CMU-MOSEI: Updated csd files give different feature tensors compared to before

[anita-hu]

On Jul 4 2020, I stored processed aligned tensors for the visual and audio features and used the raw words to create bert features. I noticed recently that a newly created dataset using the same code results in different features (I am using COVAREP, VisualFacet42 and TimestampedWords). It seems the csd files (in http://immortal.multicomp.cs.cmu.edu/CMU-MOSEI) have different file sizes and were last updated on Jul 20. I was wondering what changed with the feature csd files? I noticed the labels are the same when loaded as tensors but features are very different.

Thank you!

[A2Zadeh]

Hi @anita-hu,

What we are hoping to achieve with the SDK is to keep updating the computational sequences with the newest features and most recent updates. These updates are mostly meant to make the features better and better over time. If by any chance you found any bugs in the new csd, please let me know.

[anita-hu]

Hi @A2Zadeh,

Thank you for your response. Could you provide details on what was changed in the most recent MOSEI csd files after version 1.2.0? I noticed that the loaded features are different which led to not being able to recreate training results I had before using the previous version.