multimodal based HuggingFace implementation for backpropagation vector.

Input

5124-dim embedding

Encoder

63 x GRU with 36 heads

Output

precision projection