🥶VILIO🥶

State-of-the-art Visio-Linguistic Models 🥶

Updates

06/2021 - Hateful Memes CSV Files

• The CSV files that were used for the scores in the vilio paper are now available here

06/2021 - Inference on any meme

• Thanks to the initiative by katrinc, here are two notebooks for using Vilio to perform pure inference on any meme you want :)
• Just adapt the example input dataset / input model to use a different meme / pretrained model🥶
• GPU: https://www.kaggle.com/muennighoff/vilioexample-nb
• CPU: https://www.kaggle.com/muennighoff/vilioexample-nb-cpu

Ordering

Vilio aims to replicate the organization of huggingface's transformer repo at: https://github.com/huggingface/transformers

• /bash Shell files to reproduce hateful memes results

• /data By default, directory for loading in data & saving checkpoints

• /ernie-vil Ernie-vil sub-repository written in PaddlePaddle

• /fts_lmdb Scripts for handling .lmdb extracted features

• /fts_tsv Scripts for handling .tsv extracted features

• /notebooks Jupyter Notebooks for demonstration & reproducibility

• /py-bottm-up-attention Sub-repository for tsv feature extraction forked & adapted from here

• src/vilio All implemented models (also see below for a quick overview of models)

• /utils Pandas & ensembling scripts for data handling

• entry.py files Scripts used to access the models and apply model-specific data preparation

• pretrain.py files Same purpose as entry files, but for pre-training; Point of entry for pre-training

• hm.py Training code for the hateful memes challenge; Main point of entry

• param.py Args for running hm.py

Usage

Follow SCORE_REPRO.md for reproducing performance on the Hateful Memes Task.
Follow GETTING_STARTED.md for using the framework for your own task.
See the paper at: https://arxiv.org/abs/2012.07788

Architectures

🥶 Vilio currently provides the following architectures with the outlined language transformers:

1. E - ERNIE-VIL ERNIE-ViL: Knowledge Enhanced Vision-Language Representations Through Scene Graph
   • ERNIE: Enhanced Language Representation with Informative Entities
2. D - DeVLBERT DeVLBert: Learning Deconfounded Visio-Linguistic Representations
   • BERT: Bidirectional Transformers
3. O - OSCAR Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks
   • BERT: Bidirectional Transformers
4. U - UNITER UNITER: UNiversal Image-TExt Representation Learning
   • BERT: Bidirectional Transformers
   • RoBERTa: Robustly Optimized BERT Pretraining Approach
5. V - VisualBERT VisualBERT: A Simple and Performant Baseline for Vision and Language
   • ALBERT: A Lite BERT
   • BERT: Bidirectional Transformers
   • RoBERTa: Robustly Optimized BERT Pretraining Approach
6. X - LXMERT LXMERT: Learning Cross-Modality Encoder Representations from Transformers
   • ALBERT: A Lite BERT
   • BERT: Bidirectional Transformers
   • RoBERTa: Robustly Optimized BERT Pretraining Approach

To-do's

• Clean-up import statements, python paths & find a better way to integrate transformers (Right now all import statements only work if in main folder)
• Enable loading and running models just via import statements (and not having to clone the repo)
• Find a way to better include ERNIE-VIL in this repo (PaddlePaddle to Torch?)
• Move tokenization in entry files to model-specific tokenization similar to transformers

Attributions

The code heavily borrows from the following repositories, thanks for their great work:

• https://github.com/huggingface/transformers
• https://github.com/facebookresearch/mmf
• https://github.com/airsplay/lxmert

Citation

@article{muennighoff2020vilio,
  title={Vilio: State-of-the-art visio-linguistic models applied to hateful memes},
  author={Muennighoff, Niklas},
  journal={arXiv preprint arXiv:2012.07788},
  year={2020}
}