Fix coral, kliep, utils

Author: antoinedemathelin

adapt/feature_based/_coral.py

@@ -196,8 +196,17 @@ def fit_embeddings(self, Xs, Xt):
 
         self.Cs_ = self.lambda_ * cov_Xs + np.eye(Xs.shape[1])
         self.Ct_ = self.lambda_ * cov_Xt + np.eye(Xt.shape[1])
-        Xs_emb = np.matmul(Xs, linalg.inv(linalg.sqrtm(self.Cs_)))
-        Xs_emb = np.matmul(Xs_emb, linalg.sqrtm(self.Ct_))
+        
+        Cs_sqrt_inv = linalg.inv(linalg.sqrtm(self.Cs_))
+        Ct_sqrt = linalg.sqrtm(self.Ct_)
+        
+        if np.iscomplexobj(Cs_sqrt_inv):
+            Cs_sqrt_inv = Cs_sqrt_inv.real
+        if np.iscomplexobj(Ct_sqrt):
+            Ct_sqrt = Ct_sqrt.real
+        
+        Xs_emb = np.matmul(Xs, Cs_sqrt_inv)
+        Xs_emb = np.matmul(Xs_emb, Ct_sqrt)
 
         if self.verbose:
             new_cov_Xs = np.cov(Xs_emb, rowvar=False)
@@ -280,8 +289,16 @@ def predict_features(self, X, domain="tgt"):
         if domain in ["tgt", "target"]:
             X_emb = X
         elif domain in ["src", "source"]:
-            X_emb = np.matmul(X, linalg.inv(linalg.sqrtm(self.Cs_)))
-            X_emb = np.matmul(X_emb, linalg.sqrtm(self.Ct_))
+            Cs_sqrt_inv = linalg.inv(linalg.sqrtm(self.Cs_))
+            Ct_sqrt = linalg.sqrtm(self.Ct_)
+
+            if np.iscomplexobj(Cs_sqrt_inv):
+                Cs_sqrt_inv = Cs_sqrt_inv.real
+            if np.iscomplexobj(Ct_sqrt):
+                Ct_sqrt = Ct_sqrt.real
+            
+            X_emb = np.matmul(X, Cs_sqrt_inv)
+            X_emb = np.matmul(X_emb, Ct_sqrt)
         else:
             raise ValueError("`domain `argument "
                              "should be `tgt` or `src`, "

adapt/instance_based/_kliep.py

@@ -288,6 +288,8 @@ def fit_estimator(self, X, y,
         else:
             if sample_weight is not None:
                 sample_weight /= (sample_weight.sum() + EPS)
+            if sample_weight.sum() <= 0:
+                sample_weight = np.ones(len(sample_weight))
             bootstrap_index = np.random.choice(
             len(X), size=len(X), replace=True,
             p=sample_weight)
@@ -359,7 +361,7 @@ def _fit(self, Xs, Xt, sigma):
             alpha += b * ((((1-np.dot(np.transpose(b), alpha)) /
                             (np.dot(np.transpose(b), b) + EPS))))
             alpha = np.maximum(0, alpha)
-            alpha /= np.dot(np.transpose(b), alpha)
+            alpha /= (np.dot(np.transpose(b), alpha) + EPS)
             objective = np.mean(np.log(np.dot(A, alpha) + EPS))
             k += 1
 
@@ -389,7 +391,7 @@ def _cross_val_jscore(self, Xs, Xt, sigma, cv):
                                         centers,
                                         sigma),
                     alphas
-                )
+                ) + EPS
             ))
             cv_scores.append(j_score)
         return cv_scores

adapt/utils.py

@@ -364,6 +364,7 @@ class GradientHandler(Layer):
     """
     def __init__(self, lambda_init=1., name="g_handler"):
         super().__init__(name=name)
+        self.lambda_init=1.
         self.lambda_ = tf.Variable(lambda_init,
                                    trainable=False,
                                    dtype="float32")
@@ -384,6 +385,21 @@ def call(self, x):
         return _grad_handler(x, self.lambda_)
 
 
+    def get_config(self):
+        """
+        Return config dictionnary.
+        
+        Returns
+        -------
+        dict
+        """
+        config = super().get_config().copy()
+        config.update({
+            'lambda_init': self.lambda_init
+        })
+        return config
+
+
 def make_classification_da(n_samples=100, 
                            n_features=2,
                            random_state=2):