| ---
|
| license: apache-2.0
|
| tags:
|
| - adamw
|
| - batchnorm
|
| - blip
|
| - huge
|
| - kaiming-normal
|
| - matching
|
| - mish
|
| - sliding-window
|
| - step
|
| - tensor-fusion
|
| ---
|
|
|
| # finetune.py
|
|
|
| ## Model Overview
|
|
|
| A **huge**-scale implementation of the **blip** architecture, built for **matching** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: blip
|
| - **Scale**: huge
|
| - **Attention**: sliding window
|
| - **Fusion strategy**: tensor fusion
|
| - **Task head**: matching
|
| - **Activation**: mish
|
| - **Normalization**: batchnorm
|
| - **Initialization**: kaiming normal
|
|
|
| ## Training
|
|
|
| - **Optimizer**: adamw
|
| - **LR scheduler**: step
|
|
|
| ## Files
|
|
|
| - `finetune.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|