cdoersch

Going to reopen this because I may improve this myself in the near future. Contributions are also welcome.

@ZHAOZHIHAO I created a PR (#285 ) which should fix these issues (and also fixes a bug with the prior PR). If you have a chance, could you test the...

I merged #285, as I think #284 actually made the occlusion estimation situation worse, and I don't want people training on a dataset with those issues. I'm still interested in...

I guess we didn't test this codepath; this is more intended for evaluating checkpoints you've trained yourself. The logic for restoring the Experiment object from the checkpoint is here: https://github.com/google-deepmind/tapnet/blob/main/utils/experiment_utils.py#L160...

Interesting -- I'm fairly sure that isn't a correct initialization for opt_state (it should be an Adam state), but for evaluation I guess it doesn't matter? IIRC Jaxline will call...

Batched inference is supported. The error you're getting, however, suggests that you're taking a query point from one video and using transforms.convert_grid_coordinates to convert it to a different *framerate* for...

Internally we train using JAX. Maintaining a second training codebase would be non-trivial work for us and is beyond the capacity of our team. Sorry about that.

Could you post a full stack trace? I don't see anything about this setup that shouldn't work.

We weren't able to reproduce this error internally; I understand that the model should be loadable using the current instructions. I'm closing this, but feel free to reopen if similar...

Now that tapir_clustering.py is fixed, I've run the colab at head and verified that the code will plot more than 20 tracks. Your snippets above look correct to me--I don't...