Less is More: ClipBERT for Video-and-Language Learning via Sparse Sampling
CVPR 2021, Oral, Best Student Paper Honorable Mention.
Jie Lei*, Linjie Li*, Luowei Zhou, Zhe Gan, Tamara L. Berg, Mohit Bansal, Jingjing Liu
Official PyTorch code for ClipBERT, an efficient framework for end-to-end learning for image-text and video-text tasks. It takes raw videos/images + text as inputs, and outputs task predictions. ClipBERT is designed based on 2D CNNs and transformers, and uses a sparse sampling strategy to enable efficient end-to-end video-and-language learning. In this repository, we support end-to-end pretraining and finetuning for the following tasks:
It is also feasible and easy to add other image-text or video-text tasks for pretraining and finetuning.
[NEW] If you are interested in ClipBERT, you might also be interested in our recent work, Singularity (paper, code), with a single-frame trained model, it achieves state-of-the-art results a set of video-language tasks.
We provide a Docker image for easier reproduction. Please install the following:
Our scripts require the user to have the docker group membership so that docker commands can be run without sudo. We only support Linux with NVIDIA GPUs. We test on Ubuntu 18.04 and V100 cards. We use mixed-precision training hence GPUs with Tensor Cores are recommended.
Create a folder that stores pretrained models, all the data, and results.
PATH_TO_STORAGE=/path/to/your/data/
mkdir -p $PATH_TO_STORAGE/txt_db # annotations
mkdir -p $PATH_TO_STORAGE/vis_db # image and video
mkdir -p $PATH_TO_STORAGE/finetune # finetuning results
mkdir -p $PATH_TO_STORAGE/pretrained # pretrained modelsDownload pretrained models.
Our e2e pretrained ClipBERT model (849MB), can be downloaded with the following command.
bash scripts/download_pretrained.sh $PATH_TO_STORAGEThis pretrained model can be used for finetuning on video-text tasks and image-text tasks.
For your convenience, this script will also download bert-base-uncased and grid-feat-vqa
model weights, which are used as initialization for pretraining.
Launch the Docker container for running the experiments.
# docker image should be automatically pulled
source launch_container.sh $PATH_TO_STORAGE/txt_db $PATH_TO_STORAGE/vis_db \
$PATH_TO_STORAGE/finetune $PATH_TO_STORAGE/pretrainedThe launch script respects $CUDA_VISIBLE_DEVICES environment variable.
Note that the source code is mounted into the container under /clipbert instead
of built into the image so that user modification will be reflected without
re-building the image. (Data folders are mounted into the container separately
for flexibility on folder structures.)
Tasks: MSRVTT retrieval, DiDeMo and ActivityNet Captions paragprah-to-video retrieval, MSRVTT MC Test.
Download data.
# outside the container
# download videos + annotations for $DSET
bash scripts/download_$DSET.sh $PATH_TO_STORAGE$DSET can be one of msrvtt, didemo, anet.
Finetuning.
# inside the container
horovodrun -np 4 python src/tasks/run_video_retrieval.py \
--config $CONFIG_PATH \
--output_dir $OUTPUT_DIR
# for single GPU
python src/tasks/run_video_retrieval.py \
--config $CONFIG_PATH \
--output_dir $OUTPUT_DIR$CONFIG_PATH should be set to one of the .json config files available at src/configs
prefixed with _ret. For example, you can use src/configs/msrvtt_ret_base_resnet50.json
for MSRVTT retrieval.
Run inference.
# inside the container
horovodrun -np 4 python src/tasks/run_video_retrieval.py \
--do_inference 1 --output_dir $OUTPUT_DIR \
--inference_split val --inference_model_step $STEP \
--inference_txt_db $TXT_DB \
--inference_img_db $IMG_DB --inference_batch_size 64 \
--inference_n_clips $INFERENCE_N_CLIPS$STEP is an integer, it tells the script to use the checkpoint
$OUTPUT_DIR/ckpt/model_step_$STEP.pt for inference.
$TXT_DB and $IMG_DB are path to annotation file and video data. You can use
TXT_DB=/txt/downstream/msrvtt_retrieval/msrvtt_retrieval_val.jsonl and
IMG_DB=/img/msrvtt for inference on MSRVTT retrieval val split.
The results will be written under $OUTPUT_DIR. You can use different $INFERENCE_N_CLIPS
for inference, such as 1 or 16. Using more clips will have a large impact
on inference speed and memory usage. You may want to use smaller batch sizes if larger
values are set.
After MSRVTT retrieval model is trained, you can use it for inference for the MSRVTT MC Test task as well, which is essentially a retrieval task in a multiple-choice task setup.
