| ---
|
| license: mit
|
| tags:
|
| - clip
|
| - cross-attention
|
| - gelu
|
| - instancenorm
|
| - lion
|
| - multitask
|
| - onecycle
|
| - orthogonal
|
| - sliding-window
|
| - small
|
| ---
|
|
|
| # model.py
|
|
|
| ## Model Overview
|
|
|
| A **small**-scale implementation of the **clip** architecture, built for **multitask** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: clip
|
| - **Scale**: small
|
| - **Attention**: sliding window
|
| - **Fusion strategy**: cross attention
|
| - **Task head**: multitask
|
| - **Activation**: gelu
|
| - **Normalization**: instancenorm
|
| - **Initialization**: orthogonal
|
|
|
| ## Training
|
|
|
| - **Optimizer**: lion
|
| - **LR scheduler**: onecycle
|
|
|
| ## Files
|
|
|
| - `model.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
