-04 3.79e-02 3.75e-02 5.47e-18 3.99e-01
2800| 2.99e-04 9.80e-05 3.23e-05 3.76e-02 3.76e-02 1.43e-17 4.11e-01
2900| 1.62e-03 9.50e-04 2.36e-03 3.44e-02 3.69e-02 2.24e-17 4.23e-01
3000| 5.68e-04 3.09e-04 3.79e-04 3.79e-02 3.75e-02 4.01e-18 4.35e-01
3100| 1.02e-02 5.00e-03 1.26e-03 4.00e-02 4.14e-02 2.91e-17 4.47e-01
3200| 9.33e-03 4.59e-03 3.19e-03 4.10e-02 3.75e-02 6.83e-18 4.60e-01
3300| 1.20e-03 6.14e-04 7.47e-04 3.67e-02 3.75e-02 9.50e-18 4.72e-01
3400| 4.41e-02 2.47e-02 2.23e-02 1.80e-02 4.17e-02 2.27e-17 4.84e-01
3500| 6.01e-03 3.32e-03 2.26e-03 2.76e-02 2.52e-02 5.72e-17 4.96e-01
3600| 7.45e-04 8.20e-05 9.78e-05 3.70e-02 3.69e-02 1.19e-17 5.08e-01
3700| 1.16e-04 4.97e-05 1.93e-04 3.72e-02 3.70e-02 2.38e-17 5.20e-01
3800| 3.30e-03 1.48e-03 5.65e-04 3.74e-02 3.67e-02 6.22e-18 5.32e-01
3900| 7.25e-03 3.55e-03 1.21e-03 3.47e-02 3.35e-02 3.64e-18 5.44e-01
4000| 1.61e-03 3.41e-04 4.68e-04 3.70e-02 3.65e-02 2.64e-17 5.60e-01
4100| 5.24e-02 1.16e-02 1.15e-02 6.02e-02 4.74e-02 6.24e-18 5.73e-01
4200| 4.50e-03 2.61e-03 1.55e-03 4.08e-02 3.91e-02 1.08e-18 5.84e-01
4300| 7.94e-03 4.51e-03 8.34e-04 3.66e-02 3.75e-02 3.63e-17 6.00e-01
4400| 3.09e-03 1.80e-03 2.32e-03 3.97e-02 4.22e-02 3.98e-17 6.12e-01
4500| 6.47e-04 2.21e-04 1.00e-03 3.60e-02 3.70e-02 6.48e-18 6.24e-01
4600| 1.02e-02 4.87e-03 7.15e-03 4.29e-02 3.51e-02 1.17e-17 6.35e-01
4700| 6.71e-04 3.42e-04 2.42e-04 3.68e-02 3.71e-02 1.07e-18 6.48e-01
4800| 2.58e-03 1.50e-03 2.71e-04 3.53e-02 3.56e-02 1.59e-17 6.59e-01
4900| 2.89e-01 1.71e-01 7.00e-02 -1.56e-01 -6.99e-02 1.67e-17 6.71e-01
5000| 6.05e-03 2.44e-03 7.90e-04 3.51e-02 3.59e-02 1.61e-17 6.83e-01
----------------------------------------------------------------------------
Status: Solved/Inaccurate
Hit max_iters, solution may be inaccurate, returning best found solution.
Timing: Solve time: 6.83e-01s
Lin-sys: nnz in L factor: 25455, avg solve time: 7.81e-05s
Cones: avg projection time: 3.44e-07s
Acceleration: avg step time: 3.39e-05s
----------------------------------------------------------------------------
Error metrics:
dist(s, K) = 5.0672e-20, dist(y, K*) = 0.0000e+00, s'y/|s||y| = 3.5284e-16
primal res: |Ax + s - b|_2 / (1 + |b|_2) = 1.8102e-05
dual res: |A'y + c|_2 / (1 + |c|_2) = 4.0469e-06
rel gap: |c'x + b'y| / (1 + |c'x| + |b'y|) = 9.8806e-06
----------------------------------------------------------------------------
c'x = 0.0391, -b'y = 0.0391
============================================================================
----------------------------------------------------------------------------
SCS v2.1.4 - Splitting Conic Solver
(c) Brendan O'Donoghue, Stanford University, 2012
----------------------------------------------------------------------------
Lin-sys: sparse-direct, nnz in A = 20101
eps = 1.00e-05, alpha = 1.50, max_iters = 5000, normalize = 1, scale = 1.00
acceleration_lookback = 0, rho_x = 1.00e-03
Variables n = 102, constraints m = 202
Cones: primal zero / dual free vars: 100
soc vars: 102, soc blks: 1
Setup time: 6.42e-02s
----------------------------------------------------------------------------
Iter | pri res | dua res | rel gap | pri obj | dua obj | kap/tau | time (s)
----------------------------------------------------------------------------
0| 2.91e+19 2.29e+19 1.00e+00 -6.29e+19 1.45e+18 5.11e+19 5.68e-02
100| 3.87e-05 2.47e-07 5.24e-09 5.82e-02 5.82e-02 2.21e-17 7.08e-02
140| 6.52e-06 1.66e-07 5.32e-09 5.82e-02 5.82e-02 2.23e-17 7.57e-02
