| ---
|
| license: mit
|
| tags:
|
| - approx-gelu
|
| - batchnorm
|
| - kaiming
|
| - lamb
|
| - linear
|
| - linear-warmup
|
| - mae
|
| - multitask
|
| - small
|
| - tensor-fusion
|
| ---
|
|
|
| # train.py
|
|
|
| ## Model Overview
|
|
|
| A **small**-scale implementation of the **mae** architecture, built for **multitask** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: mae
|
| - **Scale**: small
|
| - **Attention**: linear
|
| - **Fusion strategy**: tensor fusion
|
| - **Task head**: multitask
|
| - **Activation**: approx gelu
|
| - **Normalization**: batchnorm
|
| - **Initialization**: kaiming
|
|
|
| ## Training
|
|
|
| - **Optimizer**: lamb
|
| - **LR scheduler**: linear warmup
|
|
|
| ## Files
|
|
|
| - `train.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
