Shvetsova, N., Chen, B., Rouditchenko, A., Thomas, S., Kingsbury, B., Feris, R., Harwath, D., Glass, J. and Kuehne, H. Everything at Once – Multi-modal Fusion Transformer for Video Retrieval. In CVPR, 2022.

arXiv preprint arXiv:2112.04446

Accepted at CVPR 2022!

Repository contains:

• the code to conduct all experiments reported in the paper
• model weights to obtain main results
• data for fine-tuning and evaluation on the YouCook2 and MSR-VTT datasets

1. Create an environment:

   conda create python=3.6 -y -n everything_at_once
   conda activate everything_at_once 
   conda install -y pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=10.2 -c pytorch
   pip install gensim==3.8.0 sacred==0.8.2 humanize==3.14.0 transformers==4.10.2 librosa==0.8.1 timm==0.4.12
   pip install neptune-contrib==0.28.1 --ignore-installed certifi
   

2. If needed, download data.tar with features and spectrograms to fine-tune and evaluate on YouCook2 and MSR-VTT here. Extract a tar: tar -xvf data.tar

3. If needed, create pretrained_models folder and download model weights here:

   • Everything-At_Once (ResNet-152,ResNeXt-101)
   • Everything-At_Once (ResNet-152,ResNeXt-101), fine-tuned on YouCook2
   • Everything-At_Once (ResNet-152,ResNeXt-101), fine-tuned on MSR-VTT
   • Everything-At_Once (S3D)
   • Everything-At_Once (CLIP)
   • Everything-At_Once (ResNet-152,ResNeXt-101), text-video only

   Extract a tar:

   cd pretrained_models
   tar -xvf everything_at_once_tva.tar
   

To evaluate a pretrained everything-at-once model on the MSR-VTT dataset, run:

python test.py --n_gpu 1  \
  --config configs/evaluation/msrvtt_at_once.yaml \
  --resume pretrained_models/everything_at_once_tva/latest_model.pth

On the YouCook2 dataset:

python test.py --n_gpu 1  \
  --config configs/evaluation/youcook_at_once.yaml \
  --resume pretrained_models/everything_at_once_tva/latest_model.pth

Check out configs/evaluation folder to find more configs for evaluating models trained with S3D or CLIP features, or using other strategies to process long videos.

To fine-tune the HowTo100M-pretrained model on the MSR-VTT dataset, run:

python train.py \
  --config configs/finetuning/finetune_msrvtt.yaml \
  --resume pretrained_models/everything_at_once_tva/latest_model.pth

Add --neptune key if you want to log experiments using neptune.ai (See Experiment Logging)

On the YouCook2 dataset:

python train.py \
  --config configs/finetuning/finetune_youcook.yaml \
  --resume pretrained_models/everything_at_once_tva/latest_model.pth

Add --neptune key if you want to log experiments using neptune.ai (See Experiment Logging)

Check out configs/finetunning/clip folder to find configs for fine-tuning with CLIP features.

1. Downloading HowTo100M and feature extraction. Please note that HowTo100M videos require a huge storage, and features alone take up terabytes of space. Features extraction (ResNet-152,ResNeXt-101) and audio spectrogram extraction were carefully described in https://github.com/roudimit/AVLnet/blob/main/training.md. We will release the code for S3D and CLIP feature extraction.

2. Review configs/pretraining/everything_at_once_tva.yaml and make sure csv, features_path, features_path_audio, and caption_path point on the correct paths. CSV file should contain one column named 'path' with a list of videos. An example of the CSV file that we used in the training can be found here (HowTo100M_1166_videopaths.txt).

3. Train python train.py --config configs/pretraining/everything_at_once_tva.yaml

Add --neptune key if you want to log experiments using neptune.ai (See Experiment Logging)

Check out configs/pretraining folder to find more configs for different ablation experiments.

This repository uses Sacred with a neptune.ai for logging and tracking experiments. If you want to activate this:

1. Create a neptune.ai account.
2. Create a project, copy in your credentials (api_token, project_name) in train.py
3. Add --neptune key to the training (e.g. python train.py --neptune ..)

If you want to use the model on your own data, please follow steps described in https://github.com/roudimit/AVLnet for features extraction and audio spectrogram extraction.

You may also take a look at everything_at_once_tva.yaml, where some comments about how to define n_video_tokens and num_audio_STFT_frames are provided.

If you use this code in your research, please cite:

@inproceedings{shvetsova2022everything,
  title={Everything at Once-Multi-Modal Fusion Transformer for Video Retrieval},
  author={Shvetsova, Nina and Chen, Brian and Rouditchenko, Andrew and Thomas, Samuel and Kingsbury, Brian and Feris, Rogerio S and Harwath, David and Glass, James and Kuehne, Hilde},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={20020--20029},
  year={2022}
}

If you have any problems with the code or have a question, please open an issue or send an email to shvetsova at em.uni-frankfurt.de. I'll try to answer as soon as possible.

The main structure of the code is based on the frozen-in-time code: https://github.com/m-bain/frozen-in-time, which itself is based on the pytorch-template https://github.com/victoresque/pytorch-template. Thanks for sharing good practices!

The code in davenet.py, layers.py, avlnet.py is partly derived from https://github.com/dharwath/DAVEnet-pytorch/, https://github.com/wnhsu/ResDAVEnet-VQ, https://github.com/antoine77340/howto100m, and https://github.com/roudimit/AVLnet, and is licensed under BSD-3 (David Harwath, Wei-Ning Hsu, Andrew Rouditchenko) and Apache License 2.0 (Antoine Miech).