zchoi / S2-Transformer

[IJCAI 2022] Official Pytorch code for paper “S2 Transformer for Image Captioning”

S² Transformer for Image Captioning [IJCAI 2022]

Official code implementation for the paper S² Transformer for Imgae Captioning
Pengpeng Zeng, Haonan Zhang, Jingkuan Song, and Lianli Gao

Table of Contents

• Environment setup
• Data Preparation
• Training
• Evaluation
• Reference and Citation
• Acknowledgements

Environment setup

Clone this repository and create the m2release conda environment using the environment.yml file:

conda env create -f environment.yaml
conda activate m2release

Then download spacy data by executing the following command:

python -m spacy download en_core_web_md

Note

Python 3 is required to run our code. If you suffer network problems, please download en_core_web_md library from here, unzip and place it to /your/anaconda/path/envs/m2release/lib/python*/site-packages/

Data Preparation

• Annotation. Download the annotation file m2_annotations [1]. Extract and put it in the project root directory.
• Feature. Download processed image features ResNeXt-101 and ResNeXt-152 features [2] (code 9vtB), put it in the project root directory.

Update: Image features on OneDrive

Training

Run python train_transformer.py using the following arguments:

Argument	Possible values
--exp_name	Experiment name
--batch_size	Batch size (default: 50)
--workers	Number of workers, accelerate model training in the xe stage.
--head	Number of heads (default: 8)
--resume_last	If used, the training will be resumed from the last checkpoint.
--resume_best	If used, the training will be resumed from the best checkpoint.
--features_path	Path to visual features file (h5py)
--annotation_folder	Path to annotations
--num_clusters	Number of pseudo regions

For example, to train the model, run the following command:

python train_transformer.py --exp_name S2 --batch_size 50 --m 40 --head 8 --features_path /path/to/features --num_clusters 5

or just run:

bash train.sh

Note

We apply torch.distributed to train our model, you can set the worldSize in train_transformer.py to determine the number of GPUs for your training.

Evaluation

Offline Evaluation.

Run python test_transformer.py to evaluate the model using the following arguments:

python test_transformer.py --batch_size 10 --features_path /path/to/features --model_path /path/to/saved_transformer_models/ckpt --num_clusters 5

Tip

We have removed the SPICE evaluation metric during training because it is time-cost. You can add it when evaluating the model: download this file and put it in /path/to/evaluation/, then uncomment codes in init.py.

We provide checkpoint here, you will get the following results (second row):

Model	B@1	B@4	M	R	C	S
Our Paper (ResNext101)	81.1	39.6	29.6	59.1	133.5	23.2
Reproduced Model (ResNext101)	81.2	39.9	29.6	59.1	133.7	23.3

Online Evaluation

We also report the performance of our model on the online COCO test server with an ensemble of four S² models. The detailed online test code can be obtained in this repo.

Reference and Citation

Reference

[1] Cornia, M., Stefanini, M., Baraldi, L., & Cucchiara, R. (2020). Meshed-memory transformer for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.
[2] Xuying Zhang, Xiaoshuai Sun, Yunpeng Luo, Jiayi Ji, Yiyi Zhou, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Rstnet: Captioning with adaptive attention on visual and non-visual words. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15465–15474, 2021.

Citation

@inproceedings{S2,
  author    = {Pengpeng Zeng* and
               Haonan Zhang* and
               Jingkuan Song and 
               Lianli Gao},
  title     = {S2 Transformer for Image Captioning},
  booktitle = {IJCAI},
  pages     = {1608--1614}
  year      = {2022}
}

Acknowledgements

Thanks Zhang et.al for releasing the visual features (ResNeXt-101 and ResNeXt-152). Our code implementation is also based on their repo.
Thanks for the original annotations prepared by M² Transformer, and effective visual representation from grid-feats-vqa.