Skip to content

Latest commit

 

History

History
209 lines (162 loc) · 10.1 KB

README.md

File metadata and controls

209 lines (162 loc) · 10.1 KB
Raw

ViCHA: Efficient Vision-Language Pretraining with Visual Concepts and Hierarchical Alignment.

This is the official PyTorch implementation of the ViCHA paper

Vision and Language Pretraining has become the prevalent approach for tackling multimodal downstream tasks. The current trend is to move towards ever larger models and pretraining datasets. This computational headlong rush does not seem reasonable in the long term to move toward sustainable solutions, and de facto excludes academic laboratories with limited resources. In this work, we propose a new framework, dubbed ViCHA, that efficiently exploits the input data to boost the learning by: (a) a new hierarchical cross-modal alignment loss, (b) new self-supervised scheme based on masked image modeling, (c) leveraging image-level annotations, called Visual Concepts, obtained with existing foundation models such as CLIP to boost the performance of the image encoder. Although pretrained on four times less data, our ViCHA strategy outperforms other approaches on several downstream tasks such as Image-Text Retrieval, VQA, Visual Reasoning, Visual Entailment and Visual Grounding.

Requirements:

  • pytorch 1.8.2
  • transformers 4.8.1
  • timm 0.4.9
  • clip

Installation:

Create conda env from envs/env.yaml:

conda env create -n vicha --file envs/env.yaml
conda activate vicha

Then install clip:

pip install git+https://github.com/openai/CLIP.git

Download:

Model weights

  • Pre-trained checkpoint [1M] / [4M]
  • Finetuned checkpoint for retrieval on MSCOCO 1M / [4M]
  • Finetuned checkpoint for retrieval on Flickr30k 1M / [4M]
  • Finetuned checkpoint for VQA 1M / [4M]
  • Finetuned checkpoint for NLVR2 1M / [4M]
  • Finetuned checkpoint for visual entailment on SNLI-VE 1M / [4M]
  • Finetuned checkpoint for visual grounding on RefCOCO+ 1M

Dataset

You need to donwnload the dataset from their original websites. Our data directory is structured as follows:

path_to_data_dir/

------coco:
            karpathy   v2_mscoco_train2014_annotations.json         v2_OpenEnded_mscoco_test-dev2015_questions.json  val2014
            test2015   v2_mscoco_val2014_annotations.json           v2_OpenEnded_mscoco_train2014_questions.json
            train2014  v2_OpenEnded_mscoco_test2015_questions.json  v2_OpenEnded_mscoco_val2014_questions.json

------sbu:
            annot.json  dataset  images_train

------visual_genome:
            annotations
            images: 
                  VG_100K  VG_100K_2

------cc3m:
            cc3m.json  train  Train_GCC-training.tsv  val  Validation_GCC-1.1.0-Validation.tsv

------flicker30k:
            flickr30k-images  karpathy

------snli:
            data:
                  features  flickr30k_images  images  results_20130124.token  results.tsv  snli_1.0

------nlvr:
            dev  images  nlvr  nlvr2  README.md  test1

Create augmented datasets (with VCs):

We provide the necessary scripts to extract, embedd and find the best visual concepts for each image (and you can donwnload them as described later):

  1. Extract the textual VCs from all the captions of the datasets: extract_keywords.py
  2. Embedd all textual VCs using CLIP text encoder: save_kwords_embeddings.py
  3. Embedd all images using CLIP visual encoder: save_image_embeddings.py
  4. Create the augmented dataset using: create_clip_da_captions_from_saved.py (if you don't have enough memory use create_clip_da_captions.py)

Create CLIP-Filtered dataset (800K images)

ViCHA (800K) is trained on; 100% of COCO, 50% of VG captions for each image (we keep all images), and 70% of SBU:

  1. To filter the dataset while keeping the same number of images (useful in case of Visual Genome dataset): filter_dataset_per_image.py
  2. To filter a given dataset and keeping a percentage of pairs: filter_dataset_percentage.py

Download

You can either create the augmented dataset as described above, or download them as follows;

Pretraining

Launch Pretraining on 1M images using 4 GPUs:

python -m torch.distributed.launch --nproc_per_node=4 --use_env Pretrain.py \
--config ./configs/Pretrain.yaml --output_dir path_to_log_dir/pretrain \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/json_pretrain
  • path_to_log_dir: the directory where you save the logs and model weights
  • path_to_data_dir: the path to the pretraining dataset (VG, COCO, SBU, CC3M)
  • path_to_json_dir: the path to datasets json (caption, path to image, VCs...)

you can resume the training or start from a checkpoint using --resume and --checkpoint arguments.

