| ---
|
| license: apache-2.0
|
| tags:
|
| - batchnorm
|
| - blip
|
| - flash
|
| - gelu
|
| - generation
|
| - large
|
| - lion
|
| - orthogonal
|
| - polynomial
|
| - tucker
|
| ---
|
|
|
| # inference.py
|
|
|
| ## Model Overview
|
|
|
| A **large**-scale implementation of the **blip** architecture, built for **generation** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: blip
|
| - **Scale**: large
|
| - **Attention**: flash
|
| - **Fusion strategy**: tucker
|
| - **Task head**: generation
|
| - **Activation**: gelu
|
| - **Normalization**: batchnorm
|
| - **Initialization**: orthogonal
|
|
|
| ## Training
|
|
|
| - **Optimizer**: lion
|
| - **LR scheduler**: polynomial
|
|
|
| ## Files
|
|
|
| - `inference.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
