Created by: lcy-seso

I use this PR https://github.com/PaddlePaddle/Paddle/pull/3085 to print parameter values and gradients in event handle.

My usage is as the follows:

def show_parameter_status(parameters):
    # for debug print
    for p in parameters:
        value = parameters.get(p)
        grad = parameters.get_grad(p)

        avg_abs_value = np.average(np.abs(value))
        avg_abs_grad = np.average(np.abs(grad))

        logger.info(
            ("%s avg_abs_value=%.6f avg_abs_grad=%.6f "
             "min_value=%.6f max_value=%.6f min_grad=%.6f max_grad=%.6f") %
            (p, avg_abs_value, avg_abs_grad, value.min(), value.max(),
             grad.min(), grad.max()))

• But I found when set use_gpu=True in train.init(), get_grad always returns an all-zero matrix.
• when I change use_gpu=False and keep all the other things unchanged, it returns a non-zeros matrix.
• My training cost decreases so I think the gradient matrices should be non-zeros.