trainer.py 文件源码

def ordinal_loss(y, mask):
    xp = cuda.get_array_module(y.data)
    volatile = y.volatile
    b, c, n = y.data.shape
    max_y = F.broadcast_to(F.max(y, axis=1, keepdims=True), y.data.shape)
    y = y - max_y
    sum_y = F.broadcast_to(F.expand_dims(F.sum(y, axis=1), 1), y.data.shape)
    down_tri = np.tri(c, dtype=np.float32)
    up_tri = down_tri.T
    w1 = Variable(xp.asarray(down_tri.reshape(c, c, 1, 1)), volatile=volatile)
    w2 = Variable(xp.asarray(up_tri.reshape(c, c, 1, 1)), volatile=volatile)
    h = F.exp(F.expand_dims(y, -1))
    h1 = F.convolution_2d(h, w1)
    h1 = F.convolution_2d(F.log(h1), w1)
    h2 = F.convolution_2d(h, w2)
    h2 = F.convolution_2d(F.log(h2), w2)
    h = F.reshape(h1 + h2, (b, c, n))
    return F.sum((h - sum_y - y) * mask) / b