FIX: drop dead func call; move compute to where needed

Author: sappelhoff

.gitignore

@@ -25,4 +25,6 @@ picard.egg-info/
 .tags
 tags
 .coverage
-python_picard.egg-info/
+python_picard.egg-info/
+
+junit-results.xml

picard/_core_picard.py

@@ -89,8 +89,6 @@ def core_picard(X, density=Tanh(), ortho=False, extended=False, m=7,
         # Compute the relative gradient and the Hessian off-diagonal
         G = np.inner(psiY, Y) / T
         del psiY
-        # Compute the squared signals
-        Y_square = Y ** 2
         # Compute the kurtosis and update the gradient accordingly
         if extended:
             K = np.mean(psidY, axis=1) * np.diag(C)
@@ -115,10 +113,13 @@ def core_picard(X, density=Tanh(), ortho=False, extended=False, m=7,
             diag = psidY_mean[:, None] * np.ones(N)[None, :]
             h = 0.5 * (diag + diag.T - h_off[:, None] - h_off[None, :])
             h[h < lambda_min] = lambda_min
+            del psidY
         else:
+            # Compute the squared signals (only needed for this Hessian)
+            Y_square = Y ** 2
             h = np.inner(psidY, Y_square) / T
             h = _regularize_hessian(h, h_off, lambda_min)
-        del psidY, Y_square
+            del psidY, Y_square
         # Project the gradient if ortho
         if ortho:
             G = (G - G.T) / 2

picard/densities.py

@@ -44,7 +44,6 @@ def log_lik(self, Y):
         alpha = self.alpha  # noqa
         if ne is None:
             absY = np.abs(Y)
-            np.exp(-2. * alpha * absY)
             return absY + np.log1p(np.exp(-2. * alpha * absY)) / alpha
         return ne.evaluate('abs(Y) + log1p(exp(-2. * alpha * abs(Y))) / alpha')