Multi-Modal Self-Supervision using GDT and StiCa

This is an official pytorch implementation of papers: Multi-modal Self-Supervision from Generalized Data Transformations and Space-Time Crop & Attend: Improving Cross-modal Video Representation Learning. In this repository, we provide PyTorch code for pretraining and testing our proposed GDT and StiCa models.

If you find GDT and STiCA useful in your research, please use the following BibTeX entries for citation.

@misc{patrick2020multimodal,
      title={Multi-modal Self-Supervision from Generalized Data Transformations}, 
      author={Mandela Patrick and Yuki M. Asano and Polina Kuznetsova and Ruth Fong and João F. Henriques and Geoffrey Zweig and Andrea Vedaldi},
      year={2021},
      booktitle={International Conference on Computer Vision (ICCV)},
}

@misc{m2021spacetime,
    title={Space-Time Crop & Attend: Improving Cross-modal Video Representation Learning},
    author={Mandela Patrick and Yuki M. Asano and Bernie Huang and Ishan Misra and Florian Metze and Joao Henriques and Andrea Vedaldi},
    year={2021},
    booktitle={International Conference on Computer Vision (ICCV)},
}

Highlights

(1) GDT: Formulate and generalize most pretext tasks in a NCE objective.

Using this formulation, we test various pretext tasks previously unexplored and achieve SOTA downstream performance.

(2) STiCA: Importance of incorporating within-modal invariance in cross-modal learning

We show how to efficiently incorporate within-modal invariance learning using feature crops and achieve SOTA downstream performance.

Model Zoo

We provide GDT models pretrained on Kinetics-400 (K400), HowTo100M (HT100M), and Instagram-65M (IG65M) datasets, and StiCa models pretrained on Kinetics-400 (K400).

name	dataset	# of frames	spatial crop	HMDB51 Top1	UCF101 Top1	url
GDT	K400	30	112	62.3	90.9	model
GDT	HT100M	30	112	67.4	94.1	model
GDT	IG65M	30	112	72.8	95.2	model

name	dataset	# of frames	spatial crop	HMDB51 Top1	UCF101 Top1	url
STiCA	K400	60	112	67.0	93.1	Coming Soon

Installation

This repo was tested with Ubuntu 16.04.5 LTS, Python 3.7.5, PyTorch 1.3.1, Torchvision 0.4.1, and CUDA 10.0.

Step 1

• Clone this repo to your local machine

Step 2

• Install required packages using conda env create -f environment.yml

Step 3

• Activate conda environment using conda activate GDT

Step 4

• Install kornia library pip install kornia==0.1.4

Step 5

• See below for how to pretrain GDT / StiCa or benchmark pretrained models

Data Preperation

For Kinetics-400/600, HMDB-51 and UCF-101 datasets:

1. Ensure all datasets are in the format:

$ROOT_DIR/$SPLIT/$CLASS/*

To prepare How-To-100M dataset, do the following:

1. Download the word2vec matrix and dictionary, unzip the file, and place in datasets/data folder.

wget https://www.rocq.inria.fr/cluster-willow/amiech/word2vec.zip
unzip word2vec.zip
mv word2vec.pth datasets/data/word2vec.pth 

2. Download the csv files of captions.

wget https://www.rocq.inria.fr/cluster-willow/amiech/howto100m/howto100m_captions.zip
unzip howto100m_captions.zip

3. Download the preprocessed HowTo100M videos (12TB in total) by filling this Google form: https://forms.gle/hztrfnFQUJWBtiki8.