# inside the container
horovodrun -np 4 python src/tasks/run_msrvtt_mc.py \
--do_inference 1 --output_dir $OUTPUT_DIR \
--inference_split val --inference_model_step $STEP \
--inference_txt_db /txt/downstream/msrvtt_retrieval_mc/msrvtt_retrieval_mc_test.jsonl \
--inference_img_db /img/msrvtt --inference_batch_size 64 \
--inference_n_clips $INFERENCE_N_CLIPSTasks: TGIF-QA action, transition, and frameQA tasks; MSRVTT-QA.
Download data.
# outside the container
# download MSRVTT videos, and QA + retrieval annotations
bash scripts/download_msrvtt.sh $PATH_TO_STORAGE
# download TGIF-QA videos and annotations
bash scripts/download_tgif_qa.sh $PATH_TO_STORAGE Finetuning.
# inside the container
horovodrun -np 4 python src/tasks/run_video_qa.py \
--config $CONFIG_PATH \
--output_dir $OUTPUT_DIR$CONFIG_PATH should be set to one of the .json config files available at src/configs
contains the substring _qa. For example, you can use src/configs/msrvtt_qa_base_resnet50.json
for MSRVTT-QA.
Run inference.
# inside the container
horovodrun -np 4 python src/tasks/run_video_qa.py \
--do_inference 1 --output_dir $OUTPUT_DIR \
--inference_split val --inference_model_step $STEP \
--inference_txt_db $TXT_DB \
--inference_img_db $IMG_DB --inference_batch_size 64 \
--inference_n_clips $INFERENCE_N_CLIPS$STEP is an integer, which tells the script to use the checkpoint
$OUTPUT_DIR/ckpt/model_step_$STEP.pt for inference.
$TXT_DB and $IMG_DB are path to annotation file and video data. You can use
TXT_DB=/txt/downstream/msrvtt_retrieval/msrvtt_qa_val.jsonl and
IMG_DB=/img/msrvtt for inference on MSRVTT QA val split.
The results will be written under $OUTPUT_DIR. You can use different $INFERENCE_N_CLIPS
for inference, such as 1 or 16. Using more clips will have a large impact
on inference speed and memory usage. You may want to use smaller batch sizes if larger
values are set.
Download data
# outside the container
# download COCO and VG data
bash scripts/download_coco_vg.sh $PATH_TO_STORAGE
# download VQA annotations
bash scripts/download_vqa.sh $PATH_TO_STORAGEFinetuning
# inside the container
horovodrun -np 4 python src/tasks/run_vqa.py \
--config src/configs/vqa_base_resnet50.json \
--output_dir $OUTPUT_DIRInference
# inside the container
horovodrun -np 4 python src/tasks/run_vqa.py \
--do_inference 1 --output_dir $OUTPUT_DIR \
--inference_split val --inference_model_step $STEP \
--inference_txt_db $TXT_DB \
--inference_img_db $IMG_DB \
--inference_batch_size 64Download data
# outside the container
bash scripts/download_coco_vg.sh $PATH_TO_STORAGEPretraining
#inside the container
horovodrun -np 8 python src/pretrain/run_pretrain.py \
--config src/configs/pretrain_image_text_base_resnet50_mlm_itm.json \
--output_dir $OUTPUT_DIR ClipBERT takes raw video and text as inputs, there is no need to do feature extraction.
However, to improve data loading speed, we use LMDB to store the raw image and video files.
You can use the following script to convert a list of videos with file extensions mp4 and avi into LMDB:
# outside the container
python src/preprocessing/file2lmdb.py \
--data_root /path/to/videos \
--lmdb_save_dir /path/to/save/lmdb \
--ext avi mp4 \
--file_type video For images, use appropriate file extensions for --ext and --file_type image.
Text annotation files are reorganized into jsonl files,
see example preprocessed files downloaded by the scripts in scripts/.
If you find this code useful for your research, please consider citing:
@inproceedings{lei2021less,
title={Less is More: ClipBERT for Video-and-Language Learningvia Sparse Sampling},
author={Lei, Jie and Li, Linjie and Zhou, Luowei and Gan, Zhe and Berg, Tamara L. and Bansal, Mohit and Liu, Jingjing},
booktitle={CVPR},
year={2021}
}We thank Yen-Chun Chen, Ruotian Luo, and other members and interns at Microsoft Multimodal AI for their helpful discussions. We also thank the anonymous reviewers for their constructive feedback.
This code used resources from transformers, UNITER, HERO, grid-feats-vqa, SlowFast, Detectron2. The code is implemented using PyTorch, with multi-GPU support from Horovod and mixed precision support from apex. We thank the authors for open-sourcing their awesome projects.
MIT