change definition of weights

src/loaders/prec_inn.py

@@ -139,7 +139,7 @@ def compute_preprocessing(
         input_obs.append(mass)
 
     if append_delta_r:
-        drinv = lambda x: x #np.minimum(1/(x+1e-7), 20)
+        drinv = lambda x: np.minimum(1/(x+1e-7), 20)
         if mult > 3:
             input_obs.append(drinv(dr(obs.phi[:,2], obs.phi[:,3], obs.eta[:,2], obs.eta[:,3])))
         if mult > 4:

src/plots.py

@@ -137,7 +137,7 @@ def plot_roc(self, file: str):
         Args:
             file: Output file name
         """
-        scores = -np.concatenate((1/self.weights_true, self.weights_fake), axis=0)
+        scores = -np.concatenate((self.weights_true, self.weights_fake), axis=0)
         labels = np.concatenate((
             np.ones_like(self.weights_true),
             np.zeros_like(self.weights_fake)
@@ -231,20 +231,17 @@ def plot_single_weight_hist(
         if self.bayesian:
             true_hists = np.stack([
                 np.histogram(
-                    self.weights_true[:,i] / np.mean(self.weights_true[:,i]),
-                    bins=bins
+                    self.weights_true[:,i], bins=bins
                 )[0] for i in range(self.weights_true.shape[1])
             ], axis=1)
             fake_hists = np.stack([
                 np.histogram(
-                    self.weights_fake[:,i] / np.mean(self.weights_fake[:,i]),
-                    bins=bins
+                    self.weights_fake[:,i], bins=bins
                 )[0] for i in range(self.weights_fake.shape[1])
             ], axis=1)
             combined_hists = np.stack([
                 np.histogram(
-                    weights_combined[:,i] / np.mean(weights_combined[:,i]),
-                    bins=bins
+                    weights_combined[:,i], bins=bins
                 )[0] for i in range(weights_combined.shape[1])
             ], axis=1)
 
@@ -263,14 +260,6 @@ def plot_single_weight_hist(
                 np.quantile(combined_hists, 0.159, axis=1),
                 np.quantile(combined_hists, 0.841, axis=1)
             ), axis=0)
-
-            #y_true = np.mean(true_hists, axis=1)
-            #y_true_err = np.std(true_hists, axis=1)
-            #y_fake = np.mean(fake_hists, axis=1)
-            #y_fake_err = np.std(fake_hists, axis=1)
-            #y_combined = np.mean(combined_hists, axis=1)
-            #y_combined_err = np.std(combined_hists, axis=1)
-
         else:
             y_true = np.histogram(self.weights_true, bins=bins)[0]
             y_true_err = None

src/train.py

@@ -286,7 +286,7 @@ def predict_single(self):
             ])
             y_true_sig = y_true.sigmoid().flatten()
             y_fake_sig = y_fake.sigmoid().flatten()
-            w_true = (1 - y_true_sig) / y_true_sig
+            w_true = y_true_sig / (1 - y_true_sig)
             w_fake = y_fake_sig / (1 - y_fake_sig)
             min_size = min(len(y_true), len(y_fake))
             clf_score = self.loss(