search

nupurkmr9 / S2M2_fewshot

Other
112 stars 20 forks source link

How to prepare dataset for tiredImagenet #16

Closed YounkHo closed 3 years ago

YounkHo commented 3 years ago

Hello! Thanks for this incredible contribution.

I want to know how to train on tired-imagenet, I've downloaded tired-imagenet from here, but the format is pkl and no train/vaild provided.

nupurkmr9 commented 3 years ago

Hi, you can find the split at https://github.com/renmengye/few-shot-ssl-public/tree/master/fewshot/data/tiered_imagenet_split

Hope this helps :)

YounkHo commented 3 years ago

I've already downloaded this in fact and how should I use it to train a model? As what I do as mini-ImageNet?

karayzhou commented 3 years ago

hello, thanks for your great works, i want to know some training processes about the tieredImageNet pretrain model. How long has your model been trained in tieimagenet，and how many gpus were used，can you give some details, thank you

nupurkmr9 commented 3 years ago

Hi @Dreamcreationman, Sorry for the late reply. I added tieredImageNet later on in the repo and I do not have the script similar to make_json.py saved now. But you can probably write a script to create the json with 'image_names' and 'image_labels' for tieredImageNet based on the train, val split provided on the above repo. I will try to add the script in the repo as well.

@karayzhou, we trained tieredImageNet for 100 epochs with rotation self-supervision and then fine-tuned with manifold-mixup for another 100 epochs. It was trained with other hyper-params same as miniImageNet i.e. batch-size 64. I don't remember which GPU we used, but I guess something like 4 V100 GPUs should probably be enough for training.

karayzhou commented 3 years ago

Thank you 😄

nupurkmr9 @.***> 于2021年6月1日周二 下午11:00写道：

Hi @Dreamcreationman https://github.com/Dreamcreationman, Sorry for the late reply. I added tieredImageNet later on in the repo and I do not have the script similar to make_json.py saved now. But you can probably write a script to create the json with 'image_names' and 'image_labels' for tieredImageNet based on the train, val split provided on the above repo. I will try to add the script in the repo as well.

@karayzhou https://github.com/karayzhou, we trained tieredImageNet for 100 epochs with rotation self-supervision and then fine-tuned with manifold-mixup for another 100 epochs. It was trained with other hyper-params same as miniImageNet i.e. batch-size 64.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/nupurkmr9/S2M2_fewshot/issues/16#issuecomment-852196979, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALYRVCESDPMQ6Y5JVYY7U2DTQTYYRANCNFSM44JT5YSQ .