GD-VCR

Code for Broaden the Vision: Geo-Diverse Visual Commonsense Reasoning (EMNLP 2021).

Research Questions and Aims:

How well can a model perform on the images which requires geo-diverse commonsense to understand?
What are the reasons behind performance disparity on Western and non-Western images?
We aim to broaden researchers' vision on a realistic issue existing all over the world, and call upon researchers to consider more inclusive commonsense knowledge and better model transferability on various cultures.

In this repo, GD-VCR dataset and codes about 1) general model evaluation, 2) detailed controlled experiments, and 3) dataset construction are provided.

Repo Structure

GD-VCR
 ├─X_VCR				  --> storing GD-VCR/VCR data
 ├─configs
 │  └─vcr
 │     └─fine-tune-qa.json		  --> part of configs for evaluation
 ├─dataloaders
 │  └─vcr.py			          --> load GD-VCR/VCR data based on configs
 ├─models
 │  └─train.py		                  --> fine-tune/evaluate models
 │
 ├─val.jsonl			          --> GD-VCR dataset
 ├─val_addition_single.jsonl		  --> additional low-order QA pairs

GD-VCR dataset

First download the original VCR dataset to X_VCR:

cd X_VCR
wget https://s3.us-west-2.amazonaws.com/ai2-rowanz/vcr1annots.zip
wget https://s3.us-west-2.amazonaws.com/ai2-rowanz/vcr1images.zip
unzip vcr1annots.zip
unzip vcr1images.zip

Then download the GD-VCR dataset to X_VCR:

cd X_VCR
mv val.jsonl orig_val.jsonl
wget https://gdvcr.s3.us-west-1.amazonaws.com/MC-VCR_sample.zip
unzip MC-VCR_sample.zip

cd ..
mv val.jsonl X_VCR/
mv val_addition_single.jsonl X_VCR/

The detailed items in our GD-VCR dataset are almost the same as VCR. Please refer to VCR website for detailed explanations.

VisualBERT

Prepare Environment

Prepare environment as mentioned in the original repo of VisualBERT.

Fine-tune model on original VCR

Download the task-specific pre-trained checkpoint on original VCR vcr_pre_train.th to GD-VCR/visualbert/trained_models.

Then, use the command to fine-tune:

export PYTHONPATH=$PYTHONPATH:GD-VCR/visualbert/
export PYTHONPATH=$PYTHONPATH:GD-VCR/

cd GD-VCR/visualbert/models

CUDA_VISIBLE_DEVICES=0 python train.py -folder ../trained_models -config ../configs/vcr/fine-tune-qa.json

For convenience, we provide a trained checkpoint [Link] for quick evaluation.

Evaluation on GD-VCR

CUDA_VISIBLE_DEVICES=0 python train.py -folder ../trained_models -config ../configs/vcr/eval.json \
        [-region REGION] \
        [-scene SCENE] \
        [-single_or_multiple SINGLE_OR_MULTIPLE] \
        [-orig_or_new ORIG_OR_NEW] \
	[-addition_annotation_analysis] \
        [-grounding]

Here are the explanations of several important attributions:

REGION: One of the regions west, east-asia, south-asia, africa.
SCENE: One of the scenario (e.g., wedding).
SINGLE_OR_MULTIPLE: Whether studying single(low-order) or multiple(high-order) cognitive questions.
addition_annotation_analysis: Whether studying GD-VCR or additional annotated questions. If yes, you can choose to set SINGLE_OR_MULTIPLE to specify which types of questions you want to investigate.
ORIG_OR_NEW: Whether studying GD-VCR or original VCR dev set.
grounding: Whether analyzing grounding results by visualizing attention weights.

Given our fine-tuned VisualBERT model above, the evaluation results are shown below:

Models	Overall	West	South Asia	East Asia	Africa
VisualBERT	53.27	62.91	52.04	45.39	51.85

ViLBERT

Prepare Environment

Prepare environment as mentioned in the original repo of ViLBERT.

Extract image features

We make use of the docker made for LXMERT. Detailed commands are shown below:

cd GD-VCR
git clone https://github.com/jiasenlu/bottom-up-attention.git
mv generate_tsv.py bottom-up-attention/tools
mv generate_tsv_gt.py bottom-up-attention/tools

docker pull airsplay/bottom-up-attention
docker run --name gd_vcr --runtime=nvidia -it -v /PATH/TO/:/PATH/TO/ airsplay/bottom-up-attention /bin/bash
[Used to enter into the docker]

cd /PATH/TO/GD-VCR/bottom-up-attention
pip install json_lines
pip install jsonlines
pip install python-dateutil==2.5.0

python ./tools/generate_tsv.py --cfg experiments/cfgs/faster_rcnn_end2end_resnet.yml --def models/vg/ResNet-101/faster_rcnn_end2end_final/test.prototxt --out ../vilbert_beta/feature/VCR/VCR_resnet101_faster_rcnn_genome.tsv --net data/faster_rcnn_models/resnet101_faster_rcnn_final.caffemodel --total_group 1 --group_id 0 --split VCR
python ./tools/generate_tsv_gt.py --cfg experiments/cfgs/faster_rcnn_end2end_resnet.yml --def models/vg/ResNet-101/faster_rcnn_end2end_final/test_gt.prototxt --out ../vilbert_beta/feature/VCR/VCR_gt_resnet101_faster_rcnn_genome.tsv --net data/faster_rcnn_models/resnet101_faster_rcnn_final.caffemodel --total_group 1 --group_id 0 --split VCR_gt
[Used to extract features]

