base.py 文件源码

def noisy_and(self, num_classes, trainable=True):
        """ Multiple Instance Learning (MIL), flexible pooling function
        :param num_classes: int, determine number of output maps
        """
        assert self.input.get_shape()[3] == num_classes  # input tensor should have map depth equal to # of classes
        scope = 'noisyAND'
        with tf.variable_scope(scope):
            a = self.const_variable(name='a', shape=[1], value=1.0, trainable=trainable)
            b = self.const_variable(name='b', shape=[1, num_classes], value=0.0, trainable=trainable)
            mean = tf.reduce_mean(self.input, axis=[1, 2])
            self.input = (tf.nn.sigmoid(a * (mean - b)) - tf.nn.sigmoid(-a * b)) / (
                tf.sigmoid(a * (1 - b)) - tf.sigmoid(-a * b))
        print(scope + ' output: ' + str(self.input.get_shape()))