Add tests using the weighted manhattan distance as a regularizer

tests/sklearn/test_sklearn_softmaxregression.py

@@ -10,6 +10,7 @@
 from sklearn.linear_model import LogisticRegression
 
 from ceml.sklearn import generate_counterfactual
+from ceml.backend.jax.costfunctions import LMadCost 
 
 
 def test_softmaxregression():
@@ -26,13 +27,22 @@ def test_softmaxregression():
     x_orig = X_test[1:4][0,:]
     assert model.predict([x_orig]) == 2
 
+    # Create weighted manhattan distance cost function
+    md = np.median(X_train, axis=0)
+    mad = np.median(np.abs(X_train - md), axis=0)
+    regularization_mad = LMadCost(x_orig, mad)
+
     # Compute counterfactual
     features_whitelist = None
 
     x_cf, y_cf, delta = generate_counterfactual(model, x_orig, 0, features_whitelist=features_whitelist, regularization="l1", C=1.0, optimizer="bfgs", return_as_dict=False)
     assert y_cf == 0
     assert model.predict(np.array([x_cf])) == 0
 
+    x_cf, y_cf, delta = generate_counterfactual(model, x_orig, 0, features_whitelist=features_whitelist, regularization=regularization_mad, C=1.0, optimizer="bfgs", return_as_dict=False)
+    assert y_cf == 0
+    assert model.predict(np.array([x_cf])) == 0
+
     x_cf, y_cf, delta = generate_counterfactual(model, x_orig, 0, features_whitelist=features_whitelist, regularization="l1", C=1.0, optimizer="cg", return_as_dict=False)
     assert y_cf == 0
     assert model.predict(np.array([x_cf])) == 0

tests/tfkeras/test_tfkeras_softmaxregression.py

@@ -13,7 +13,7 @@
 from sklearn.metrics import accuracy_score
 
 from ceml.tfkeras import generate_counterfactual
-from ceml.backend.tensorflow.costfunctions import NegLogLikelihoodCost
+from ceml.backend.tensorflow.costfunctions import NegLogLikelihoodCost, LMadCost
 from ceml.model import ModelWithLoss
 
 
@@ -61,6 +61,11 @@ def get_loss(self, y_target, pred=None):
     x_orig = X_test[1,:]
     assert model.predict(np.array([x_orig])) == 1
 
+    # Create weighted manhattan distance cost function
+    md = np.median(X_train, axis=0)
+    mad = np.median(np.abs(X_train - md), axis=0)
+    regularization_mad = LMadCost(x_orig, mad)
+
     # Compute counterfactual
     features_whitelist = None
 
@@ -70,6 +75,11 @@ def get_loss(self, y_target, pred=None):
     assert y_cf == 0
     assert model.predict(np.array([x_cf])) == 0
 
+    optimizer = "bfgs"
+    x_cf, y_cf, delta = generate_counterfactual(model, x_orig, y_target=0, features_whitelist=features_whitelist, regularization=regularization_mad, C=0.01, optimizer=optimizer, optimizer_args=optimizer_args, return_as_dict=False)
+    assert y_cf == 0
+    assert model.predict(np.array([x_cf])) == 0
+
     optimizer = "nelder-mead"
     x_cf, y_cf, delta = generate_counterfactual(model, x_orig, y_target=0, features_whitelist=features_whitelist, regularization="l1", C=0.01, optimizer=optimizer, optimizer_args=optimizer_args, return_as_dict=False)
     assert y_cf == 0

tests/torch/test_torch_softmaxregression.py

@@ -11,7 +11,7 @@
 from sklearn.metrics import accuracy_score
 
 from ceml.torch import generate_counterfactual
-from ceml.backend.torch.costfunctions import NegLogLikelihoodCost
+from ceml.backend.torch.costfunctions import NegLogLikelihoodCost, LMadCost
 from ceml.model import ModelWithLoss
 
 
@@ -74,6 +74,11 @@ def get_loss(self, y_target, pred=None):
     x_orig = X_test[1,:]
     assert model.predict(torch.from_numpy(np.array([x_orig]))).numpy() == 1
 
+    # Create weighted manhattan distance cost function
+    md = np.median(X_train, axis=0)
+    mad = np.median(np.abs(X_train - md), axis=0)
+    regularization_mad = LMadCost(torch.from_numpy(x_orig), torch.from_numpy(mad))
+
     # Compute counterfactual
     features_whitelist = None
 
@@ -83,6 +88,12 @@ def get_loss(self, y_target, pred=None):
     assert y_cf == 0
     assert model.predict(torch.from_numpy(np.array([x_cf], dtype=np.float32))).numpy() == 0
 
+    optimizer = "bfgs"
+    optimizer_args = {"max_iter": 1000, "args": {"lr": 0.9, "momentum": 0.9}}
+    x_cf, y_cf, delta = generate_counterfactual(model, x_orig, y_target=0, features_whitelist=features_whitelist, regularization=regularization_mad, C=0.001, optimizer=optimizer, optimizer_args=optimizer_args, return_as_dict=False)
+    assert y_cf == 0
+    assert model.predict(torch.from_numpy(np.array([x_cf], dtype=np.float32))).numpy() == 0
+
     optimizer = "nelder-mead"
     x_cf, y_cf, delta = generate_counterfactual(model, x_orig, y_target=0, features_whitelist=features_whitelist, regularization="l1", C=0.001, optimizer=optimizer, optimizer_args=optimizer_args, return_as_dict=False)
     assert y_cf == 0