Multimodal Neural Machine Translation

Reference: Iacer Calixto's MultimodalNMT

This is the implementation of four different multi-modal neural machine translation models described in the research papers (1) and (2). They are based on the Pytorch port of OpenNMT, an open-source (MIT) neural machine translation system.

Related Work

• papers

Dataset

• Multi30k Dataset

Requirements

The code is successfully tested on PyTorch=1.3.1 and torchtext=0.2.3. If you have any questions, you could see my issues or feel free to add your issue.

In case one of the two are missing or not up-to-date and assuming you installed pytorch using the conda package manager and torchtext using pip, you might want to run the following:

conda install -c soumith pytorch
pip install torchtext==0.2.3
pip install -r requirements.txt

Run the Code

Step 0: Extract the image features for the Multi30k data set.

If you are using image features extracted by someone else, you can skip this step.

We assume you have downloaded the Multi30k data set and have the training, validation and test images locally (make sure you download the test2016 test set). Together with the image files, you need text files with the image file names in the training, validation, and test sets, respectively. These are named train_images.txt,val_images.txt, and test_images.txt, and are part of the original Flickr30k data set. If you download them from the WMT Multi-modal MT shared task website, you might need to adjust the file names accordingly.

In order to extract the image features, run the following script:

python extract_image_features.py --gpuid 0 --pretrained_cnn vgg19_bn --splits=train,valid,test --images_path ./path/to/flickr30k/images/ --train_fnames ./path/to/flickr30k/train_images.txt --valid_fnames ./path/to/flickr30k/val_images.txt --test_fnames ./path/to/flickr30k/test2016_images.txt

This will use GPU 0 to extract features with the pre-trained VGG19 with batch normalisation, for the training, validation and test sets of the Flickr30k. Change the name of the pre-trained CNN to any of the CNNs available under this repository, and the model will automatically use this CNN to extract features. This script will extract both global and local visual features.

Step 1: Preprocess the data

That is the same way as you would do with a text-only NMT model. Important: the preprocessing script only uses the textual portion of the multi-modal machine translation data set!

In here, we assume you have downloaded the Multi30k data set and extracted the sentences in its training, validation and test sets. After pre-processing them (e.g. tokenising, lowercasing, and applying a BPE model), feed the training and validation sets to the preprocess.py script, as below.

python preprocess.py -train_src ./path/to/flickr30k/train.norm.tok.lc.10000bpe.en -train_tgt ./path/to/flickr30k/train.norm.tok.lc.10000bpe.de -valid_src ./path/to/flickr30k/val.norm.tok.lc.10000bpe.en -valid_tgt ./path/to/flickr30k/val.norm.tok.lc.10000bpe.de -save_data ./data/m30k

Step 2: Train the model

To train a multi-modal NMT model, use the train_mm.py script. In addition to the parameters accepted by the standard train.py (that trains a text-only NMT model), this script expects the path to the training and validation image features, as well as the multi-modal model type (one of imgd, imge, imgw, or src+img).

For a complete description of the different multi-modal NMT model types, please refer to the papers where they are described (1) and (2).

python train_mm.py -data data/m30k -save_model model_snapshots/IMGD_ADAM -gpuid 0 -epochs 25 -batch_size 40 -path_to_train_img_feats ./flickr30k_train_vgg19_bn_cnn_features.hdf5 -path_to_valid_img_feats ./flickr30k_valid_vgg19_bn_cnn_features.hdf5 -optim adam -learning_rate 0.002 -use_nonlinear_projection --multimodal_model_type imgd

In case you want to continue training from a previous checkpoint, simply run (for example):

MODEL_SNAPSHOT=IMGD_ADAM_acc_60.79_ppl_8.38_e4.pt
python train_mm.py -data data/m30k -save_model model_snapshots/IMGD_ADAM -gpuid 0 -epochs 25 -batch_size 40 -path_to_train_img_feats /path/to/flickr30k/features/flickr30k_train_vgg19_bn_cnn_features.hdf5 -path_to_valid_img_feats /path/to/flickr30k/features/flickr30k_valid_vgg19_bn_cnn_features.hdf5 -optim adam -learning_rate 0.002 -use_nonlinear_projection --multimodal_model_type imgd -train_from model_snapshots/${MODEL_SNAPSHOT}

As an example, if you wish to train a doubly-attentive NMT model (referred to as src+img), try the following command:

python train_mm.py -data data/m30k -save_model model_snapshots/NMT-src-img_ADAM -gpuid 0 -epochs 25 -batch_size 40 -path_to_train_img_feats /path/to/flickr30k/features/flickr30k_train_vgg19_bn_cnn_features.hdf5 -path_to_valid_img_feats /path/to/flickr30k/features/flickr30k_valid_vgg19_bn_cnn_features.hdf5 -optim adam -learning_rate 0.002 -use_nonlinear_projection --decoder_type doubly-attentive-rnn --multimodal_model_type src+img

Step 3: Translate new sentences

To translate a new test set, simply use translate_mm.py similarly as you would use the original translate.py script, with the addition of the path to the file containing the test image features. In the example below, we translate the Multi30k test set used in the 2016 run of the WMT Multi-modal MT Shared Task.

MODEL_SNAPSHOT=IMGD_ADAM_acc_60.79_ppl_8.38_e4.pt
python translate_mm.py -src ~/exp/opennmt_imgd/data_multi30k/test2016.norm.tok.lc.bpe10000.en -model model_snapshots/${MODEL_SNAPSHOT} -path_to_test_img_feats ~/resources/multi30k/features/flickr30k_test_vgg19_bn_cnn_features.hdf5 -output model_snapshots/${MODEL_SNAPSHOT}.translations-test2016

Todo

• Strange Results

Citation

If you use the multi-modal NMT models in this repository, please consider citing the research papers where they are described (1) and (2):

@InProceedings{CalixtoLiu2017EMNLP,
  Title                    = {{Incorporating Global Visual Features into Attention-Based Neural Machine Translation}},
  Author                   = {Iacer Calixto and Qun Liu},
  Booktitle                = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
  Year                     = {2017},
  Address                  = {Copenhagen, Denmark},
  Url                      = {http://aclweb.org/anthology/D17-1105}
}

@InProceedings{CalixtoLiuCampbell2017ACL,
  author    = {Calixto, Iacer  and  Liu, Qun  and  Campbell, Nick},
  title     = {{Doubly-Attentive Decoder for Multi-modal Neural Machine Translation}},
  booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  month     = {July},
  year      = {2017},
  address   = {Vancouver, Canada},
  publisher = {Association for Computational Linguistics},
  pages     = {1913--1924},
  url       = {http://aclweb.org/anthology/P17-1175}
}

If you use OpenNMT, please cite as below.

OpenNMT technical report

@inproceedings{opennmt,
  author    = {Guillaume Klein and
               Yoon Kim and
               Yuntian Deng and
               Jean Senellart and
               Alexander M. Rush},
  title     = {OpenNMT: Open-Source Toolkit for Neural Machine Translation},
  booktitle = {Proc. ACL},
  year      = {2017},
  url       = {https://doi.org/10.18653/v1/P17-4012},
  doi       = {10.18653/v1/P17-4012}
}