SG-MLP

The Official Pytorch Implementation for Switch Gated Multi-Layer Perceptron(SG-MLP).

SG-MLP, a novel and attentionless architecture for Natural Language Understanding(NLU), achieves decent results in the GLUE benchmark without any help of the Attention Mechanism in both Pre-Training and FineTuning steps. The following repositiory contains demos, pretrained models, and supplementaries necessary for reproducing the results.

Model Config & Pretrained Weights

We trained a total of three models SG-MLP Small, SG-MLP Base and SG-MLP Large.
The following are the configuration for each models. Pretrained weights for all models are available here.

SG-MLP	Parameter	Tokenizer	Corpus	Train Steps
SG-MLP Small	67 M	bert-base-cased	Book Corpus + Wiki	110,000
SG-MLP Base	125 M	roberta-base	C4	200,000
SG-MLP Large	170 M	roberta-base	C4	200,000

1. Load SG-MLP Base

from SGMLP.models.model import build_base_model
from SGMLP.utils import apply_weight

PATH = '/weights/SGMLP_Base.pth'
base_model = build_base_model()
base_model = apply_weight(base_model,PATH)

2. Load SG-MLP Large

from SGMLP.models.model import build_large_model
from SGMLP.utils import apply_weight

PATH = '/weights/SGMLP_Large.pth'
large_model = build_large_model()
large_model = apply_weight(large_model,PATH)

Masked Language Modeling(MLM) Demo

SG-MLP trained on the C4 corpus, learns to predict proper grammer and commonsense knowledge. Refer to the codes below, or the Colab Notebook for a MLM demo ran by our model. (Make sure you have our pretrained model downloaded for the demo)

from SGMLP.models.model import build_large_model
from SGMLP.utils import SGMLP_inference, apply_weight

PATH = '/weights/SGMLP_Large.pth'
large_model = build_large_model(output_logits = True)
large_model = apply_weight(large_model,PATH)

SGMLP_inference('A bird has <mask> legs.',large_model)

Contributors

• 김승원 - Seungone Kim
• 손규진 - GUIJIN SON
• 주세준 - SE JUNE JOO
• 조우진 - WOOJIN CHO
• 채형주 - Hyungjoo Chae

References

@InProceedings{Zhu_2015_ICCV,
    title = {Aligning Books and Movies: Towards Story-Like Visual Explanations by Watching Movies and Reading Books},
    author = {Zhu, Yukun and Kiros, Ryan and Zemel, Rich and Salakhutdinov, Ruslan and Urtasun, Raquel and Torralba, Antonio and Fidler, Sanja},
    booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
    month = {December},
    year = {2015}
}

@InProceedings{wikitext,
    author={Stephen, Merity and Caiming ,Xiong and James, Bradbury and Richard Socher}
    year=2016
}

@article{2019t5,
    author = {Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu},
    title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
    journal = {arXiv e-prints},
    year = {2019},
    archivePrefix = {arXiv},
    eprint = {1910.10683},
}

@InProceedings{wang2019glue,
  title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding},
  author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R.},
  note={In the Proceedings of ICLR.},
  year={2019}
}