Probabilistic Cross-Modal Embedding (PCME) CVPR 2021

Official Pytorch implementation of PCME | Paper

Sanghyuk Chun1 Seong Joon Oh1 Rafael Sampaio de Rezende2 Yannis Kalantidis2 Diane Larlus2

## Updates
  • 23 Jun, 2021: Initial upload.

Installation

Install dependencies using the following command.

pip install cython && pip install -r requirements.txt
python -c 'import nltk; nltk.download("punkt", download_dir="/opt/conda/nltk_data")'
git clone https://github.com/NVIDIA/apex && cd apex && pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Dockerfile

You can use my docker image as well

docker pull sanghyukchun/pcme:torch1.2-apex-dali

Please Add --model__cache_dir /vector_cache when you run the code

Configuration

All experiments are based on configuration files (see config/coco and config/cub). If you want to change only a few options, instead of re-writing a new configuration file, you can override the configuration as the follows:

python <train | eval>.py --dataloader__batch_size 32 --dataloader__eval_batch_size 8 --model__eval_method matching_prob

See config/parser.py for details

Dataset preparation

COCO Caption

We followed the same split provided by VSE++. Dataset splits can be found in datasets/annotations.

Note that we also need instances_<train | val>2014.json for computing PMRP score.

CUB Caption

Download images from this link, and download caption from reedscot/cvpr2016. You can use the image path and the caption path separately in the code.

Evaluate pretrained models

NOTE: the current implementation of plausible match R-Precision (PMRP) is not efficient:
It first dumps all ranked items for each item to a local file, and compute R-precision.
We are planning to re-implement efficient PMRP as soon as possible.

COCO Caption

# Compute recall metrics
python evaluate_recall_coco.py ./config/coco/pcme_coco.yaml \
    --dataset_root <your_dataset_path> \
    --model_path model_last.pth \
    # --model__cache_dir /vector_cache # if you use my docker image


# Compute plausible match R-Precision (PMRP) metric
python extract_rankings_coco.py ./config/coco/pcme_coco.yaml \
    --dataset_root <your_dataset_path> \
    --model_path model_last.pth \
    --dump_to <dumped_ranking_file> \
    # --model__cache_dir /vector_cache # if you use my docker image

python evaluate_pmrp_coco.py --ranking_file <dumped_ranking_file>
Method I2T PMRP I2T [email protected] T2I PMRP T2I [email protected] Model file
PCME 45.0 68.8 46.0 54.6 link
PVSE K=1 40.3 66.7 41.8 53.5 -
PVSE K=2 42.8 69.2 43.6 55.2 -
VSRN 41.2 76.2 42.4 62.8 -
VSRN + AOQ 44.7 77.5 45.6 63.5 -

CUB Caption

python evaluate_cub.py ./config/cub/pcme_cub.yaml \
    --dataset_root <your_dataset_path> \
    --caption_root <your_caption_path> \
    --model_path model_last.pth \
    # --model__cache_dir /vector_cache # if you use my docker image

NOTE: If you just download file from reedscot/cvpr2016, then caption_root will be cvpr2016_cub/text_c10

If you want to test other probabilistic distances, such as Wasserstein distance or KL-divergence, try the following command:

python evaluate_cub.py ./config/cub/pcme_cub.yaml \
    --dataset_root <your_dataset_path> \
    --caption_root <your_caption_path> \
    --model_path model_last.pth \
    --model__eval_method <distance_method> \
    # --model__cache_dir /vector_cache # if you use my docker image

You can choose distance_method in ['elk', 'l2', 'min', 'max', 'wasserstein', 'kl', 'reverse_kl', 'js', 'bhattacharyya', 'matmul', 'matching_prob']

How to train

NOTE: we train each model with mixed-precision training (O2) on a single V100.
Since, the current code does not support multi-gpu training, if you use different hardware, the batchsize should be reduced.
Please note that, hence, the results couldn't be reproduced if you use smaller hardware than V100.

COCO Caption

python train_coco.py ./config/coco/pcme_coco.yaml --dataset_root <your_dataset_path> \
    # --model__cache_dir /vector_cache # if you use my docker image

It takes about 46 hours in a single V100 with mixed precision training.

CUB Caption

We use CUB Caption dataset (Reed, et al. 2016) as a new cross-modal retrieval benchmark. Here, instead of matching the sparse paired image-caption pairs, we treat all image-caption pairs in the same class as positive. Since our split is based on the zero-shot learning benchmark (Xian, et al. 2017), we leave out 50 classes from 200 bird classes for the evaluation.

  • Reed, Scott, et al. "Learning deep representations of fine-grained visual descriptions." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
  • Xian, Yongqin, Bernt Schiele, and Zeynep Akata. "Zero-shot learning-the good, the bad and the ugly." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017.

hyperparameter search

We additionally use cross-validation splits by (Xian, et el. 2017), namely using 100 classes for training and 50 classes for validation.

python train_cub.py ./config/cub/pcme_cub.yaml \
    --dataset_root <your_dataset_path> \
    --caption_root <your_caption_path> \
    --dataset_name cub_trainval1 \
    # --model__cache_dir /vector_cache # if you use my docker image

Similarly, you can use cub_trainval2 and cub_trainval3 as well.

training with full training classes

python train_cub.py ./config/cub/pcme_cub.yaml \
    --dataset_root <your_dataset_path> \
    --caption_root <your_caption_path> \
    # --model__cache_dir /vector_cache # if you use my docker image

It takes about 4 hours in a single V100 with mixed precision training.

How to cite

@inproceedings{chun2021pcme,
    title={Probabilistic Embeddings for Cross-Modal Retrieval},
    author={Chun, Sanghyuk and Oh, Seong Joon and De Rezende, Rafael Sampaio and Kalantidis, Yannis and Larlus, Diane},
    year={2021},
    booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
}

GitHub

https://github.com/naver-ai/pcme