about original buffer

[Kciiiman]

Hello, regarding the code in distill_mtt.py, I found that the net_eval we use is a pre-trained buffer on the entire dataset. If we rely on the buffer previously trained on the complete dataset, will it mask the independent verification ability of the synthetic data on the model training effect, and will it overestimate the effectiveness of the synthetic data?

net_eval = get_network(model_eval, channel, num_classes, im_size).to(args.device) # get a random model

                    for b, p in zip(buffer[it_eval][0], net_eval.state_dict().items()):
                        # load the same weights
                        p[1].copy_(b.data)

                    eval_labs = label_syn
                    with torch.no_grad():
                        image_save = image_syn
                    image_syn_eval, label_syn_eval = copy.deepcopy(image_save.detach()), copy.deepcopy(eval_labs.detach()) # avoid any unaware modification
                    saved_epoch_eval_train = args.epoch_eval_train
                    args.epoch_eval_train = args.override_epoch_eval_train
                    teacher_net, acc_train, acc_test = evaluate_synset(it_eval, net_eval, image_syn_eval, label_syn_eval, testloader, args, texture=args.texture)
                    args.epoch_eval_train = saved_epoch_eval_train