def subsample_()

in datasets.py [0:0]

• 20 lines of code
• 7 McCabe index (conditional complexity)

    def subsample_(self, subsample_what):
        perm = torch.randperm(len(self)).tolist()

        if subsample_what == "groups":
            min_size = min(list(self.group_sizes))
        else:
            min_size = min(list(self.class_sizes))

        counts_g = [0] * self.nb_groups * self.nb_labels
        counts_y = [0] * self.nb_labels
        new_i = []
        for p in perm:
            y, g = self.y[self.i[p]], self.g[self.i[p]]

            if (
                subsample_what == "groups"
                and counts_g[self.nb_groups * int(y) + int(g)] < min_size
            ) or (subsample_what == "classes" and counts_y[int(y)] < min_size):
                counts_g[self.nb_groups * int(y) + int(g)] += 1
                counts_y[int(y)] += 1
                new_i.append(self.i[p])

        self.i = new_i
        self.count_groups()