Usage

GDT pretraining

To pretrain audio-visual GDT on K-400

Multi-node distributed training with SLURM cluster:

sbatch pretraining_scripts/pretrain_gdt_k400.sh ${HYPOTHESIS_DESC} ${HYPOTHESIS} 

Single-node distributed training:

python -m torch.distributed.launch --master_port=$RANDOM --nproc_per_node=2 --use_env main_gdt.py --batch_size $BS --lr $LR --hypothesis {1,2,3,4,5,6,7,8,9}

To pretrain video-text GDT on HT100M

Multi-node training with SLURM cluster:

sbatch pretraining_scripts/pretrain_gdt_ht100m.sh ${HYPOTHESIS_DESC} ${HYPOTHESIS} 

Single-node distributed training:

python -m torch.distributed.launch --master_port=$RANDOM --nproc_per_node=2 --use_env main_gdt.py --batch_size $BS --lr $LR --hypothesis {1,2,3,4,5,6,7,8,9} --dataset ht100m --decode_audio False --model vid_text_gdt --sample_rate 2

$HYPOTHESIS refers to the hypotheses explored in GDT. We experiment with the following:

1 - cross-modal baseline (cross_modal_baseline)
2 - variant to time reversal (v_reversal)
3 - invariant to time reversal (i_reversal)
4 - variant to time shift (v_shift)
5 - invariant to time shift (i_shift)
6 - variant to time reversal and variant to time shift (v_reversal_v_shift)
7 - invariant to time reversal, variant to time shift (i_reversal_v_shift)
8 - variant to time reversal, and invariant to time shift (v_reversal_i_shift)
9 - invariant to time reversal, invariant to time shift (i_reversal_i_shift)

Please modify the following in SLURM script:

• SBATCH directives (e.g. partition, nodes, constraint,)
• SAV_FOLDER
• --root_dir (path of K-400 / HT100M train directory)

All experiments were run with 8 nodes (64 GPUs, volta32). Please scale batch-size and learning-rate appropriately.

STiCA pretraining

To pretrain audio-visual STiCA on K-400

Multi-node training with SLURM cluster:

sbatch scripts/pretrain_stica.sh $NUM_FRAMES $AUD_NUM_SEC $NUM_LARGE_CROPS $NUM_SMALL_CROPS $NUM_SMALL_TCROPS $NUM_LARGE_TCROPS $NUM_LAYER

Single-node distributed training:

python -m torch.distributed.launch --master_port=$RANDOM --nproc_per_node=2 --use_env main_stica.py --batch_size $BS --base_lr $LR

Hyper-parameters:

NUM_FRAMES - number of frames (e.g. 30)
AUD_NUM_SEC - number of seconds (30f: 1sec, 60f: 2s)
NUM_LARGE_CROPS - num of large feature spatial crops (e.g. 2)
NUM_SMALL_CROPS - num of small feature spatial crops (e.g. 4)
NUM_SMALL_TCROPS - num of large feature spatial crops (e.g. 1)
NUM_LARGE_TCROPS - num of small feature spatial crops (e.g. 2)
NUM_LAYER - num of transformer pooling layers (0 == GAP, >1 is num. of transformer layers)
e.g. sbatch scripts/pretrain_stica.sh 30 1 2 4 1 2 0

Please modify the following in SLURM script:

• SBATCH directives (e.g. partition, nodes, constraint,)
• SAV_FOLDER
• --root_dir (path of K-400 / HT100M train directory)

All experiments were run with 8 nodes (64 GPUs, volta32). Please scale batch-size and learning-rate appropriately.

Benchmarking

To evaluate pretraining on video action recognition on UCF-101 and HMDB-51 datasets,

Locally:

python3 eval_video.py --dataset {ucf101, hmdb51} --fold {1,2,3} --weights-path {WEIGHTS_PATH} --model ${vid_text_gdt, stica, av_gdt}

On SLURM:

bash scripts/eval.sh ${WEIGHTS_PATH} ${OUTPUT_DIR} ${CKPT_NUM} ${CLIP_LEN} ${vid_text_gdt, stica, av_gdt} ${1, 2, 3}

Modify --root_dir, --ucf101-annotation-path, and --hmdb51-annotation-path in eval_video.py.

License

The majority of this work is licensed under CC-NC 4.0 International license.

Contributing

We actively welcome your pull requests. Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.