Pre-trained models with R101, NLN, etc

Hi, thank you very much for this wonderful codebase, the pre-trained checkpoints, and the enhancements since the first release last Fall 2019.

You do not owe us anything, and certainly not anything more. But could you please inform us if the larger checkpoints may ever be linked in the `Model Zoo`? The `R101` Kinetics models, as well as those trained on more frames, the ones with `NLN` etc -- have been marked as `coming soon` for a many months now.

These results are included in your original paper, so the model checkpoints surely exist. If there are any issues in releasing these models, I know perhaps you can not tell us. But if these weights may be sitting in a directory, waiting to be linked, I am sure we are not the only ones who would appreciate trying these larger checkpoints -- in places where the `R50` pre-trained Kinetics models already do quite well.

PS I was able to use your `R101` AVA model and make a hack to adapt it back to Kinetics format. That works ok, but not as good as we are not trying to make predictions for bounding boxes.

Of course we can train Kinetics from scratch, but this is expensive, the dataset is not so easy to build, and training from scratch may not make a good checkpoint on the first attempt.

Thank you! This great project is very much appreciated.