Squash me

pysindy/optimizers/constrained_sr3.py

@@ -274,7 +274,7 @@ def _create_var_and_part_cost(
             cost = cost + cp.norm2(np.ravel(self.thresholds) @ xi) ** 2
         return xi, cost
 
-    def _update_coef_cvxpy(self, xi, cost, var_len, x, y, coef_sparse):
+    def _update_coef_cvxpy(self, xi, cost, var_len, coef_sparse, tol):
         if self.use_constraints:
             if self.inequality_constraints and self.equality_constraints:
                 # Process inequality constraints then equality constraints
@@ -304,22 +304,22 @@ def _update_coef_cvxpy(self, xi, cost, var_len, x, y, coef_sparse):
         try:
             prob.solve(
                 max_iter=self.max_iter,
-                eps_abs=self.tol,
-                eps_rel=self.tol,
+                eps_abs=tol,
+                eps_rel=tol,
                 verbose=self.verbose_cvxpy,
             )
         # Annoying error coming from L2 norm switching to use the ECOS
         # solver, which uses "max_iters" instead of "max_iter", and
         # similar semantic changes for the other variables.
         except TypeError:
             try:
-                prob.solve(abstol=self.tol, reltol=self.tol, verbose=self.verbose_cvxpy)
+                prob.solve(abstol=tol, reltol=tol, verbose=self.verbose_cvxpy)
             except cp.error.SolverError:
                 print("Solver failed, setting coefs to zeros")
-                xi.value = np.zeros(coef_sparse.shape[0] * coef_sparse.shape[1])
+                xi.value = np.zeros(var_len)
         except cp.error.SolverError:
             print("Solver failed, setting coefs to zeros")
-            xi.value = np.zeros(coef_sparse.shape[0] * coef_sparse.shape[1])
+            xi.value = np.zeros(var_len)
 
         if xi.value is None:
             warnings.warn(
@@ -438,7 +438,7 @@ def _reduce(self, x, y):
             var_len = coef_sparse.shape[0] * coef_sparse.shape[1]
             xi, cost = self._create_var_and_part_cost(var_len, x_expanded, y)
             coef_sparse = self._update_coef_cvxpy(
-                xi, cost, var_len, x_expanded, y, coef_sparse
+                xi, cost, var_len, coef_sparse, self.tol
             )
             objective_history.append(self._objective(x, y, 0, coef_full, coef_sparse))
         else:

pysindy/optimizers/trapping_sr3.py

@@ -391,7 +391,7 @@ def _objective(self, x, y, coef_sparse, A, PW, q):
             )
         return 0.5 * np.sum(R2) + 0.5 * np.sum(A2) / self.eta + L1
 
-    def _solve_sparse_relax_and_split(self, xi, cost, var_len, coef_prev):
+    def _solve_sparse_relax_and_split(self, xi, cost, var_len, coef_prev, tol):
         """Solve coefficient update with CVXPY if threshold != 0"""
         if self.use_constraints:
             if self.inequality_constraints:
@@ -411,20 +411,16 @@ def _solve_sparse_relax_and_split(self, xi, cost, var_len, coef_prev):
         try:
             prob.solve(
                 max_iter=self.max_iter,
-                eps_abs=self.eps_solver,
-                eps_rel=self.eps_solver,
+                eps_abs=tol,
+                eps_rel=tol,
                 verbose=self.verbose_cvxpy,
             )
         # Annoying error coming from L2 norm switching to use the ECOS
         # solver, which uses "max_iters" instead of "max_iter", and
         # similar semantic changes for the other variables.
         except TypeError:
             try:
-                prob.solve(
-                    abstol=self.eps_solver,
-                    reltol=self.eps_solver,
-                    verbose=self.verbose_cvxpy,
-                )
+                prob.solve(abstol=tol, reltol=tol, verbose=self.verbose_cvxpy)
             except cp.error.SolverError:
                 print("Solver failed, setting coefs to zeros")
                 xi.value = np.zeros(var_len)
@@ -641,7 +637,7 @@ def _reduce(self, x, y):
                         )
                         # sparse relax_and_split
                         coef_sparse = self._solve_sparse_relax_and_split(
-                            xi, cost, r * n_features, coef_prev
+                            xi, cost, r * n_features, coef_prev, self.eps_solver
                         )
                     else:
                         pTp = np.dot(Pmatrix.T, Pmatrix)