Altered Quadratic class

Author: hoodaty

skglm/datafits/single_task.py

@@ -27,7 +27,10 @@ class Quadratic(BaseDatafit):
     """
 
     def __init__(self):
-        pass
+        self.sample_weights = None
+
+    def set_sample_weights(self, sample_weights):
+        self.sample_weights = sample_weights
 
     def get_spec(self):
         spec = (
@@ -40,24 +43,34 @@ def params_to_dict(self):
 
     def get_lipschitz(self, X, y):
         n_features = X.shape[1]
-
-        lipschitz = np.zeros(n_features, dtype=X.dtype)
-        for j in range(n_features):
-            lipschitz[j] = (X[:, j] ** 2).sum() / len(y)
-
+        if self.sample_weights is None:
+            lipschitz = np.zeros(n_features, dtype=X.dtype)
+            for j in range(n_features):
+                lipschitz[j] = (X[:, j] ** 2).sum() / len(y)
+        else:
+            lipschitz = np.zeros(n_features, dtype=X.dtype)
+            for j in range(n_features):
+                lipschitz[j] = np.sum(self.sample_weights * (X[:, j] ** 2))/len(y)
         return lipschitz
 
     def get_lipschitz_sparse(self, X_data, X_indptr, X_indices, y):
         n_features = len(X_indptr) - 1
-        lipschitz = np.zeros(n_features, dtype=X_data.dtype)
-
-        for j in range(n_features):
-            nrm2 = 0.
-            for idx in range(X_indptr[j], X_indptr[j + 1]):
-                nrm2 += X_data[idx] ** 2
-
-            lipschitz[j] = nrm2 / len(y)
-
+        if self.sample_weights is None:
+            lipschitz = np.zeros(n_features, dtype=X_data.dtype)
+            for j in range(n_features):
+                nrm2 = 0.
+                for idx in range(X_indptr[j], X_indptr[j + 1]):
+                    nrm2 += X_data[idx] ** 2
+
+                lipschitz[j] = nrm2 / len(y)
+        else:
+           lipschitz = np.zeros(n_features, dtype=X_data.dtype)
+           for j in range(n_features):
+                nrm2 = 0.
+                for idx in range(X_indptr[j], X_indptr[j + 1]):
+                    nrm2 += self.sample_weights[idx]*X_data[idx] ** 2
+
+                lipschitz[j] = nrm2 / len(y)
         return lipschitz
 
     def initialize(self, X, y):
@@ -75,33 +88,57 @@ def initialize_sparse(self, X_data, X_indptr, X_indices, y):
             self.Xty[j] = xty
 
     def get_global_lipschitz(self, X, y):
-        return norm(X, ord=2) ** 2 / len(y)
+        if self.sample_weights is None:
+           return norm(X, ord=2) ** 2 / len(y)
+        else:
+           return (norm(X@self.sample_weights, axis=1, ord=2) ** 2)/len(y)
 
     def get_global_lipschitz_sparse(self, X_data, X_indptr, X_indices, y):
-        return spectral_norm(X_data, X_indptr, X_indices, len(y)) ** 2 / len(y)
+        if self.sample_weights is None:
+            return spectral_norm(X_data, X_indptr, X_indices, len(y)) ** 2 / len(y)
+        else:
+            return spectral_norm(X_data@self.sample_weights, X_indptr, X_indices, len(y)) ** 2 /len(y)
 
     def value(self, y, w, Xw):
-        return np.sum((y - Xw) ** 2) / (2 * len(Xw))
+        if self.sample_weights is None:
+            return np.sum((y - Xw) ** 2) / (2 * len(Xw))
+        else:
+            return np.sum(self.sample_weights * (y - Xw) ** 2) / (2 * len(Xw))
 
     def gradient_scalar(self, X, y, w, Xw, j):
-        return (X[:, j] @ Xw - self.Xty[j]) / len(Xw)
+        if self.sample_weights is None:
+            return (X[:, j] @ Xw - self.Xty[j]) / len(Xw)
+        else:
+            return self.sample_weights * (X[:, j] @ Xw - self.Xty[j])
 
     def gradient_scalar_sparse(self, X_data, X_indptr, X_indices, y, Xw, j):
         XjTXw = 0.
-        for i in range(X_indptr[j], X_indptr[j+1]):
-            XjTXw += X_data[i] * Xw[X_indices[i]]
-        return (XjTXw - self.Xty[j]) / len(Xw)
+        if self.sample_weights is None:
+            for i in range(X_indptr[j], X_indptr[j+1]):
+                XjTXw += X_data[i] * Xw[X_indices[i]]
+            return (XjTXw - self.Xty[j]) / len(Xw)
+        else:
+            XjTXw_weighted = 0.
+            for i in range(X_indptr[j], X_indptr[j+1]):
+                XjTXw_weighted += X_data[i] * Xw[X_indices[i]] * self.sample_weights[X_indices[i]]
+            return (XjTXw_weighted - self.Xty[j]) / len(Xw)
 
-    def full_grad_sparse(
-            self, X_data, X_indptr, X_indices, y, Xw):
+    def full_grad_sparse(self, X_data, X_indptr, X_indices, y, Xw):
         n_features = X_indptr.shape[0] - 1
         n_samples = y.shape[0]
         grad = np.zeros(n_features, dtype=Xw.dtype)
-        for j in range(n_features):
-            XjTXw = 0.
-            for i in range(X_indptr[j], X_indptr[j + 1]):
-                XjTXw += X_data[i] * Xw[X_indices[i]]
-            grad[j] = (XjTXw - self.Xty[j]) / n_samples
+        if self.sample_weights is None:
+            for j in range(n_features):
+               XjTXw = 0.
+               for i in range(X_indptr[j], X_indptr[j + 1]):
+                    XjTXw += X_data[i] * Xw[X_indices[i]]
+               grad[j] = (XjTXw - self.Xty[j]) / n_samples
+        else:
+            for j in range(n_features):
+                XjTXw_weighted = 0.
+                for i in range(X_indptr[j], X_indptr[j + 1]):
+                    XjTXw_weighted += X_data[i] * Xw[X_indices[i]] * self.sample_weights[X_indices[i]]
+                grad[j] = (XjTXw_weighted - self.Xty[j]) / n_samples
         return grad
 
     def intercept_update_step(self, y, Xw):