This repository contains the code to reproduce the results in our CVPR 2020 paper Transform and Tell: Entity-Aware News Image Captioning. We propose an end-to-end model which generates captions for images embedded in news articles. News images present two key challenges: they rely on real-world knowledge, especially about named entities; and they typically have linguistically rich captions that include uncommon words. We address the first challenge by associating words in the caption with faces and objects in the image, via a multi-modal, multi-head attention mechanism. We tackle the second challenge with a state-of-the-art transformer language model that uses byte-pair-encoding to generate captions as a sequence of word parts.
On the GoodNews dataset, our model outperforms the previous state of the art by a factor of four in CIDEr score (13 to 54). This performance gain comes from a unique combination of language models, word representation, image embeddings, face embeddings, object embeddings, and improvements in neural network design. We also introduce the NYTimes800k dataset which is 70% larger than GoodNews, has higher article quality, and includes the locations of images within articles as an additional contextual cue.
A live demo can be accessed here. In the demo, you can provide the URL to a New York Times article. The server will then scrape the web page, extract the article and image, and feed them into our model to generate a caption.
Please cite with the following BibTeX:
@InProceedings{Tran_2020_CVPR,
author = {Tran, Alasdair and Mathews, Alexander and Xie, Lexing},
title = {Transform and Tell: Entity-Aware News Image Captioning},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}
# Install Anaconda for Python and then create a dedicated environment.
# This will make it easier to reproduce our experimental numbers.
conda env create -f environment.yml
conda activate tell
# This step is only needed if you want to use the Jupyter notebook
python -m ipykernel install --user --name tell --display-name "tell"
# We also pin the apex version, which is used for mixed precision training
cd libs/apex
git submodule init && git submodule update .
pip install -v --no-cache-dir --global-option="--pyprof" --global-option="--cpp_ext" --global-option="--cuda_ext" ./
# Install our package
cd ../.. && python setup.py develop
# Spacy is used to calcuate some of the evaluation metrics
spacy download en_core_web_lg
# We use nltk to tokenize the generated text to compute linguistic metrics
python -m nltk.downloader punktThe quickest way to get the data is to send an email to first.last@anu.edu.au
(where first is alasdair and last is tran) to request the MongoDB dump
that contains the dataset. Alternatively, see here for
instructions on how to get the data from scratch, which will take a few days.
Once we have obtained the data from the authors, which consists of two
directories expt and data, you can simply put them at the root of this
repo.
# First, let's start an empty local MongoDB server on port 27017. Below
# we set the cache size to 10GB of RAM. Change it depending on your system.
mkdir data/mongodb
mongod --bind_ip_all --dbpath data/mongodb --wiredTigerCacheSizeGB 10
# Next let's restore the NYTimes200k and GoodNews datasets
mongorestore --db nytimes --host=localhost --port=27017 --drop --gzip --archive=data/mongobackups/nytimes-2020-04-21.gz
mongorestore --db goodnews --host=localhost --port=27017 --drop --gzip --archive=data/mongobackups/goodnews-2020-04-21.gz
# Next we unarchive the image directories. For each dataset, you can see two
# directories: `images` and `images_processed`. The files in `images` are
# the orignal files scraped from the New York Times. You only need this
# if you want to recompute the face and object embeddings. Otherwise, all
# the experiments will use the images in `images_processed`, which have
# already been cropped and resized.
tar -zxf data/nytimes/images_processed.tar.gz -C data/nytimes/
tar -zxf data/goodnews/images_processed.tar.gz -C data/goodnews/
# We are now ready to train the models!You can see an example of how we read the NYTimes800k samples from the MongoDB database here. Here's a minimum working example in Python:
import os
from PIL import Image
from pymongo import MongoClient
# Assume that you've already restored the database and the mongo server is running
client = MongoClient(host='localhost', port=27017)
# All of our NYTimes800k articles sit in the database `nytimes`
db = client.nytimes
# Here we select a random article in the training set.
article = db.articles.find_one({'split': 'train'})
# You can visit the original web page where this article came from
url = article['web_url']
# Each article contains a lot of fields. If you want the title, then
title = article['headline']['main'].strip()
# If you want the article text, then you will need to manually merge all
# paragraphs together.
sections = article['parsed_section']
paragraphs = []
for section in sections:
if section['type'] == 'paragraph':
paragraphs.append(section['text'])
article_text = '\n'.join(paragraphs)
# To get the caption of the first image in the article
pos = article['image_positions'][0]
caption = sections[pos]['text'].strip()
# If you want to load the actual image into memory
image_dir = 'data/nytimes/images_processed' # change this accordingly
image_path = os.path.join(image_dir, f"{sections[pos]['hash']}.jpg")
image = Image.open(image_path)
# You can also load the pre-computed FaceNet embeddings of the faces in the image
facenet_embeds = sections[pos]['facenet_details']['embeddings']
# Object embeddings are stored in a separate collection due to a size limit in mongo
obj = db.objects.find_one({'_id': sections[pos]['hash']})
object_embeds = obj['object_features']# Train the full model on NYTimes800k. This takes around 4 days on a Titan V GPU.
# The training will populate the directory expt/nytimes/9_transformer_objects/serialization
CUDA_VISIBLE_DEVICES=0 tell train expt/nytimes/9_transformer_objects/config.yaml -f
# Once training is finished, the best model weights are stored in
# expt/nytimes/9_transformer_objects/serialization/best.th
# We can use this to generate captions on the NYTimes800k test set. This
# takes about one hour.
CUDA_VISIBLE_DEVICES=0 tell evaluate expt/nytimes/9_transformer_objects/config.yaml -m expt/nytimes/9_transformer_objects/serialization/best.th
# Compute the evaluation metrics on the test set
python scripts/compute_metrics.py -c data/nytimes/name_counters.pkl expt/nytimes/9_transformer_objects/serialization/generations.jsonlThere are also other model variants which are ablation studies. Check our paper for more details, but here's a summary:
| Experiment | Word Embedding | Language Model | Image Attention | Weighted RoBERTa | Location-Aware | Face Attention | Object Attention |
|---|---|---|---|---|---|---|---|
1_lstm_glove |
GloVe | LSTM | ✔ | ||||
2_transformer_glove |
GloVe | Transformer | ