jtengyp/NeuralBabyTalk

Pytorch code of for our CVPR 2018 paper "Neural Baby Talk"

Neural Baby Talk

Docker Setup

This repository provides a Dockerfile for setting up all dependencies and preprocessed data for COCO experiments (normal / robust / NOC). Docker support for Flickr30k experiments is not yet supported. To build the Dockerfile, just execute this from project root:

docker build -t nbt .

Before running the container, you need to get COCO dataset downloaded and kept somewhere in your filesystem. Declare two environment variables:

1. $COCO_IMAGES: path to a directory with sub-directories of images as train2014, val2014, test2015, etc...
2. $COCO_ANNOTATIONS: path to a directory with annotation files like instances_train2014.json, captions_train2014.json etc...

These directories will be attached as "volumes" to our docker container for Neural Baby Talk to use within. Run the docker image within a container in an interactive mode (bash session). Get nvidia-docker and execute this command to run the fresh built docker image.

nvidia-docker run --name nbt_container -it \
     -v $COCO_IMAGES:/workspace/neuralbabytalk/data/coco/images \
     -v $COCO_ANNOTATIONS:/workspace/neuralbabytalk/data/coco/annotations \
     --shm-size 8G -p 8888:8888 nbt /bin/bash

Ideally, shared memory size (--shm-size) of 8GB would be enough. Tune it according to your requirements / machine specifications.

Saved Checkpoints: All checkpoints will be saved in /workspace/neuralbabytalk/save. From outside the container, execute this to get your checkpoints from this container into the main filesystem: The container would expose port 8888, which can be used to host tensorboard visualizations.

docker container cp nbt_container:workspace/neuralbabytalk/save /path/to/local/filesystem/save

Skip directly to Training and Evaluation section to execute specified commands within the container.

requirement

Inference:

• pytorch
• torchvision
• torchtext

Data Preparation:

• stanford-corenlp-wrapper
• stanford-corenlp

Evaluation:

• coco-caption: Download the modified version of coco-caption and put it under tools/

Demo

Without detection bbox

With detection bbox

Constraint beam search

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.

Training and Evaluation

Data Preparation

Head to data/README.md, and prepare the data for training and evaluation.

Pretrained model

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

Standard Image Captioning

Training (COCO)

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

Evaluation (COCO)

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

Training (Flickr30k)

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

Evaluation (Flickr30k)

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

Robust Image Captioning

Training

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

Evaluation (robust-coco)

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

Novel Object Captioning

Training

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

Evaluation (noc-coco)

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

Multi-GPU Training

This codebase also support training with multiple GPU. To enable this feature, simply add --mGPUs Ture in the commnad.

Self-Critic Training and Fine-Tuning CNN

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!

More Visualization Results

Reference

If you use this code as part of any published research, please acknowledge the following paper

@inproceedings{Lu2018Neural,
author = {Lu, Jiasen and Yang, Jianwei and Batra, Dhruv and Parikh, Devi},
title = {Neural Baby Talk},
booktitle = {CVPR},
year = {2018}
}

Acknowledgement

We thank Ruotian Luo for his self-critical.pytorch repo.