REFACTOR IN PROCESS

No I'm serious. Don't touch this.

VisualRoBERTa

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)

Introduction

The first public Vietnamese visual linguistic foundation model(s). This work was carried out only by myself under supervision of Dr Pham Quang Nhat Minh @ Aimesoft and Dr Tran Giang Son @ USTH. Thanks to Mr Nguyen Anh Duong @ VietAI for TPU supports.

Keywords: computer vision, natural language processing, visual linguistic, image text, pretrain, Vietnamese, foundation, multi-modal, machine learning

Results

On UIT-ViIC test set

Baseline models are the best models in UIT-ViIC paper.

On VQA test set

Baseline model is the best model in IC paper.

Citation

To cite this repos or the models' weights or the theory,

@software{dinhanhx_VisualRoBERTa_2022,
	title        = {{VisualRoBERTa}},
	author       = {dinhanhx},
	year         = 2022,
	month        = 9,
	url          = {https://github.com/dinhanhx/VisualRoBERTa}
}

⚠ This entry will be updated when the white paper is published or released to the public.

Setup Dependencies

• For TPU, you just can pip install requirements.txt
• For GPU, besides reading requirements.txt, you gotta remove any command related to TPU, XLA, then follow original PyTorch docs.

Download Dataset

In training (run) files (such as run_ptrain.py), paths to data folders are hardcoded

⚠ TranslateCOCO2017 also contains json files from UIT-ViIC.

Download links:

• MS COCO
• Translate COCO 2017 this work
• ViVQA
• UIT-ViIC

You are encouraged to read src/data.py to understand dataset structure and renamed paths to something suitable for your systems.

Train models

It's quite simple, just simple go with

python -m exp.run_<task_name_go_here>.py

for example, python run_pretrain.py will pretrain the model.

You are encouraged to read these files to understand what they do before training.

• For TPU, just run it like normal
• For GPU, you gotta remove/modify anything related to TPU such as xla, tpu, xm, xla_spawn_debug, DistributedSampler...

⚠ Hardcoded file paths might be updated.

Kill leftover processes

pgrep -f "python -m exp.run_pretrain" | xargs kill -9

Evaluate models

It's also simple, just simple go with

python -m exp.eval_<dataset_go_here>.py

for example, python eval_vqa.py will infer the models to produce the answers, NOT to compute metrics.

You are encouraged to read these files to understand what they do before evaluation.

⚠ Hardcoded file paths might be updated.