This repository includes the code of the ICCV 2019 paper Generating Diverse and Descriptive Image Captions Using Visual Paraphrases.
We refer to our model as DCVP (Diverse and descriptive Captioning using Visual Paraphrases).
Our code is based on Ruotian Luo's implementation of Self-critical Sequence Training for Image Captioning, available here. Please refer to the project for more details.
Python 2.7 (because there is no coco-caption version for python 3)
PyTorch 0.4 (along with torchvision, it seems to be OK with other versions like PyTorch 1.0)
coco-caption, PIL, h5py
Download preprocessed coco captions from link from Karpathy's homepage. Extract dataset_coco.json from the zip file and copy it in to data/. This file provides preprocessed captions and also standard train-val-test splits.
Then do:
$ python scripts/prepro_labels.py --input_json data/dataset_coco.json --output_json data.json --output_h5 dataprepro_labels.py will map all words that occur <= 5 times to a special UNK token, and create a vocabulary for all the remaining words. The image information and vocabulary are dumped into data.json and discretized caption data are dumped into data_label.h5.
Download the coco images from link. We need 2014 training images and 2014 val. images. You should put the train2014/ and val2014/ in the same directory, denoted as $IMAGE_ROOT.
Download pre-extracted feature from link. You can either download adaptive one (used in our paper) or fixed one.
For example:
mkdir data/bu_data; cd data/bu_data
wget https://storage.googleapis.com/bottom-up-attention/trainval.zip
unzip trainval.zip
Then:
python script/make_bu_data.py --output_dir data/cocobuThis will create data/cocobu_fc, data/cocobu_att and data/cocobu_box.
We provide h5 file for visual paraphrase pairs using len(0), yngve(0), Tdiv(0.1, 0.3) and IR(2) scoring functions.
Please refer to our paper for the details to select visual paraphrase pairs.
The visual paraphrase pairs h5 file should have the same keys as the h5 file of preprocessing labels (data_label.h5). And values will be an HDF5 dataset from numpy arrays in shape (|P|, 2), where |P| is the number of valid paraphrase pairs for the specific image and scoring function. If one pair is [0, 3], it means the first caption (numbered 0) and the fourth caption (numbered 3) of the image constitutes a valid paraphrase pair. The order of captions should follow the order in data_label.h5.
Download Link on BaiduYun (password: yppa ).
Download visual paraphrase pairs h5 file (in folder visual_paraphrase_pairs/), df_processed.pkl(document frequency file), and text-image retrieval model from VSE++ for evaluation on R@K metric.
We also provide preprocessed captions used for the paper.
Folder self_retrieval_model/ contains the cross-modal retrieval model used for IR scoring function to construct visual paraphrase pairs. See DiscCaptioning project for training the model.
We provide model checkpoints of DCVP(Tdiv, 0.1), DCVP(Tdiv, 0.3) and DCVP(IR, 2), and model output json results on Karpathy test set.
If you only need the evaluation code, please download evaluation.zip.
$ python train_pair.py --id tdiv_0.1 --caption_model att2in2p --pair_h5 <PAIR_H5_PATH> --batch_size 16 --learning_rate 5e-4 --checkpoint_path cp_tdiv_0.1 --save_checkpoint_every 6000 --max_epochs 30<PAIR_H5_PATH> is the path of h5 file containing information of visual paraphrase pairs, such as pair_tdiv_0.1.h5.
The train script will dump checkpoints into the folder specified by --checkpoint_path (default = save/). We only save the best-performing checkpoint on validation and the latest checkpoint to save disk space.
To resume training, you can specify --start_from option to be the path saving infos.pkl and model.pth (usually you could just set --start_from and --checkpoint_path to be the same).
If you have tensorflow, the loss histories are automatically dumped into --checkpoint_path, and can be visualized using tensorboard.
If you'd like to evaluate BLEU/CIDEr/SPICE scores during training in addition to validation cross entropy loss, use --language_eval 1 option, but don't forget to download the coco-caption code into the directory.
For more options, see opts.py. Please refer to self-critical.pytorch repo for more details.
$ python eval.py --model cp_tdiv_0.1/model-best.pth --infos_path cp_tdiv_0.1/infos_tdiv_0.1.pkl --beam_size 3The default split to evaluate is test. The default inference method is greedy decoding (--sample_max 1), to sample from the posterior, set --sample_max 0.
Beam Search. Beam search can increase the performance of the search for greedy decoding sequence by ~5%. However, this is a little more expensive. To turn on the beam search, use --beam_size N, N should be greater than 1.
In default, the code will evaluate conventional automatic metrics results of final captions (from the second decoder) and preliminary captions (from the first decoder) generated from the two-stage decoder.
If you have model output results in json, please use eval_diversity.py and eval_retrieval.py from this repo or evaluate.zip from download link above to evaluate results on Dist-N metric for caption diversity and R@K metric for caption descriptiveness(distinctness, discriminability).
The evaluation of R@K relies on raw images from Karpathy test set and a pretrained cross-model retrieval model (We use VSE++ with fine-tuned ResNet-152. Other state-of-the-art retrieval models such as ViLBERT-multi-task can also be used).
If you find our work or this repo useful, please consider citing:
@InProceedings{Liu_2019_ICCV,
author = {Liu, Lixin and Tang, Jiajun and Wan, Xiaojun and Guo, Zongming},
title = {Generating Diverse and Descriptive Image Captions Using Visual Paraphrases},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}
Thanks Ruotian Luo for his self-critical.pytorch and DiscCaptioning repositories.