RSTNet: Relationship-Sensitive Transformer Network

This repository contains the reference code for the paper RSTNet: Captioning with Adaptive Attention on Visual and Non-Visual Words (CVPR 2021).

Tips

Sometimes I may not be able to answer issues in time.
if you are in a hurry, you can add my wechat: zcclovelr with the remark 'RSTNet'.

Environment setup

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

conda env create -f environment.yml
conda activate m2release

Then download spacy data by executing the following command:

python -m spacy download en

Note: Python 3.6 is required to run our code.

Data preparation

To run the code, annotations and visual features for the COCO dataset are needed.

First, most annotations have been prepared by [1], please download annotations.zip and rename the extracted folder as m2_annotations, please download image_info_test2014.json and put it into m2_annotations.

Then, visual features are computed with the code provided by [2]. To reproduce our result, please download the COCO features file X-101-features.tgz and rename the extracted folder as X101-features. Note that this visual features are huge, you can alternatively save the features as float16 for storage space saving by executing the following command:

python switch_datatype.py

Finally, in order to solve the shape difference and match the feat shape with region feat shape (50 regions), please execute the following command to reshape the visual to 49(7x7) and save all visual features as a h5py file.

python feats_process.py

Note that, you can also use my processed offline image features COCO-X-101-grid.hdf5 with extraction code wsvg and my processed online image features X101_grid_feats_coco_test.hdf5 with extraction code qzwm for convenience.

Besides, if you want to extract grid features of your custom image dataset, you can refer to the code grid-feats-vqa .

Training procedure

Run python train_language.py and python train_transformer.py in sequence using the following arguments:

For example, to train our BERT-based language model with the parameters used in our experiments, use

python train_language.py --exp_name bert_language --batch_size 50 --features_path /path/to/features --annotation_folder /path/to/annotations

to train our rstnet model with the parameters used in our experiments, use

python train_transformer.py --exp_name rstnet --batch_size 50 --m 40 --head 8 --features_path /path/to/features --annotation_folder /path/to/annotations

The figure below shows the changes of cider value during the training of rstnet. You can also visualize the training details by calling the tensorboard files in tensorboard_logs.

Evaluation

Run python test_transformer.py to evaluate the rstnet or python test_language.py to evaluate the language model using the following arguments:

Note that, you can also download our pretrained model files in the Pre-trained_Models folder to reproduce the our reported results. The results of offline evaluation (Karpathy test split of MS COCO) are as follows：

References

[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] Jiang, H., Misra, I., Rohrbach, M., Learned-Miller, E., & Chen, X. (2020). In defense of grid features for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.

Acknowledgements

Thank Cornia et.al for their open source code meshed-memory-transformer , on which our implements are based.
Thank Jiang et.al for the significant discovery in visual representation [2], which has given us a lot of inspiration.