sigmoid based HuggingFace implementation for model loss.

Input

2438-dim embedding

Encoder

120 x Attention-UNet with 12 heads

Output

mAP projection