zxcvfd13502 / ddas_code Goto Github PK

Hi
Your article is very interesting. However, I just want to check my search policies. It seems that Figure4 in your article may help with this.
May I know how can I do the sanity check experiment.? Do you have a plan to release the code anything related to it?
I tried to plot policy probability vs epoch. It is not the same as yours. Can you help?
Thanks.

in the example , search is with model wresnet40x2, but traininig is wresnet28x10, why don't use the same model while searching and training?

Hi,
Does any place has load pretrained models from imagenet or somewhere else?

proxy task models or Train models?

if not have any pretrainde models, why?

decrease init tp，the search result of tp is also decrease; increase init tp，the search result of tp is also increase; it is highly dependent on init tp?

Thanks for your good work! When I run the code, an problem arise, 'ModuleNotFoundError: No module named 'theconf''. Could you provide the thecof file?

Hi. Thank you for your work.

When I tried to visualize the augmented images, it appears that each operation has resulted in same kind of augmentation.
For example, ShearX.0, ShearX.4, ShearX.1
0, and ShearX.14 are similar. The same goes for Rotation, ShearY, and translateX and Y.
Do you have any suggestions? Please have a look at the image.

Hi Aoming,
thanks for sharing this nice work!
I have two question about search.py. Looking forward for your reply.

1. Why you set the hyper parameter:explore_ratio as 0.99999 in search.py, which may cause the aug_types being random chose.
2. In your arxiv paper, you say"We firstly search for augmentation policy on a proxy task with a small network, Wide-ResNet-40-2 [40], on part of the dataset for 20 epochs, with 5000 images for training and 5000 images for validation.". However, I only get 4000 for training and 500 for validation in search.py.