ViLBERT

Code and pre-trained models for ViLBERT: Pretraining Task-Agnostic VisiolinguisticRepresentations for Vision-and-Language Tasks.

*Note: This codebase is still in beta release to replicate the paper's preformance. *

Repository Setup

1. Create a fresh conda environment, and install all dependencies.

conda create -n vilbert python=3.6
conda activate vilbert
git clone https://github.com/jiasenlu/vilbert_beta
cd vilbert_beta
pip install -r requirements.txt

2. Install pytorch

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

3. Install apx, follows https://github.com/NVIDIA/apex

4. compile tools

cd tools/refer
make

Data Setup

Check README.md under data for more details. Check vlbert_tasks.yml for more details.

Pre-trained model for Evaluation

Model	Objective	Link
ViLBERT 2-Layer	Conceptual Caption	Google Drive
ViLBERT 4-Layer	Conceptual Caption	Google Drive
ViLBERT 6-Layer	Conceptual Caption	Google Drive
ViLBERT 8-Layer	Conceptual Caption	Google Drive
ViLBERT 6-Layer	VQA	Google Drive
ViLBERT 6-Layer	VCR	Google Drive
ViLBERT 6-Layer	RefCOCO+	Google Drive
ViLBERT 6-Layer	Image Retrieval	Google Drive

Evaluation

Zero-Shot Image Retrieval

We can directly use the Pre-trained ViLBERT model for zero-shot image retrieval tasks on Flickr30k.

1: Download the pretrained model with objective Conceptual Caption and put it under save

2: Update featyres_h5path1 and val_annotations_jsonpath in vlbert_task.yml to load the Flickr30k testset image feature and jsonfile (defualt is training feature).

3: Use the following command to evaluate pre-trained 6 layer ViLBERT model. (only support single GPU for evaluation now):

python eval_retrieval.py --bert_model bert-base-uncased --from_pretrained save/bert_base_6_layer_6_connect/pytorch_model_9.bin --config_file config/bert_base_6layer_6conect.json --task 3 --split test --batch_size 1 --zero_shot

Image Retrieval

1: Download the pretrained model with objective Image Retrieval and put it under save

2: Update featyres_h5path1 and val_annotations_jsonpath in vlbert_task.yml to load the Flickr30k testset image feature and jsonfile (defualt is training feature).

3: Use the following command to evaluate pre-trained 6 layer ViLBERT model. (only support single GPU for evaluation now):

python eval_retrieval.py --bert_model bert-base-uncased --from_pretrained save/RetrievalFlickr30k_bert_base_6layer_6conect-pretrained/pytorch_model_19.bin --config_file config/bert_base_6layer_6conect.json --task 3 --split test --batch_size 1

VQA

1: Download the pretrained model with objective VQA and put it under save

2: To test on held out validation split, use the following command:

python eval_tasks.py --bert_model bert-base-uncased --from_pretrained save/VQA_bert_base_6layer_6conect-pretrained/pytorch_model_19.bin --config_file config/bert_base_6layer_6conect.json --task 0 --split minval

VCR

1: Download the pretrained model with objective VCR and put it under save

2: To test on VCR Q->A

python eval_tasks.py --bert_model bert-base-uncased --from_pretrained save/VCR_Q-A-VCR_QA-R_bert_base_6layer_6conect-pretrained/pytorch_model_19.bin --config_file config/bert_base_6layer_6conect.json --task 1 --split val

3: To test on VCR QA->R

python eval_tasks.py --bert_model bert-base-uncased --from_pretrained save/VCR_Q-A-VCR_QA-R_bert_base_6layer_6conect-pretrained/pytorch_model_19.bin --config_file config/bert_base_6layer_6conect.json --task 2 --split val

RefCOCO+

1: Download the pretrained model with objective RefCOCO+ and put it under save

2: We use the Pre-computed detections/masks from MAttNet for fully-automatic comprehension task, Check the MAttNet repository for more details.

3: To test on the RefCOCO+ val set and use the following command:

python eval_tasks.py --bert_model bert-base-uncased --from_pretrained save/refcoco+_bert_base_6layer_6conect-pretrained/pytorch_model_19.bin --config_file config/bert_base_6layer_6conect.json --task 4

Visiolinguistic Pre-training

Once you extracted all the image features, to train a 6-layer ViLBERT model on conceptual caption:

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 train_concap.py --from_pretrained bert-base-uncased --bert_model bert-base-uncased --conf
ig_file config/bert_base_6layer_6conect.json --learning_rate 1e-4 --train_batch_size 512 --save_name pretrained

Train ViLBERT for DownStream Tasks

VQA

To fintune a 6-layer ViLBERT model for VQA with 8 GPU. --tasks 0 means VQA tasks. Check vlbert_tasks.yml for more settings for VQA tasks.

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 4e-5 --num_workers 16 --tasks 0 --save_name pretrained

VCR

Similarly, to finetune a 6-layer vilbert model for VCR task, run the following commands. Here we joint train Q->A and QA->R tasks, so the tasks is specified as --tasks 1-2

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

Image Retrieval

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 4e-5 --num_workers 9 --tasks 3 --save_name pretrained

Refer Expression

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 4e-5 --num_workers 16 --tasks 4 --save_name pretrained

• For single GPU training, use smaller batch size and simply remove -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0

References

If you find this code is useful for your research, please cite our paper

@article{lu2019vilbert,
  title={ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks},
  author={Lu, Jiasen and Batra, Dhruv and Parikh, Devi and Lee, Stefan},
  journal={arXiv preprint arXiv:1908.02265},
  year={2019}
}

Why does ViLBERT look like ?