Thank you for open-sourcing this work.

From the paper:

• What does $C$ contain? The classes on which SWAG pre-training was performed?
• How is $p$ obtained?

Is it possible to also accompany this repository with a notebook that shows how to adapt the pre-trained models for zero-shot transfer? I don't think there is a good example available on the internet that shows how to extend an image recognition model like SWAG for zero-shot classification.

@mannatsingh

This pull request makes it possible to run your model inside a Docker environment, which makes it easier for other people to run it. We're using an open source tool called Cog to make this process easier.

This also means we can make a web page where other people can try out your model! View it here: https://replicate.com/facebookresearch/swag. We enable selecting different models for inference, and you can find the docker file under the tab ‘run model with docker’.

We have added some examples to the page, but do claim the page so you can own the page, customise the Example gallery as you like, push any future update to the web demo, and we'll feature it on our website and tweet about it too.

In case you're wondering who I am, I'm from Replicate, where we're trying to make machine learning reproducible. We got frustrated that we couldn't run all the really interesting ML work being done. So, we're going round implementing models we like. 😊

Hello,

I've tried reproducing the reported fine-tuning for RegNetY-16GF but obtained only 50% top-1 accuracy on ImageNet-1K at the end of training. I used the following hyper-parameters, taken from the paper:

# lr: 6e-3
# lr schedule: cosine
# batch size: 512
# weight-decay: 0
# resolution: 384x384
# training steps: 20000 ~= 8 epochs
# mixup: 0.1
# sync batchnorm
# EMA: decay 1e-4

I have some doubt regarding the number of training steps, which seems to be too low. 20K training steps is equivalent to 8 epochs:

num_epochs = batch_size x num_steps / num_images = 512 x 20000/1281167 ~= 8

Could you please confirm if this was indeed your setting?

Thank you in advance!

Hello,

This is great work, congratulations! And thank you for having released the models.

I would like to ask some questions please:

1. Is there a particular reason for choosing 518x518 as crop size for fine-tuning ViT H/14? This value doesn't seem to be very standard to be (I would have expected something like 500x500 or 512x512). Could you tell me if you did a hyper-parameter search for this value?
2. How much do you expect the accuracy to drop if we use 500x500 instead of 518x518?

Thank you very much in advance for your response.

The paper uses 28K canonical hashtags (each correspond to a set of wordnet synsets). I was wondering whether that can be released to foster future research in this area! Thanks

I would like to know if you could share your Classy Vision fine-tuning recipes (i.e., the *.json configs) for the trained weights listed in the README.

Thank you in advance. Best regards.