SoftLabelCCRF

This repository contains implementation of Phrase Grounding by Soft-Label Chain Conditional Random Field in EMNLP-IJCNLP 2019.

Setup

This repo is tested on Python 3.6 and PyTorch 1.0.0

Install dependencies: numpy, cv2, progressbar, allennlp

Download and extract the Flickr30kEntities dataset

Clone this repo. In this repo,

• Create directory feats and cache
• Create symlink to [path-to-Flickr30kEntities-dataset]

Image Feature Extraction and Caching

Clone the Bottom-Up Attention repo

Create directory [path-to-bottom-up-attention-repo]/data/flickr30k. In this directory,

• Create symlink to [path-to-Flickr30kEntities-dataset]
• Create symlink to [path-to-SoftLabelCCRF-repo]/split
• Create symlink to [path-to-SoftLabelCCRF-repo]/feats

Replace [path-to-bottom-up-attention-repo]/tools/generate_tsv.py with the one provided by us in [path-to-SoftLabelCCRF-repo]/tools/generate_tsv.py

Under [path-to-bottom-up-attention-repo], run something like

python ./tools/generate_tsv.py \
    --gpu 1,2,3,4,5,6,7 \
    --def models/vg/ResNet-101/faster_rcnn_end2end_final/test.prototxt \
    --cfg experiments/cfgs/faster_rcnn_end2end_resnet.yml \
    --net data/faster_rcnn_models/resnet101_faster_rcnn_final.caffemodel \
    --split flickr30k \
    --out data/flickr30k/feats/feats.tsv

Region proposals and features will appear in [path-to-SoftLabelCCRF-repo]/feats

Under [path-to-SoftLabelCCRF-repo], run something like

python3.6 SoftLabelCCRF/cache/cache.py 1-7 29783 1000 1000

Per-image caches will appear in [path-to-SoftLabelCCRF-repo]/cache

Training and Testing

CUDA_VISIBLE_DEVICES=1 python3.6 SoftLabelCCRF/run.py \
    --output_dir [output_dir] \
    --do_train \
    --do_test

Additionally, you may use

• --kld to switch from Hard-Label to Soft-Label training
• --crf to enable CRF layer. Be sure to specify --decode [viterbi|smoothing]
• --tran_context [none|m|mlr|mlrg] to add transition score context
• --no_box_regression to disable bounding box regression. Be sure to set --gamma 0.0
• --visualize to visualize grounding results

Other flags can be found in the script

Trained Models

Handle	Description	Flags	Size
HL	Hard-Label Model	N/A	477.6 MB
HL-CCRF	Hard-Label Chain CRF Model	--crf --decode viterbi	478.1 MB
SL	Soft-Label Model	--kld	477.6 MB
SL-CCRF	Soft-Label Chain CRF Model	--kld --crf --decode viterbi	478.1 MB
SL-CCRF-M	SL-CCRF, M	--kld --crf --decode viterbi --tran_context m	478.2 MB
SL-CCRF-MLR	SL-CCRF, MLR	--kld --crf --decode viterbi --tran_context mlr	478.6 MB
SL-CCRF-MLRG	SL-CCRF, MLRG	--kld --crf --decode viterbi --tran_context mlrg	478.8 MB
SL-CCRF-M-NBR	SL-CCRF, M, no box regression	--kld --crf --decode viterbi --tran_context m --no_box_regression --gamma 0.0	478.2 MB

Citation

If you find this repo useful, please consider citing the following work

@inproceedings{liu-hockenmaier-2019-phrase,
    title = "Phrase Grounding by Soft-Label Chain Conditional Random Field",
    author = "Liu, Jiacheng  and
      Hockenmaier, Julia",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
    month = nov,
    year = "2019",
    address = "Hong Kong, China",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/D19-1515",
    pages = "5115--5125",
    abstract = "The phrase grounding task aims to ground each entity mention in a given caption of an image to a corresponding region in that image. Although there are clear dependencies between how different mentions of the same caption should be grounded, previous structured prediction methods that aim to capture such dependencies need to resort to approximate inference or non-differentiable losses. In this paper, we formulate phrase grounding as a sequence labeling task where we treat candidate regions as potential labels, and use neural chain Conditional Random Fields (CRFs) to model dependencies among regions for adjacent mentions. In contrast to standard sequence labeling tasks, the phrase grounding task is defined such that there may be multiple correct candidate regions. To address this multiplicity of gold labels, we define so-called Soft-Label Chain CRFs, and present an algorithm that enables convenient end-to-end training. Our method establishes a new state-of-the-art on phrase grounding on the Flickr30k Entities dataset. Analysis shows that our model benefits both from the entity dependencies captured by the CRF and from the soft-label training regime. Our code is available at {\textbackslash}url{github.com/liujch1998/SoftLabelCCRF}",
}