dongzelian / t-nas Goto Github PK

hi, i have viewed the code, but i have some questions about the settings of the training for the task-specific architecture. Since the model have seen specific the test task（like 5 test classes) T while updating the Theta, why the model still use the whole meta test tasks (only 20 test classes) set for evaluation? I think it is a data leak. Could you explain about it?
[https://github.com/dongzelian/T-NAS/blob/master/train_tnas_miniimagenet.py](https://github.com/dongzelian/T-NAS/blob/master/train_tnas_miniimagenet.py

cant find learner.py file in dir,so that the project does not work properly,
can you give me this file by email([email protected]) ，thanks?

line 128
train_accs_theta, train_accs_w = meta_train(train_loader, maml, device, epoch, writer)
However,the function "meta_train" returns only one value named "accs", so is there any solutions? Look forward to your reply.

As title.

Dear Dongze Lian:
Sorry to be a bother. I am a Grade-One graduate student from NWPU, your excellent work has inspired me a lot. Thank you! However, I have met some problems when I reappeared your results.
i ) For part of AutoMAML, my results is 49.44% for 1-shot and 61.96% for 5-shot, which is about 2 percent below your result, can you give me some suggestions about the super-parameters' value?
ii ) For part of 'train_tnas_from_scrach', there lost some file(such as 'MiniImagenet_task.py' and 'meta_nas_train.py'), can you send for me? Thank you, my email address is [email protected]
I will be appreciated for your reply, thank you very much!
Gaoyitao

It's a nice work. Looking forward your source code!

# batch_size here means total episode number
mini = MiniImagenet(args.data_path, mode='train', n_way=args.n_way, k_shot=args.k_spt,
                    k_query=args.k_qry,
                    batch_size=args.batch_size, resize=args.img_size)
mini_valid = MiniImagenet(args.data_path, mode='train', n_way=args.n_way, k_shot=args.k_spt,
                    k_query=args.k_qry,
                    batch_size=args.test_batch_size, resize=args.img_size)

train_loader = DataLoader(mini, args.meta_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True)
valid_loader = DataLoader(mini_valid, args.meta_test_batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True)

mini and mini_valid are sampled from train dataset, but mini is used for meta_train and mini_valid is used for meta_test. So if mini_valid is a part of mini, it is unfair. And I divide the train dataset into 80% and 20%. mini is sampled from 80% and mini_valid is sampled from 20%. Finally acc by running train_tnas.sh is 0.57 but the original acc by runing train_tnas.sh is 0.60.

Hi, thanks for your inspiring work

I have one question about the automl algorithm in the appendix:

while in meta train, there are two repeat inner weights update procedures. Do these two update procedures sever different purposes: the first one for outer weights $\widetilde{w} $ update and the second one for outer architecture $\theta$ update. That is to say, the first step from M for the second inner weight update starts at the finishing of the first outer weight $\widetilde{w} $ update, i.e. $\widetilde{w} = w^{m}_i$(m is the beginning in the second inner weight update)