----------------------------------------------------------------------------
Status: Solved
Timing: Solve time: 7.57e-02s
Lin-sys: nnz in L factor: 25455, avg solve time: 1.01e-04s
Cones: avg projection time: 7.41e-07s
Acceleration: avg step time: 8.07e-08s
----------------------------------------------------------------------------
Error metrics:
dist(s, K) = 4.4318e-19, dist(y, K*) = 1.1102e-16, s'y/|s||y| = 3.7957e-17
primal res: |Ax + s - b|_2 / (1 + |b|_2) = 6.5182e-06
dual res: |A'y + c|_2 / (1 + |c|_2) = 1.6561e-07
rel gap: |c'x + b'y| / (1 + |c'x| + |b'y|) = 5.3244e-09
----------------------------------------------------------------------------
c'x = 0.0582, -b'y = 0.0582
============================================================================
Error:
/usr/local/lib/python3.7/dist-packages/jax/_src/api.py in jacfun(*args, **kwargs)
970 tree_map(partial(_check_input_dtype_jacfwd, holomorphic), dyn_args)
971 pushfwd = partial(_jvp, f_partial, dyn_args)
--> 972 y, jac = vmap(pushfwd, out_axes=(None, -1))(_std_basis(dyn_args))
973 tree_map(partial(_check_output_dtype_jacfwd, holomorphic), y)
974 example_args = dyn_args[0] if isinstance(argnums, int) else dyn_args
<ipython-input-60-2f66503cf065> in <lambda>(*args, **kwargs)
16 return lambda *args, **kwargs: CvxpyLayer(prob,
17 parameters=[_F, _dx_F, _dt_F, _ground_truth],
---> 18 variables=[_lam])(*args, **kwargs)[0]
/usr/local/lib/python3.7/dist-packages/cvxpylayers/jax/cvxpylayer.py in CvxpyLayer(problem, parameters, variables, gp)
74 )
75 else:
---> 76 data, _, _ = problem.get_problem_data(solver=cp.SCS)
77 compiler = data[cp.settings.PARAM_PROB]
78 param_ids = [p.id for p in param_order]
/usr/local/lib/python3.7/dist-packages/cvxpy/problems/problem.py in get_problem_data(self, solver, gp, enforce_dpp, verbose)
599
600 data, solver_inverse_data = solving_chain.solver.apply(
--> 601 self._cache.param_prog)
602 inverse_data = self._cache.inverse_data + [solver_inverse_data]
603 self._compilation_time = time.time() - start
/usr/local/lib/python3.7/dist-packages/cvxpy/reductions/solvers/conic_solvers/scs_conif.py in apply(self, problem)
219 # Apply parameter values.
220 # Obtain A, b such that Ax + s = b, s \in cones.
--> 221 c, d, A, b = problem.apply_parameters()
222 data[s.C] = c
223 inv_data[s.OFFSET] = d
/usr/local/lib/python3.7/dist-packages/cvxpy/reductions/dcp2cone/cone_matrix_stuffing.py in apply_parameters(self, id_to_param_value, zero_offset, keep_zeros)
203 self.reduced_A, param_vec, self.x.size,
204 nonzero_rows=self._A_mapping_nonzero, with_offset=True,
--> 205 problem_data_index=self.problem_data_index)
206 return c, d, A, np.atleast_1d(b)
207
/usr/local/lib/python3.7/dist-packages/cvxpy/cvxcore/python/canonInterface.py in get_matrix_from_tensor(problem_data_tensor, param_vec, var_length, nonzero_rows, with_offset, problem_data_index)
236 if nonzero_rows is not None and nonzero_rows.size > 0:
237 A_nrows, _ = A.shape
--> 238 A_rows, A_cols = nonzero_csc_matrix(A)
239 A_vals = np.append(A.data, np.zeros(nonzero_rows.size))
240 A_rows = np.append(A_rows, nonzero_rows % A_nrows)
/usr/local/lib/python3.7/dist-packages/cvxpy/cvxcore/python/canonInterface.py in nonzero_csc_matrix(A)
144 # this function returns (rows, cols) corresponding to nonzero entries in
145 # A; an entry that is explicitly set to zero is treated as nonzero
--> 146 assert not np.isnan(A.data).any()
147
148 # scipy drops rows, cols with explicit zeros; use nan as a sentinel
AssertionError:
