accuracy based ONNX implementation for deep attention-head.

Input

5832-dim embedding

Encoder

86 x ResNet with 30 heads

Output

recall projection