| ---
|
| license: bsd-3-clause
|
| tags:
|
| - adafactor
|
| - bilinear
|
| - deit
|
| - gelu-tanh
|
| - layernorm
|
| - matching
|
| - multi-query
|
| - step
|
| - tiny
|
| - xavier-uniform
|
| ---
|
|
|
| # predict.py
|
|
|
| ## Model Overview
|
|
|
| A **tiny**-scale implementation of the **deit** architecture, built for **matching** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: deit
|
| - **Scale**: tiny
|
| - **Attention**: multi query
|
| - **Fusion strategy**: bilinear
|
| - **Task head**: matching
|
| - **Activation**: gelu tanh
|
| - **Normalization**: layernorm
|
| - **Initialization**: xavier uniform
|
|
|
| ## Training
|
|
|
| - **Optimizer**: adafactor
|
| - **LR scheduler**: step
|
|
|
| ## Files
|
|
|
| - `predict.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
