Home

Awesome

image caption generation by chainer

Notes after two years.

I created this basically when I was almost undergrad and then things changed a lot! I feel a bit ashamed to show my dirty coding :) Well, the pretrained models are still effective and probably good if you just want to try generating captions in English/Chinese/Japanese. However, when it comes to training, probably not the best one, and many algorithmic improvements happened after I stop the maintenance of this code base…

If you want to train image captioning, I highly recommend pytorch rather than chainer. I stopped using chainer because it always broke the compatibility when I upgrade it. For example, this code is written in chainer 1.x and never works in 2.x, 3.x, 4.x, … and 6.x (see how quickly they change the version! ). *1

If you still want to stay on this code, here’s some note that you should know

*1 This does NOT mean i do not like chainer anymore. It's still great tool to quickly prototype from scratch. I just want to say, if you want to minimize the maintenance effort for a long term, probably it is not the right tool, because of the frequent major version up that breaks the downward compatibility.

descrtiption

This repository contains an implementation of typical image caption generation based on neural network (i.e. CNN + RNN). The model first extracts the image feature by CNN and then generates captions by RNN. CNN is ResNet50 and RNN is a standard LSTM .

The training data is MSCOCO. I preprocessed MSCOCO images by extracting CNN features in advance. Then I trained the language model to generate captions. Not only English, I trained on Japanese and Chinese.

I made pre-trained models available. For English captions, the model achieves CIDEr of 0.692 (Otheres are Bleu1: 0.657, Bleu2: 0.471, Bleu3: 0.327, Bleu4: 0.228, METEOR: 0.213, ROUGE_L: 0.47) for the MSCOCO validation dataset. The scores are increased a little bit when the beam search is used. For example, CIDEr is 0.716 with beam size of 5. If you want to achieve a better score, CNN has to be fine-tuned, but I haven’t tried because it’s computationally heavier.

<img src="sample.png" >

requirements

and some more packages. Make sure to use this version of chainer by pip install chainer==1.24.0. Different versions easily break the compatability.

I use conda to manage the environment and here's what I did.

conda create  -y  --prefix=conda python=2.7
conda deactivate
conda activate ./conda
conda install --yes numpy scipy matplotlib pandas pillow #normal tools
CUDA_PATH=/usr/local/cuda-8.0/ pip install chainer==1.24.0 --no-cache-dir
CUDA_PATH=/usr/local/cuda-8.0/ pip install cupy==2.0
pip install h5py
pip install nltk

You can also see my env.yml that is basically the exact environment of mine.

citation:

If you find this implementation useful, please consider to cite:

@article{multilingual-caption-arxiv,
title={{Using Artificial Tokens to Control Languages for Multilingual Image Caption Generation}},
author={Satoshi Tsutsui, David Crandall},
journal={arXiv:1706.06275},
year={2017}
}

@inproceedings{multilingual-caption,
author={Satoshi Tsutsui, David Crandall},
booktitle = {CVPR Language and Vision Workshop},
title = {{Using Artificial Tokens to Control Languages for Multilingual Image Caption Generation}},
year = {2017}
}

I just want to generate caption!

OK, first, you need to download the models and other preprocessed files.

bash download.sh

Then you can generate caption.

#English
python sample_code_beam.py \
--rnn-model ./data/caption_en_model40.model \
--cnn-model ./data/ResNet50.model \
--vocab ./data/MSCOCO/mscoco_caption_train2014_processed_dic.json \
--gpu -1 \
--img ./sample_imgs/COCO_val2014_000000185546.jpg \

#Japanese trained from machine translated Japanese (https://github.com/apple2373/mt-mscoco)
python sample_code_beam.py \
--rnn-model ./data/caption_jp_mt_model40.model \
--cnn-model ./data/ResNet50.model \
--vocab ./data/MSCOCO/captions_train2014_jp_translation_processed_dic.json \
--gpu -1 \
--img ./sample_imgs/COCO_val2014_000000185546.jpg \


#Japanese by YJCaptions (https://github.com/yahoojapan/YJCaptions)
python sample_code_beam.py \
--rnn-model ./data/caption_jp_yj_model40.model \
--cnn-model ./data/ResNet50.model \
--vocab ./data/MSCOCO/yjcaptions26k_clean_processed_dic.json \
--gpu -1 \
--img ./sample_imgs/COCO_val2014_000000185546.jpg \

#Chinese trained from machine translated Chinese (https://github.com/apple2373/mt-mscoco)
python sample_code_beam.py \
--rnn-model ./data/caption_cn_model40.model \
--cnn-model ./data/ResNet50.model \
--vocab ./data/MSCOCO/captions_train2014_cn_translation_processed_dic.json \
--gpu -1 \
--img ./sample_imgs/COCO_val2014_000000185546.jpg \

See the help for other options. You can, for example, use beam search if you want.

I want to run caption generation module as a web API.

I have a simple script for that.

cd webapi
python server.py --rnn-model ../data/caption_en_model40.model \
--cnn-model ../data/ResNet50.model \
--vocab ../data/MSCOCO/mscoco_caption_train2014_processed_dic.json \
--gpu -1 \

curl -X POST -F image=@./sample_imgs/COCO_val2014_000000185546.jpg http://localhost:8090/predict
#you should get json

I want to train the model by myself.

