| ---
|
| license: cc-by-4.0
|
| tags:
|
| - batchnorm
|
| - bilinear
|
| - clip
|
| - giant
|
| - kaiming
|
| - linear-warmup
|
| - novograd
|
| - retrieval
|
| - standard
|
| - swish
|
| ---
|
|
|
| # pipeline.py
|
|
|
| ## Model Overview
|
|
|
| A **giant**-scale implementation of the **clip** architecture, built for **retrieval** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: clip
|
| - **Scale**: giant
|
| - **Attention**: standard
|
| - **Fusion strategy**: bilinear
|
| - **Task head**: retrieval
|
| - **Activation**: swish
|
| - **Normalization**: batchnorm
|
| - **Initialization**: kaiming
|
|
|
| ## Training
|
|
|
| - **Optimizer**: novograd
|
| - **LR scheduler**: linear warmup
|
|
|
| ## Files
|
|
|
| - `pipeline.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
