saliency_map.py 文件源码

def _jsma2_impl(model, x, yind, epochs, eps, clip_min, clip_max, score_fn):

    def _cond(k, xadv):
        return tf.less(k, epochs)

    def _body(k, xadv):
        ybar = model(xadv)

        dy_dx = tf.gradients(ybar, xadv)[0]

        # gradients of target w.r.t input
        yt = tf.gather_nd(ybar, yind)
        dt_dx = tf.gradients(yt, xadv)[0]

        # gradients of non-targets w.r.t input
        do_dx = dy_dx - dt_dx

        c0 = tf.logical_or(eps < 0, xadv < clip_max)
        c1 = tf.logical_or(eps > 0, xadv > clip_min)
        cond = tf.reduce_all([dt_dx >= 0, do_dx <= 0, c0, c1], axis=0)
        cond = tf.to_float(cond)

        # saliency score for each pixel
        score = cond * score_fn(dt_dx, do_dx)

        shape = score.get_shape().as_list()
        dim = _prod(shape[1:])
        score = tf.reshape(score, [-1, dim])

        a = tf.expand_dims(score, axis=1)
        b = tf.expand_dims(score, axis=2)
        score2 = tf.reshape(a + b, [-1, dim*dim])
        ij = tf.argmax(score2, axis=1)

        i = tf.to_int32(ij / dim)
        j = tf.to_int32(ij) % dim

        dxi = tf.one_hot(i, dim, on_value=eps, off_value=0.0)
        dxj = tf.one_hot(j, dim, on_value=eps, off_value=0.0)
        dx = tf.reshape(dxi + dxj, [-1] + shape[1:])

        xadv = tf.stop_gradient(xadv + dx)
        xadv = tf.clip_by_value(xadv, clip_min, clip_max)

        return k+1, xadv

    _, xadv = tf.while_loop(_cond, _body, (0, tf.identity(x)),
                            back_prop=False, name='_jsma2_batch')
    return xadv