Inference:
Data Preparation:
Evaluation:
- coco-caption: Download the modified version of coco-caption and put it under
tools/
This code also involve the implementation of constraint beam search proposed by Peter Anderson. I'm not sure my impmentation is 100% correct, but it works well in conjuction with neural baby talk code. You can refer to this paper for more details. To enable CBS while decoding, please set the following flags:
--cbs True|False : Whether use the constraint beam search.
--cbs_tag_size 3 : How many detection bboxes do we want to include in the decoded caption.
--cbs_mode all|unqiue|novel : Do we allow the repetive bounding box? `novel` is an option only for novel object detection task.
Head to data/README.md, and prepare the data for training and evaluation.
| Task | Dataset | Backend | Batch size | Link |
|---|---|---|---|---|
| Standard image captioning | COCO | Res-101 | 100 | Pre-trained Model |
| Standard image captioning | Flickr30k | Res-101 | 50 | Pre-trained Model |
| Robust image captioning | COCO | Res-101 | 100 | Pre-trained Model |
| Novel object captioning | COCO | Res-101 | 100 | Pre-trained Model |
First, modify the cofig file cfgs/normal_coco_res101.yml with the correct file path.
python main.py --path_opt cfgs/normal_coco_res101.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30
Download Pre-trained model. Extract the tar.zip file and put it under save/.
python main.py --path_opt cfgs/normal_coco_res101.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30 --inference_only True --beam_size 3 --start_from save/coco_nbt_1024
Modify the cofig file cfgs/normal_flickr_res101.yml with the correct file path.
python main.py --path_opt cfgs/normal_flickr_res101.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30
Download Pre-trained model. Extract the tar.zip file and put it under save/.
python main.py --path_opt cfgs/normal_flickr_res101.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30 --inference_only True --beam_size 3 --start_from save/flickr30k_nbt_1024
Modify the cofig file cfgs/normal_flickr_res101.yml with the correct file path.
python main.py --path_opt cfgs/robust_coco.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30
Download Pre-trained model. Extract the tar.zip file and put it under save/.
python main.py --path_opt cfgs/robust_coco.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30 --inference_only True --beam_size 3 --start_from save/robust_coco_nbt_1024
Modify the cofig file cfgs/noc_coco_res101.yml with the correct file path.
python main.py --path_opt cfgs/noc_coco_res101.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30
Download Pre-trained model. Extract the tar.zip file and put it under save/.
python main.py --path_opt cfgs/noc_coco_res101.yml --batch_size 20 --cuda True --num_workers 20 --max_epoch 30 --inference_only True --beam_size 3 --start_from save/noc_coco_nbt_1024
This codebase also support training with multiple GPU. To enable this feature, simply add --mGPUs Ture in the commnad.
This codebase also support self-critic training and fine-tuning CNN. You are welcome to try this part and upload your trained model to the repo!
If you use this code as part of any published research, please acknowledge the following paper
@misc{Lu2018Neural,
author = {Lu, Jiasen and Yang, Jianwei and Batra, Dhruv and Parikh, Devi},
title = {Neural Baby Talk},
journal = {CVPR},
year = {2018}
}
We thank Ruotian Luo for his self-critical.pytorch repo.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent