我也遇到关于恢复训练的问题，在训练过程中能够保存、重新载入保存文件、评估都没有问题，但是训练结束后单独进行载入、评估，损失就很高，我感觉应该是模型还有东西没有保存始终没有看出问题，不知道哪里出问题了。

    def save_model(self):
        save_folder='result'
        if not os.path.exists(save_folder):
            os.makedirs(save_folder)
        for model_name, model in self.models.items():
            save_path = os.path.join(save_folder, '{}.pdparams'.format(model_name))
            to_save = model.state_dict()
            paddle.save(to_save, save_path)
        save_path = os.path.join(save_folder, "{}.pdopt".format('sgd'))
        paddle.save(self.opt.state_dict(), save_path)
        
    def load_model(self):
        model_folder='result'
        model_names=['encoder','decoder']
        for model_name in model_names:
            model_path = os.path.join(model_folder, "{}.pdparams".format(model_name))
            model_state = paddle.load(model_path)
            self.models[model_name].set_state_dict(model_state)
        opt_path = os.path.join(model_folder, "{}.pdopt".format('sgd'))
        opt_state=paddle.load(opt_path)
        self.opt.set_state_dict(opt_state)