Gray-scale network and pre-trained model?

Hi! Thanks again for the great repo.

In the paper, you mention a gray-scale model that gets within 1% of RGB model. That's very interesting.

Are there plans to also provide a link to this model? And also -- apologies on a dumb question but how do you handle the inputs? Does the input get reduced to 1 channel -- and if so is there an option already in the model code for that? Or are you projecting the grayscale into three channels -- given that some Kinetics videos are already in B&W.

I'm fine-tuning on a problem where a lot of my data is in B&W, and would be interesting to just use grayscale for pre-training. Thank you!