This is the code for the paper:

Composing Text and Image for Image Retrieval - An Empirical Odyssey
Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, James Hays
CVPR 2019.

Please note that this is not an officially supported Google product. And this is the reproduced, not the original code.

If you find this code useful in your research then please cite

@inproceedings{vo2019composing,
  title={Composing Text and Image for Image Retrieval-An Empirical Odyssey},
  author={Vo, Nam and Jiang, Lu and Sun, Chen and Murphy, Kevin and Li, Li-Jia and Fei-Fei, Li and Hays, James},
  booktitle={CVPR},
  year={2019}
}

In this paper, we study the task of image retrieval, where the input query is specified in the form of an image plus some text that describes desired modifications to the input image.

We propose a new way to combine image and text using TIRG function for the retrieval task. We show this outperforms existing approaches on different datasets.

• torchvision
• pytorch
• numpy
• tqdm
• tensorboardX

• main.py: driver script to run training/testing
• datasets.py: Dataset classes for loading images & generate training retrieval queries
• text_model.py: LSTM model to extract text features
• img_text_composition_models.py: various image text compostion models (described in the paper)
• torch_function.py: contains soft triplet loss function and feature normalization function
• test_retrieval.py: functions to perform retrieval test and compute recall performance

Download the dataset from this external website.

Make sure the dataset include these files: <dataset_path>/css_toy_dataset_novel2_small.dup.npy <dataset_path>/images/*.png

To run our training & testing:

python main.py --dataset=css3d --dataset_path=./CSSDataset --num_iters=160000 \
  --model=tirg --loss=soft_triplet --comment=css3d_tirg

python main.py --dataset=css3d --dataset_path=./CSSDataset --num_iters=160000 \
  --model=tirg_lastconv --loss=soft_triplet --comment=css3d_tirgconv

The first command apply TIRG to the fully connected layer and the second applies it to the last conv layer. To run the baseline:

python main.py --dataset=css3d --dataset_path=./CSSDataset --num_iters=160000 \
  --model=concat --loss=soft_triplet --comment=css3d_concat

Download the dataset from this external website.

Make sure the dataset include these files:

<dataset_path>/images/<adj noun>/*.jpg

For training & testing:

python main.py --dataset=mitstates --dataset_path=./mitstates \
  --num_iters=160000 --model=concat --loss=soft_triplet \
  --learning_rate_decay_frequency=50000 --num_iters=160000 --weight_decay=5e-5 \
  --comment=mitstates_concat

python main.py --dataset=mitstates --dataset_path=./mitstates \
  --num_iters=160000 --model=tirg --loss=soft_triplet \
  --learning_rate_decay_frequency=50000 --num_iters=160000 --weight_decay=5e-5 \
  --comment=mitstates_tirg

Download the dataset from this external website Download our generated test_queries.txt from here.

Make sure the dataset include these files:

<dataset_path>/labels/*.txt
<dataset_path>/women/<category>/<caption>/<id>/*.jpeg
<dataset_path>/test_queries.txt`

Run training & testing:

python main.py --dataset=fashion200k --dataset_path=./Fashion200k \
  --num_iters=160000 --model=concat --loss=batch_based_classification \
  --learning_rate_decay_frequency=50000 --comment=f200k_concat

python main.py --dataset=fashion200k --dataset_path=./Fashion200k \
  --num_iters=160000 --model=tirg --loss=batch_based_classification \
  --learning_rate_decay_frequency=50000 --comment=f200k_tirg

Our pretrained models can be downloaded below. You can find our best single model accuracy: The numbers are slightly different from the ones reported in the paper due to the re-implementation.

• CSS Model: 0.760
• Fashion200k Model: 0.161
• MITStates Model: 0.132

These saved weights might not be working correctly any more with new version, please refer to #12

All log files will be saved at ./runs/<timestamp><comment>. Monitor with tensorboard (training loss, training retrieval performance, testing retrieval performance):

tensorboard --logdir ./runs/ --port 8888

Pytorch's data loader might consume a lot of memory, if that's an issue add --loader_num_workers=0 to disable loading data in parallel.