custom stats for "binary" classification where we have more than 2 classes and want to binary classify between one and rest

binary_selected_stat.py

from mxnet import metric, nd


class BinarySelectedStatistics(metric._BinaryClassificationMetrics):
    def __init__(self):
        super().__init__()
        self.positive = 1 # default
        self.num_inst = 0
        self.sum_metric = 0.0

    def set_positive_index(self, index):
        self.positive = index

    @property
    def score(self):
        return self.sum_metric/self.num_inst
    def reset(self):
        self.reset_stats()
        self.num_inst = 0
        self.sum_metric = 0.0

    def update(self, label, outputs):
        label, outputs = metric.check_label_shapes(label, outputs, True)
        for label, pred_label in zip(label, outputs):
            pred_label = nd.argmax(pred_label, axis=1)
            pred_label = pred_label.asnumpy().astype('int32')
            label = label.asnumpy().astype('int32')
            label = label.flat
            pred_label = pred_label.flat

            # import numpy as np
            # junk_label = np.array([1,1,0,2,3,3,3,1,1,1])
            # pred_label = np.array([1,0,2,2,2,1,1,2,3,0])
            # # TP: 1, TN: 3, FP: 2, FN: 4
            # label=junk_label
            metric.check_label_shapes(label, pred_label)
            label_true = (label == self.positive)

            true_positives = (pred_label == self.positive).sum(where=label_true)
            false_positives = (pred_label == self.positive).sum(where=~label_true)
            false_negatives = (pred_label != self.positive).sum(where=label_true)
            true_negatives = (pred_label != self.positive).sum(where=~label_true)

            self.true_positives += true_positives
            self.false_positives += false_positives
            self.false_negatives += false_negatives
            self.true_negatives += true_negatives
            self.global_true_positives += true_positives
            self.global_false_positives += false_positives
            self.global_false_negatives += false_negatives
            self.global_true_negatives += true_negatives

            self.num_inst += len(pred_label)
            self.sum_metric += (pred_label == label).sum()