parallelize the cvxpy code

autokoopman/estimator/koopman.py

@@ -74,9 +74,9 @@ def swdmdc(X, Xp, U, r, Js, W):
     K = cp.Variable((n_obs, n_obs))
 
     # SW-eDMD objective
-    objective = cp.Minimize(sum([ 
-        cp.sum_squares(sf * cp.multiply((np.abs(J) @ w)[:n_obs], (xpi[:n_obs] - K @ xi[:n_obs]))) for J, w, xpi, xi in zip(Js, W, Xp, X)
-    ]))
+    weights_obj = np.vstack([(np.abs(J) @ w)[:n_obs] for J, w in zip(Js, W)]).T 
+    P = sf * cp.multiply(weights_obj, Xp[:, :n_obs].T - K @ X[:, :n_obs].T)
+    objective = cp.Minimize(cp.sum_squares(P))
 
     # unconstrained problem
     constraints = None

notebooks/weighted-cost-func.ipynb

@@ -124,7 +124,7 @@
     "    learning_weights=weights, # weight the eDMD algorithm objectives\n",
     "    scoring_weights=weights, # pass weights as required for cost_func=\"weighted\"\n",
     "    opt=\"grid\",             # grid search to find best hyperparameters\n",
-    "    n_obs=30,              # maximum number of observables to try\n",
+    "    n_obs=10,              # maximum number of observables to try\n",
     "    max_opt_iter=200,       # maximum number of optimization iterations\n",
     "    grid_param_slices=5,   # for grid search, number of slices for each parameter\n",
     "    n_splits=5,             # k-folds validation for tuning, helps stabilize the scoring\n",
@@ -156,7 +156,7 @@
     "    learning_weights=None,  # don't use eDMD weighting\n",
     "    scoring_weights=weights, # pass weights as required for cost_func=\"weighted\"\n",
     "    opt=\"grid\",             # grid search to find best hyperparameters\n",
-    "    n_obs=30,              # maximum number of observables to try\n",
+    "    n_obs=10,              # maximum number of observables to try\n",
     "    max_opt_iter=200,       # maximum number of optimization iterations\n",
     "    grid_param_slices=5,   # for grid search, number of slices for each parameter\n",
     "    n_splits=5,             # k-folds validation for tuning, helps stabilize the scoring\n",