Finetuning

Retreival

  1. Download MSCOCO or Flickr30k datasets from the original websites.
  2. Download and extract the provided dataset json files.
  3. Finetune the pre-trained checkpoint using 4 A100 GPUs:

For Flickr30K:

python -m torch.distributed.launch --nproc_per_node=4 --use_env Retrieval.py \
--config ./configs/Retrieval_flickr.yaml --output_dir path_to_log_dir/retrieval_flickr \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/data \
--checkpoint path_to_log_dir/pretrain/checkpoint_last.pth

For COCO:

python -m torch.distributed.launch --nproc_per_node=4 --use_env Retrieval.py \
--config ./configs/Retrieval_coco.yaml --output_dir path_to_log_dir/retrieval_coco \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/data \
--checkpoint path_to_log_dir/pretrain/checkpoint_last.pth
  1. Evaluate by adding the --evaluate arg and using the finetuned checkpoint.

VQA

  1. Download VQA v2 dataset and Visual Genome dataset from the original websites.
  2. Download and extract the provided dataset json files.
  3. Finetune the pre-trained checkpoint using 2 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=2 --use_env VQA.py \
--config ./configs/VQA.yaml --output_dir path_to_log_dir/vqa \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/data \
--checkpoint path_to_log_dir/pretrain/checkpoint_last.pth
  1. Evaluate on the VQA v2 server

SNLI-VE

  1. Download SNLI-VE dataset from the original website.
  2. Download and extract the provided dataset json files.
  3. Finetune the pre-trained checkpoint using 2 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=2 --use_env VE.py \
--config ./configs/VE.yaml --output_dir path_to_log_dir/ve \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/data \
--checkpoint path_to_log_dir/pretrain/checkpoint_last.pth
  1. Evaluate by adding the --evaluate arg and using the finetuned checkpoint.

NLVR

  1. Download SNLI-VE dataset from the original website.
  2. Download and extract the provided dataset json files.
  3. Pretrain with text-assignment (TA):
python -m torch.distributed.launch --nproc_per_node=2 --use_env Pretrain_nlvr.py \
--config ./configs/NLVR_pretrain.yaml --output_dir path_to_log_dir/nlvr_pretrain \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/json_pretrain \
--checkpoint path_to_log_dir/pretrain/checkpoint_last.pth
  1. Finetune the pre-trained checkpoint using 2 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=2 --use_env NLVR.py \
--config ./configs/NLVR.yaml --output_dir path_to_log_dir/nlvr\
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/data \
--checkpoint path_to_log_dir/nlvr_pretrain/checkpoint_last.pth
  1. Evaluate by adding the --evaluate arg and using the finetuned checkpoint.

Grounding

  1. Download MSCOCO dataset from the original website.
  2. Download and extract the provided dataset json files.
  3. Finetune the pre-trained checkpoint using 2 A100 GPUs:
python -m torch.distributed.launch --nproc_per_node=2 --use_env Grounding.py \
--config ./configs/Grounding.yaml --output_dir path_to_log_dir/grounding \
--data_dir path_to_data_dir --data_json_dir path_to_json_dir/data \
--checkpoint path_to_log_dir/pretrain/checkpoint_last.pth
  1. Evaluate by adding the --evaluate arg and using the finetuned checkpoint.

Visualization:

You can visualize the Grounding results using the visualization.ipynb notebook.

Citation.

@inproceedings{shukor2022efficient,
  title={Efficient Vision-Language Pretraining with Visual Concepts and Hierarchical Alignment},
  author={Shukor, Mustafa and Couairon, Guillaume and Cord, Matthieu},
  booktitle={33rd British Machine Vision Conference (BMVC)},
  year={2022}
}

Aknowledgment:

This code is based on the ALBEF repo and borrows some code from CLIP, timm.