Kinetics 700

[daniel-j-h]

Related to #6: let's get the Kinetics-700 in.

@Showmax I updated

classes.json
kinetics_test.json
kinetics_train.json
kinetics_val.json

based on the official Kinetics 700 dataset. But the json files are becoming too big to host them in the Github repository (remote: error: File resources/kinetics_train.json is 158.91 MB; this exceeds GitHub's file size limit of 100.00 MB). Where do you prefer me putting them? Into git lfs?

Then a quick question regarding categories.json: I could not find this file in the Kinetics 700 dataset; how did you generate it? Manually? How can we update it to Kinetics 700?

Opening this pull request already so we can discuss here.

Thanks!

[ondrejbiza]

Thanks for the pull request! Maybe we can write a script that downloads the jsons and generates the metadata? I think I made categories.json with a modified version of list_categories.py.

[daniel-j-h]

The json (and other files) are all bundled up here - It'd be best to host the files we care about e.g. on this Github repository (as release files), then we could download from it easily.

https://storage.googleapis.com/deepmind-media/research/Kinetics_700.zip

[ondrejbiza]

I've never done that before but I can take a look over the weekend--hopefully.

[ondrejbiza]

Hi Daniel,
I think we don't actually need to release any big files. The train.json from Kinetics-700 can be compressed into 20 MB. We just need to add a simple bash script to decompress it or instructions to do it by hand.

However, there are three small issues:

1. We don't have the classes.json file for Kinetics-700. When I made it for Kinetics-400, I think I just copied all the classes from the paper and formatted them to json by hand. Could you do the same for Kinetics-700?

2. I guess categories aren't really defined for Kinetics-700. We could probably just put everything into a single category "all" to be backwards compatible with the scripts.

3. Do you think it's worth it to keep the option to download Kinetics-400 only? I can code up a switch between the two datasets after you make the classes.json.

Thanks,
Ondrej

[karenli1995]

@ondrejba Here is the classes.json for Kinetics700:
https://gist.github.com/karenli1995/fc410a304ab2a856631fa73f6a02801f