Our next steps for the Multimodal SDK
Computational Sequence
- Refactoring some of the code ++ Loose function cleanups
- Adding compression ++ Adding the chunks on hdd
- Adding json metadata
- Allow in-hdd access
Data Processing
- Dimension names for computational sequences - done - doing final checks before commit
- BERT, and GloVe (updates) for CMU-MOSI, CMU-MOSEI, POM - delayed due to better alignment now available
- Adding passive alignment for RAVEN style models
- Adding numpy style key search [x,1,2] => video x segment number 2, fine segment number 3
Alignment Functionalities
- A faster implementation of unify based on muliple set intersection
- Intermediate goal: accomodate low-ram setups by forcing the h5py to remain on the hdd <= done, needs final testing as it broke the alignment + setitem and remove items
High priority fixes None ATM