Then, exit the dockerfile, and convert extracted features into lmdb form:

cd GD-VCR/vilbert_beta
python script/convert_lmdb_VCR.py
python script/convert_lmdb_VCR_gt.py

Fine-tune model on original VCR

Download the pre-trained checkpoint to GD-VCR/vilbert_beta/save/bert_base_6_layer_6_connect_freeze_0/.

Then, use the command to fine-tune:

cd GD-VCR/vilbert_beta
python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 train_tasks.py --bert_model bert-base-uncased --from_pretrained save/bert_base_6_layer_6_connect_freeze_0/pytorch_model_8.bin  --config_file config/bert_base_6layer_6conect.json  --learning_rate 2e-5 --num_workers 16 --tasks 1-2 --save_name pretrained

For convenience, we provide a trained checkpoint [Link] for quick evaluation.

Evaluation on GD-VCR

CUDA_VISIBLE_DEVICES=0,1 python eval_tasks.py 
		--bert_model bert-base-uncased 
		--from_pretrained save/VCR_Q-A-VCR_QA-R_bert_base_6layer_6conect-pretrained/vilbert_best.bin 
		--config_file config/bert_base_6layer_6conect.json --task 1 --split val  --batch_size 16

Note that if you want the results on original VCR dev set, you could directly change the "val_annotations_jsonpath" value of TASK1 to X_VCR/orig_val.jsonl.

Given our fine-tuned ViLBERT model above, the evaluation results are shown below:

Models	Overall	West	South Asia	East Asia	Africa
ViLBERT	58.47	65.82	62.90	46.45	62.04

Dataset Construction

Here we provide dataset construction methods in our paper:

similarity.py: Compute the similarity among answer candidates and distribute candidates to each annotated questions.
relevance_model.py: Train a model to compute the relevance between question and answer.
question_cluster.py: Infer question templates from original VCR dataset as the basis of annotation.

For sake of convenience, we provide the trained relevance computation model [Link].

Acknowledgement

We thank for VisualBERT, ViLBERT, and Detectron authors' implementation. Also, we appreciate the effort of original VCR paper's author, and our work is highly influenced by VCR.

Citation

Please cite our EMNLP paper if this repository inspired your work.

@inproceedings{yin2021broaden,
  title = {Broaden the Vision: Geo-Diverse Visual Commonsense Reasoning},
  author = {Yin, Da and Li, Liunian Harold and Hu, Ziniu and Peng, Nanyun and Chang, Kai-Wei},
  booktitle = {EMNLP},
  year = {2021}
}

Code and data for "Broaden the Vision: Geo-Diverse Visual Commonsense Reasoning" (EMNLP 2021).

Related tags

Overview

GD-VCR

Research Questions and Aims:

Repo Structure

GD-VCR dataset

VisualBERT

Prepare Environment

Fine-tune model on original VCR

Evaluation on GD-VCR

ViLBERT

Prepare Environment

Extract image features

Fine-tune model on original VCR

Evaluation on GD-VCR

Dataset Construction

Acknowledgement

Citation

Owner

Da Yin

Tensorflow AffordanceNet and AffContext implementations

【Arxiv】Exploring Separable Attention for Multi-Contrast MR Image Super-Resolution

Replication Package for "An Empirical Study of the Effectiveness of an Ensemble of Stand-alone Sentiment Detection Tools for Software Engineering Datasets"

Christmas face app for Decathlon xmas coding party!

A geometric deep learning pipeline for predicting protein interface contacts.

Learning RAW-to-sRGB Mappings with Inaccurately Aligned Supervision (ICCV 2021)

A PaddlePaddle implementation of STGCN with a few modifications in the model architecture in order to forecast traffic jam.

This is an official implementation of CvT: Introducing Convolutions to Vision Transformers.

Time-Optimal Planning for Quadrotor Waypoint Flight

Using modified BiSeNet for face parsing in PyTorch

FairEdit: Preserving Fairness in Graph Neural Networks through Greedy Graph Editing

Learning Neural Painters Fast! using PyTorch and Fast.ai

Official implementation of the PICASO: Permutation-Invariant Cascaded Attentional Set Operator

[ICCV 2021] Official Tensorflow Implementation for "Single Image Defocus Deblurring Using Kernel-Sharing Parallel Atrous Convolutions"

Mercury: easily convert Python notebook to web app and share with others

Prososdy Morph: A python library for manipulating pitch and duration in an algorithmic way, for resynthesizing speech.

Official repository for ABC-GAN

[NeurIPS-2020] Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID.

Human head pose estimation using Keras over TensorFlow.

GB-CosFace: Rethinking Softmax-based Face Recognition from the Perspective of Open Set Classification