layer.py 文件源码

def bn_backward(dout, cache):
    X, X_norm, mu, var, gamma, beta = cache

    N, D = X.shape

    X_mu = X - mu
    std_inv = 1. / np.sqrt(var + c.eps)

    dX_norm = dout * gamma
    dvar = np.sum(dX_norm * X_mu, axis=0) * -.5 * std_inv**3
    dmu = np.sum(dX_norm * -std_inv, axis=0) + dvar * np.mean(-2. * X_mu, axis=0)

    dX = (dX_norm * std_inv) + (dvar * 2 * X_mu / N) + (dmu / N)
    dgamma = np.sum(dout * X_norm, axis=0)
    dbeta = np.sum(dout, axis=0)

    return dX, dgamma, dbeta