Merge branch 'develop' into integration

os-climate · Oct 2, 2024 · eba1372 · eba1372
2 parents 3b19a2a + 2d8b13a
commit eba1372
Showing 1 changed file with 40 additions and 37 deletions.
diff --git a/sostrades_optimization_plugins/tools/cst_manager/func_manager_common.py b/sostrades_optimization_plugins/tools/cst_manager/func_manager_common.py
@@ -114,43 +114,46 @@ def get_dsmooth_dvariable(cst, alpha=3):
 
 
 def get_dsmooth_dvariable_vect(cst, alpha=3):
-    cst_array = np.array(cst)
-    max_exp = 650.0  # max value for exponent input, higher value gives infinity
-    min_exp = -300
-    alphaxcst = alpha * cst_array
-
-    max_alphax = np.amax(alphaxcst, axis=1)
-
-    k = max_alphax - max_exp
-    exp_func = np.maximum(min_exp, alpha * cst_array -
-                          np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))
-    den = np.sum(np.exp(exp_func), axis=1)
-    num = np.sum(cst_array * np.exp(exp_func), axis=1)
-
-    # Vectorized calculation
-    exp_func = np.maximum(min_exp, alpha * cst_array -
-                          np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))
-    dden = alpha * np.exp(exp_func)
-    dnum = cst_array * (alpha * np.exp(exp_func)
-                        ) + np.exp(exp_func)
-    grad_value = dnum / np.repeat(den, cst_array.shape[1]).reshape(cst_array.shape) - (np.repeat(num, cst_array.shape[1]).reshape(
-        cst_array.shape) / np.repeat(den, cst_array.shape[1]).reshape(cst_array.shape)) * (dden / np.repeat(den, cst_array.shape[1]).reshape(cst_array.shape))
-
-    # Special case for max element
-    max_elem = np.amax(cst_array * np.sign(alpha), axis=1) * np.sign(alpha)
-    non_max_idx = np.array([cst_array[i] != max_elem[i]
-                            for i in np.arange(cst_array.shape[0])]).reshape(cst_array.shape[0], cst_array.shape[1])
-    dden_max = np.sum(-alpha * non_max_idx *
-                      np.exp(np.maximum(min_exp, alpha * cst_array - np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))), axis=1)
-    dnum_max = np.sum(-alpha * cst_array * non_max_idx *
-                      np.exp(np.maximum(min_exp, alpha * cst_array - np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))), axis=1)
-    # derivative of den wto cstmax is 0
-    dden_max = dden_max + 0.0
-    dnum_max = dnum_max + 1.0 * np.exp(alpha * max_elem - k)
-    grad_val_max = dnum_max / den - (num / den) * (dden_max / den)
-
-    for i in np.arange(cst_array.shape[0]):
-        grad_value[i][np.logical_not(non_max_idx)[i]] = grad_val_max[i]
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+
+        cst_array = np.array(cst)
+        max_exp = 650.0  # max value for exponent input, higher value gives infinity
+        min_exp = -300
+        alphaxcst = alpha * cst_array
+
+        max_alphax = np.amax(alphaxcst, axis=1)
+
+        k = max_alphax - max_exp
+        exp_func = np.maximum(min_exp, alpha * cst_array -
+                              np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))
+        den = np.sum(np.exp(exp_func), axis=1)
+        num = np.sum(cst_array * np.exp(exp_func), axis=1)
+
+        # Vectorized calculation
+        exp_func = np.maximum(min_exp, alpha * cst_array -
+                              np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))
+        dden = alpha * np.exp(exp_func)
+        dnum = cst_array * (alpha * np.exp(exp_func)
+                            ) + np.exp(exp_func)
+        grad_value = dnum / np.repeat(den, cst_array.shape[1]).reshape(cst_array.shape) - (np.repeat(num, cst_array.shape[1]).reshape(
+            cst_array.shape) / np.repeat(den, cst_array.shape[1]).reshape(cst_array.shape)) * (dden / np.repeat(den, cst_array.shape[1]).reshape(cst_array.shape))
+
+        # Special case for max element
+        max_elem = np.amax(cst_array * np.sign(alpha), axis=1) * np.sign(alpha)
+        non_max_idx = np.array([cst_array[i] != max_elem[i]
+                                for i in np.arange(cst_array.shape[0])]).reshape(cst_array.shape[0], cst_array.shape[1])
+        dden_max = np.sum(-alpha * non_max_idx *
+                          np.exp(np.maximum(min_exp, alpha * cst_array - np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))), axis=1)
+        dnum_max = np.sum(-alpha * cst_array * non_max_idx *
+                          np.exp(np.maximum(min_exp, alpha * cst_array - np.repeat(k, cst_array.shape[1]).reshape(cst_array.shape))), axis=1)
+        # derivative of den wto cstmax is 0
+        dden_max = dden_max + 0.0
+        dnum_max = dnum_max + 1.0 * np.exp(alpha * max_elem - k)
+        grad_val_max = dnum_max / den - (num / den) * (dden_max / den)
+
+        for i in np.arange(cst_array.shape[0]):
+            grad_value[i][np.logical_not(non_max_idx)[i]] = grad_val_max[i]
 
     return grad_value