| ---
|
| license: mit
|
| tags:
|
| - adafactor
|
| - contrastive
|
| - efficientformer
|
| - exponential
|
| - flash
|
| - gelu
|
| - orthogonal
|
| - scalenorm
|
| - tensor-fusion
|
| - xlarge
|
| ---
|
|
|
| # pipeline.py
|
|
|
| ## Model Overview
|
|
|
| A **xlarge**-scale implementation of the **efficientformer** architecture, built for **contrastive** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: efficientformer
|
| - **Scale**: xlarge
|
| - **Attention**: flash
|
| - **Fusion strategy**: tensor fusion
|
| - **Task head**: contrastive
|
| - **Activation**: gelu
|
| - **Normalization**: scalenorm
|
| - **Initialization**: orthogonal
|
|
|
| ## Training
|
|
|
| - **Optimizer**: adafactor
|
| - **LR scheduler**: exponential
|
|
|
| ## Files
|
|
|
| - `pipeline.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
