How to get the following two files in search_joint.yaml?

[desertfex]

/scratch/project_2001654/wp/data/ntu/xview/train_data_joint1.npy
/scratch/project_2001654/wp/data/ntu/xview/train_data_joint2.npy
I want to run gcn_search.py, and the above two files are needed.
The ‘readme.md‘ file shows that ' Devide the training data into trainging set and searching parts, as it is for a general NAS'. But I still do not know how. Is there any code for this? Sorry to bother you. I would be greatly appreciated if you could give me some advice.

[xiaoiker]

Hello,

Thanks for interesting in our work.
If you want to try our searching part, there are several ways to prepare the data.

1. Easiest way is : do not shuffle the training dataset and directly use the first 2/3 to train and the last 1/3 to search. To this end, you can set a flag at the training part.
2. You can use the data_gen/ntu_gendata.py to customize your own data.
3. Use python (numpy) to read the pre-processed data, say xview/train_data_joint.npy, and divide it with a rule and save it.

If you just want to test this method, you can just use a very small part of the data. This datset is much larger than CIFAR, thus the search is really not that fast.

All the best,

Wei

[desertfex]

Thank you very much for your patience. I still have two questions.
The first:
Easiest way is : do not shuffle the training dataset and directly use the first 2/3 to train and the last 1/3 to search.
What's the meaning of 'training dataset'? For the ntu dataset, the total dataset is divided to two parts, 2/3 for training and 1/3 for testing. Do you mean that the 2/3 part is the 'training dataset'? Or the total dataset is the 'training dataset'?
The second: 'To this end, you can set a flag at the training part.'
I have read your codes. Is the flag you mentioned in the gcn_search.py:
if self.arg.phase == 'train':
self.data_loader['train'] = torch.utils.data.DataLoader(
dataset=Feeder(**self.arg.train_feeder_args),
batch_size=self.arg.batch_size,

            shuffle=False,

            num_workers=self.arg.num_worker,
            drop_last=True,
            worker_init_fn=init_seed)

Is 'shuffle=False' the flag?

非常感谢您的耐心解答！我有两个疑问：
一是您说的‘’the training dataset ‘’是什么呢？是说数据集中用来训练的那一部分吗？比如ntu的数据集，2/3用来训练，1/3用来测试，您是指用来训练的那2/3吗？ 然后把这2/3的数据的2/3用来训练(即/scratch/project_2001654/wp/data/ntu/xview/train_data_joint1.npy)，这2/3的数据的1/3用来搜索（即/scratch/project_2001654/wp/data/ntu/xview/train_data_joint2.npy）。是这个意思吗？
二是set a flag at the training part是什么意思呢？ 是将gcn_search.py文件中的下列代码中的shuffle设置为False吗？
if self.arg.phase == 'train':
self.data_loader['train'] = torch.utils.data.DataLoader(
dataset=Feeder(**self.arg.train_feeder_args),
batch_size=self.arg.batch_size,
shuffle=False,
num_workers=self.arg.num_worker,
drop_last=True,
worker_init_fn=init_seed)

期待能得到您的指点，非常感谢！

[xiaoiker]

Yes, exactly as you said.

Best

[Lyndsey-Xing]

/scratch/project_2001654/wp/data/ntu/xview/train_data_joint1.npy
/scratch/project_2001654/wp/data/ntu/xview/train_data_joint2.npy
I want to run gcn_search.py, and the above two files are needed.
The ‘readme.md‘ file shows that ' Devide the training data into trainging set and searching parts, as it is for a general NAS'. But I still do not know how. Is there any code for this? Sorry to bother you. I would be greatly appreciated if you could give me some advice.

Hello, have you solved it? How exactly is it done? I am still a little confused, I hope you can give me some tips.