lifted_struct_loss.py 文件源码

def lifted_struct_loss(f_a, f_p, alpha=1.0):
    """Lifted struct loss function.

    Args:
        f_a (~chainer.Variable): Feature vectors as anchor examples.
            All examples must be different classes each other.
        f_p (~chainer.Variable): Positive examples corresponding to f_a.
            Each example must be the same class for each example in f_a.
        alpha (~float): The margin parameter.

    Returns:
        ~chainer.Variable: Loss value.

    See: `Deep Metric Learning via Lifted Structured Feature Embedding \
        <http://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/\
        Song_Deep_Metric_Learning_CVPR_2016_paper.pdf>`_

    """
    assert f_a.shape == f_p.shape, 'f_a and f_p must have same shape.'
    n = 2 * f_a.shape[0]  # use shape[0] due to len(Variable) returns its size
    f = F.vstack((f_a, f_p))
    D_sq = squared_distance_matrix(f)

    pairs_p = np.arange(n).reshape(2, -1)  # indexes of positive pairs
    row = []
    col = []
    for i, j in pairs_p.T:
        row.append([i] * (n - 2) + [j] * (n - 2))
        col.append(np.tile(np.delete(np.arange(n), (i, j)), 2))
    row = np.ravel(row)
    col = np.ravel(col)
    pairs_n = np.vstack((row, col))

    distances_p = F.sqrt(D_sq[pairs_p[0], pairs_p[1]])
    distances_n = F.sqrt(D_sq[pairs_n[0], pairs_n[1]])
    distances_n = distances_n.reshape((n // 2, -1))
    loss_ij = F.logsumexp(alpha - distances_n, axis=1) + distances_p
    return F.sum(F.relu(loss_ij) ** 2) / n