I made the preprocessed files (e.g., extracted ResNet features) available. You can download like this.

bash download.sh train

Then you can train like this.

# Preprocessing

cd ./code/

## English
## make sure to downalod captions_train2014.json  from the orignal MSCOCO! 
python preprocess_MSCOCO_captions.py \
--input ../data/MSCOCO/captions_train2014.json \
--output ../data/MSCOCO/mscoco_caption_train2014_processed.json -\
-outdic ../data/MSCOCO/mscoco_caption_train2014_processed_dic.json \
--outfreq ../data/MSCOCO/mscoco_caption_train2014_processed_freq.json 

## Japanese from Yahoo
python preprocess_MSCOCO_captions.py \
--input ../data/MSCOCO/yjcaptions26k_clean.json \
--output ../data/MSCOCO/yjcaptions26k_clean_processed.json \
--outdic ../data/MSCOCO/yjcaptions26k_clean_processed_dic.json \
--outfreq ../data/MSCOCO/yjcaptions26k_clean_processed_freq.json \
--cut 0 \
--char True \

## Japanese from machine translation
python preprocess_MSCOCO_captions.py \
--input ../data/MSCOCO/captions_train2014_jp_translation.json \
--output ../data/MSCOCO/captions_train2014_jp_translation_processed.json \
--outdic ../data/MSCOCO/captions_train2014_jp_translation_processed_dic.json \
--outfreq ../data/MSCOCO/captions_train2014_jp_translation_processed_freq.json \
--cut 5 \
--char True \

## Chinese from machine translation
python preprocess_MSCOCO_captions.py \
--input ../data/MSCOCO/captions_train2014_cn_translation.json \
--output ../data/MSCOCO/captions_train2014_cn_translation_processed.json \
--outdic ../data/MSCOCO/captions_train2014_cn_translation_processed_dic.json \
--outfreq ../data/MSCOCO/captions_train2014_cn_translation_processed_freq.json \
--cut 5 \
--char True \

cd ../

# Train
## train English caption 
python train_caption_model.py --savedir ./experiment1 --epoch 40 --batch 120 --gpu -1 \
--vocab ./data/MSCOCO/mscoco_caption_train2014_processed_dic.json \
--captions ./data/MSCOCO/mscoco_caption_train2014_processed.json \

## train Chinese caption by machine translation
python train_caption_model.py --savedir ./experiment1cn --epoch 50 --batch 120 --gpu 0 \
--vocab ./data/MSCOCO/captions_train2014_cn_translation_processed_dic.json \
--captions ./data/MSCOCO/captions_train2014_cn_translation_processed.json\

## train Japanese caption by machine translation
python train_caption_model.py --savedir ./experiment1jp_mt --epoch 40 --batch 120 --gpu 0 \
--vocab ./data/MSCOCO/captions_train2014_jp_translation_processed_dic.json \
--captions ./data/MSCOCO/captions_train2014_jp_translation_processed.json\
--preload True

## train Japanese caption by Yahoo's 
python train_caption_model.py --savedir ./experiment1jp_yj --epoch 40 --batch 120 --gpu 0 \
--vocab ./data/MSCOCO/yjcaptions26k_clean_processed_dic.json \
--captions ./data/MSCOCO/yjcaptions26k_clean_processed.json \
--preload True

I want to train the model from my own data.

Alright, you need to do additional amount of work.

cd code
#extract features using ResNet50 \
python ResNet_feature_extractor.py --img-dir ../data/MSCOCO/train2014 \
 --out-dir ../data/MSCOCO/train2014_ResNet50_features \
  --gpu -1

--gpu is GPU id (-1 is CPU).—img-dir is the directory that you stores images. —out-dir is the directory that the ResNet features will be saved. The file name will be the same, but extension is “.npz”.

#preprocess the json files. you need to have the same structure as MSCOCO json.
python preprocess_MSCOCO_captions.py \
--input ../data/MSCOCO/captions_train2014.json \
--output ../data/MSCOCO/mscoco_caption_train2014_processed.json \
--outdic ../data/MSCOCO/mscoco_caption_train2014_processed_dic.json \
--outfreq ../data/MSCOCO/mscoco_caption_train2014_processed_freq.json \
—-cut 5 \
—-char True

input is the json file containing caption. output will be the main preprocessed output. outdic is the vocabulary file. outfreq is the internal file you don’t need it in the training. Just frequency count. cut is the cutoff frequency for minor words. Character based chunking will be used when char is True. You can use it for non-spaced languages like Japanese and Chinese.

Then you can use my script above for training.

I want to fine-tune CNN part.

Officially, this doesn’t support the CNN tuning. That’s what I did in the original paeper. However, informally, I made it to train CNN part too... but I didn't document it, now, after two years, I don't remember very much. train_image_caption_model.py is the script to train the CNN part. I also remember that I tried to use another preprocessed json than I document here. Currently I have the _dic file and the main processed file separately but I combined them. The script to generate a preprocessed file in the new format (i.e that one compatible with train_image_caption_model.py ) should be code/preprocess_captions.py. That's what I vaguely remember.