feat: RBR Implementation

experiments/experimental_setup.yaml

@@ -146,3 +146,7 @@ recourse_methods:
     hyperparams:
   roar:
     hyperparams:
+  rbr:
+    hyperparams:
+      train_data: None
+      device: "cpu"

experiments/run_experiment.py

@@ -170,6 +170,11 @@ def initialize_recourse_method(
         return Probe(mlmodel, hyperparams)
     elif method == "roar":
         return Roar(mlmodel, hyperparams)
+    elif method == "rbr":
+        hyperparams["train_data"] = data.df_train.drop(columns=["y"], axis=1)
+        dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        hyperparams["device"] = dev
+        return RBR(mlmodel, hyperparams)
     else:
         raise ValueError("Recourse method not known")
 
@@ -199,7 +204,7 @@ def create_parser():
     -r, --recourse_method: Specifies recourse methods for the experiment.
         Default: ["dice", "cchvae", "cem", "cem_vae", "clue", "cruds", "face_knn", "face_epsilon", "gs", "mace", "revise", "wachter"].
         Choices: ["dice", "ar", "causal_recourse", "cchvae", "cem", "cem_vae", "claproar", "clue", "cruds", "face_knn", "face_epsilon", "feature_tweak",
-            "focus", "gravitational", "greedy", "gs", "mace", "revise", "wachter", "cfvae", "roar", "probe"].
+            "focus", "gravitational", "greedy", "gs", "mace", "revise", "wachter", "cfvae", "roar", "probe", "rbr"].
     -n, --number_of_samples: Specifies the number of instances per dataset.
         Default: 20.
     -s, --train_split: Specifies the split of the available data used for training.
@@ -292,6 +297,7 @@ def create_parser():
             "cfvae",
             "probe",
             "roar",
+            "rbr",
         ],
         help="Recourse methods for experiment",
     )
@@ -374,6 +380,7 @@ def create_parser():
         "cfvae",
         "probe",
         "roar",
+        "rbr",
     ]
     sklearn_methods = ["feature_tweak", "focus", "mace"]
 

methods/__init__.py

@@ -8,6 +8,7 @@
     CRUD,
     FOCUS,
     MACE,
+    RBR,
     ActionableRecourse,
     CausalRecourse,
     ClaPROAR,

methods/catalog/__init__.py

@@ -16,6 +16,7 @@
 from .growing_spheres import GrowingSpheres
 from .mace import MACE
 from .probe import Probe
+from .rbr import RBR
 from .revise import Revise
 from .roar import Roar
 from .wachter import Wachter

methods/catalog/probe/library/probe.py

@@ -71,14 +71,15 @@ def perturb_sample(x, n_samples, sigma2):
     return X + eps
 
 
-def reparametrization_trick(mu, sigma2, n_samples):
+def reparametrization_trick(mu, sigma2, device, n_samples):
     # var = torch.eye(mu.shape[1]) * sigma2
-    std = torch.sqrt(sigma2)
+    std = torch.sqrt(sigma2).to(device)
     epsilon = MultivariateNormal(
         loc=torch.zeros(mu.shape[1]), covariance_matrix=torch.eye(mu.shape[1])
     )
     epsilon = epsilon.sample((n_samples,))  # standard Gaussian random noise
-    ones = torch.ones_like(epsilon)
+    epsilon = epsilon.to(device)
+    ones = torch.ones_like(epsilon).to(device)
     random_samples = mu.reshape(-1) * ones + std * epsilon
 
     return random_samples
@@ -176,7 +177,9 @@ def probe_recourse(
     costs = []
     ces = []
 
-    random_samples = reparametrization_trick(x_new, noise_variance, n_samples=1000)
+    random_samples = reparametrization_trick(
+        x_new, noise_variance, device, n_samples=1000
+    )
     invalidation_rate = compute_invalidation_rate(torch_model, random_samples)
 
     while (f_x_new <= DECISION_THRESHOLD) or (
@@ -226,7 +229,7 @@ def probe_recourse(
             optimizer.step()
 
             random_samples = reparametrization_trick(
-                x_new, noise_variance, n_samples=10000
+                x_new, noise_variance, device, n_samples=10000
             )
             invalidation_rate = compute_invalidation_rate(torch_model, random_samples)
 

methods/catalog/rbr/__init__.py

@@ -0,0 +1,3 @@
+# flake8: noqa
+
+from .model import RBR

methods/catalog/rbr/library/__init__.py

@@ -0,0 +1,3 @@
+# flake8: noqa
+
+from .rbr_loss import robust_bayesian_recourse