| ---
|
| license: mit
|
| tags:
|
| - constant-warmup
|
| - contrastive
|
| - cross-attention
|
| - large
|
| - layernorm
|
| - mae
|
| - orthogonal
|
| - relu
|
| - sgd
|
| - sliding-window
|
| ---
|
|
|
| # eval.py
|
|
|
| ## Model Overview
|
|
|
| A **large**-scale implementation of the **mae** architecture, built for **contrastive** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: mae
|
| - **Scale**: large
|
| - **Attention**: sliding window
|
| - **Fusion strategy**: cross attention
|
| - **Task head**: contrastive
|
| - **Activation**: relu
|
| - **Normalization**: layernorm
|
| - **Initialization**: orthogonal
|
|
|
| ## Training
|
|
|
| - **Optimizer**: sgd
|
| - **LR scheduler**: constant warmup
|
|
|
| ## Files
|
|
|
| - `eval.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
