bart-forecast / README.md
lhoffmann's picture
Upload 2 files
4fae0d6 verified
|
Raw History Blame Contribute Delete
720 Bytes
metadata
license: cc-by-4.0
tags:
  - coca
  - cosine
  - dilated
  - giant
  - instancenorm
  - low-rank
  - multitask
  - novograd
  - orthogonal
  - swish

finetune.py

Model Overview

A giant-scale implementation of the coca architecture, built for multitask tasks.

Architecture

  • Architecture: coca
  • Scale: giant
  • Attention: dilated
  • Fusion strategy: low rank
  • Task head: multitask
  • Activation: swish
  • Normalization: instancenorm
  • Initialization: orthogonal

Training

  • Optimizer: novograd
  • LR scheduler: cosine

Files

  • finetune.py — main artifact of this repository

License

See the license field above.