undersampling added

cnn.py

@@ -5,6 +5,8 @@ from matplotlib.pyplot \
     import figure,plot,title,ylabel,xlabel,legend,savefig,ioff
     import figure,plot,title,ylabel,xlabel,legend,savefig,ioff
 from numpy import expand_dims as dims
 from numpy import expand_dims as dims
 from numpy import unique
 from numpy import unique
+from random import shuffle
+from functools import reduce
 
 
 from prec_recall import create_pr, avg_pr
 from prec_recall import create_pr, avg_pr
 from sklearn.metrics import confusion_matrix
 from sklearn.metrics import confusion_matrix
@@ -13,6 +15,76 @@ from sklearn.utils import class_weight
 from imblearn.over_sampling import SMOTE,ADASYN
 from imblearn.over_sampling import SMOTE,ADASYN
 from tensorflow.keras.utils import to_categorical
 from tensorflow.keras.utils import to_categorical
 
 
+def undesample(X,y):
+    """
+    Balances the input and output data by removing
+    samples from the more represented classes
+
+    Parameters
+    ----------
+    X : numpy array
+        Input data.
+    y : numpy array
+        class membership.
+
+    Returns
+    -------
+    numpy array
+        Balanced input.
+    numpy array
+        Balanced output.
+
+    """
+    locs = find_classes(y)
+    shuffle_members(locs)
+    cmin,members = min(locs.items(),key=lambda x:len(x[1]))
+    # maximum number of samples to keep cardinality of classes balanced
+    n = len(members)
+    print('Undersampling to %d samples'%n)
+    indices = list(reduce(lambda x,y:x+y,map(lambda x:x[:n],locs.values())))
+    shuffle(indices)
+    return X[indices],y[indices]
+
+def shuffle_members(cm):
+    """
+    Shuffle the members of each class in place
+
+    Parameters
+    ----------
+    cm : dictionary
+        keys are classes, values are list of indices of its members.
+
+    Returns
+    -------
+    None
+
+    """
+    for c,_ in cm.items():
+        shuffle(cm[c])
+        print("cardinality of class %d is %d"%(c,len(cm[c])))
+
+def find_classes(x):
+    """
+    Find the indices of members of each class
+
+    Parameters
+    ----------
+    x : iterable 
+        heterogeneous memberships.
+
+    Returns
+    -------
+    locs : list
+        dictionary with the locations of the members for each class.
+    """
+    locs = {}
+    for i,c in enumerate(x):
+        if c in locs:
+            locs[c].append(i)
+        else:
+            locs[c] = [i]
+    return locs
+
 def encode(output,N=4):
 def encode(output,N=4):
     """
     """
     One hot encoding of the input
     One hot encoding of the input
@@ -66,6 +138,9 @@ def balance(X,y,method):
         print('CLASS WEIGHTS')
         print('CLASS WEIGHTS')
         weights = class_weight.compute_class_weight('balanced',unique(y),y)
         weights = class_weight.compute_class_weight('balanced',unique(y),y)
         return (X,y),weights
         return (X,y),weights
+    elif method == 'undersampling':
+        print('UNDERSAMPLING')
+        return undesample(X,y),None
     
     
 class CNN_Antifrag:
 class CNN_Antifrag:
     """
     """

main.py

@@ -14,11 +14,11 @@ o belongs to {0,1,2} and represents the balancing method
 """
 """
 
 
 c_ = [0,1,2,3]
 c_ = [0,1,2,3]
-b_ = [1]
-e_ = [0]
-o_ = [0,1,2]
+b_ = [0,1]
+e_ = [0,1]
+o_ = [0,1,2,3]
 
 
-n_experiences = 100
+n_experiences = 10000
 combinations = itertools.product(c_,b_,e_,o_)
 combinations = itertools.product(c_,b_,e_,o_)
 
 
 
 
@@ -55,4 +55,3 @@ f = open("out/acc_auc.txt", 'w+')
 f.write("Average accuracy: " + str(acc)+"\n")
 f.write("Average accuracy: " + str(acc)+"\n")
 f.write("Average area under the curve: " + str(auc))
 f.write("Average area under the curve: " + str(auc))
 f.close()
 f.close()
\ No newline at end of file
-

models.py

@@ -81,3 +81,5 @@ def choose_balancing_method(o):
         return 'adasyn'
         return 'adasyn'
     elif o == 2:
     elif o == 2:
         return 'class_weight'
         return 'class_weight'
+    elif o == 3:
+        return 'undersampling'
\ No newline at end of file