Created by: aiyouwoqu123

paddlepaddle 1.5.1 在训练过程中，需要加载自己预训练好的模型，来当作当前任务中的一个inference模块，但是出现了以下问题

def inference(img):
    adv_program = fluid.Program()
    with fluid.program_guard(adv_program,fluid.Program()):
        input_layer = fluid.layers.data(name='img', shape=image_shape, dtype='float32')
        model = models.__dict__[model_name]()
        out_logits = model.net(input=input_layer, class_dim=class_dim)
        out = fluid.layers.softmax(out_logits)
        place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
        exe = fluid.Executor(placel)
        exe.run(fluid.default_startup_program())
        fluid.io.load_persistables(exe_cl, pretrained_model)
        eval_program = adv_program.clone(for_test=False)
        fetch_list = [out.name]
    result = exe.run(eval_program,
                            fetch_list=fetch_list,
                            feed={"img": img})
    result = result[0][0]
    pred_label = np.argmax(result)
    pred_score = result[pred_label]
    return pred_label, pred_score
    

def main():
    ...
    with fluid.program_guard(train_g, startup):
        img = fluid.layers.data('image', shape=[3,224,224])
        perturbation = Generator(img)
        fake_img = fluid.layers.clip(perturbation, -0.3, 0.3) + img
        fake_img = fluid.layers.clip(fake_img, 0., 1.)
        infer_program = train_g.clone(for_test=True)
        p = Discriminator(fake_img)
         ones = fluid.layers.fill_constant_batch_size_like(p,shape=[-1,1,24,24],dtype='float32', value=1)
        g_cost = fluid.layers.reduce_mean(fluid.layers.square_error_cost(p, ones*0.9))
        g_avg_cost = fluid.layers.mean(g_cost)
        fluid.layers.reshape(perturbation,[perturbation.shape[0],perturbation.shape[1]*perturbation.shape[2]*perturbation.shape[3]])
        C = 0.1
        loss_perturb = fluid.layers.mean(fluid.layers.l2_normalize(fluid.layers.reshape(perturbation,[perturbation.shape[0],perturbation.shape[1]*perturbation.shape[2]*perturbation.shape[3]]),axis=1))
        pred_label, pred_score = inference(fake_img) 
    ...