gaopengcuhk / ballad Goto Github PK

Hello,

Great work! Thanks a lot for sharing the codes!

I can find the config of training on ImageNetLT, but the configs of training on PlaceLT and CIFAR10 are missing. Also, I cannot find the details of training hyperparameters in the original paper.

Could you please share the config files for PlaceLT and CIFAR10? Or you can simply tell me the details of hyperparameters here, please. Thanks for your help!

Thank you very much for sharing the source code of your great work!

Currently, this repository does not contain the LICENSE file. Could you add it so that we could know the conditions to use the code?

from clip import clip
well，where is the clip, just from the https://github.com/openai/CLIP/clip ?

Thank you for your awesome work, it's very inspiring.

I just wonder if you have any plan of releasing the checkpoint of your model? It would be nice to have them so we don't need to re-run the experiments. :)

hi, it seems like missing places_lt config and cifar_lt config in configs dir

Hi. May I ask how long does it take to train the model with setting "ResNet50×16" on ImageNet-LT using an 8xA100 or 8xV100 server?

I would appreciate it a lot if you could provide me an approximate training time.

hi , dataloader code is missing in data dir or in pwd?

Hi Authors,

Thank you for sharing your great work. I have a question related to the linear adapter proposed in the paper for Phase B.
The paper describes the linear adapter to have a bias and is followed by ReLU activation.

However, in your model description, it is just a linear layer with no bias and activation. Could you please share the best configuration of the linear adapter that helped in your experiments?

Thank you.

Hi,

I genuinely admire the compelling and innovative approach you've introduced, which has notably enhanced our understanding of the field. I'm currently conducting further research based on this research, and I wondering if you could share the code used for the CIFAR dataset. Thanks for all.

Hi @TeleeMa , thanks for your inspiring work in applying the vision-language model to long-tailed recognition.
And I have some questions about the precision of the zero-shot CLIP on the Imagenet-LT test split. In the paper, you point out the results are balanced on many-shots (59.4%), medium-shots (57.5%), and low-shots (57.6%) subsets and the overall performance (58.2%). And the performance of the original CLIP on the Imagenet are 73.3%:

So I wonder if there is any difference in settings between BALLAD and CLIP? Looking forward to your reply and thanks in advance~

Thanks for your great work!
I have a slight problem, will you release the version of CIFAR? The small dataset will be more helpful and feasible for us to train the model. Thank you!