保存模型时，将epoch  model  optimizer参数都保存下来

                    checkpoint = {
                       'epoch': self.step,
                       'model': self.refiner.state_dict(),
                       'optimizer': self.optim.state_dict(),
                     }           
                    self.save(cfg.ckpt_dir, cfg.ckpt_name, checkpoint)

    def save(self, ckpt_dir, ckpt_name,checkpoint):
        #print 'sfdfsdsfsdf'
        save_path = os.path.join(
            ckpt_dir, "{}_{}.pth".format(ckpt_name, self.step))

        torch.save(checkpoint, save_path) 
        #torch.save(self.refiner.state_dict(), save_path)  #只保存模型参数   
        #torch.save(self.refiner, save_path)  保存整个模型

在训练开始时加载保存的模型，训练模型参数初始化值使用读取的参数

        if cfg.resume > 0:
           checkpoint = torch.load(cfg.resume_dir)
           self.optim.load_state_dict(checkpoint['optimizer'])
           self.refiner.load_state_dict(checkpoint['model'])
           self.step = checkpoint['epoch']+1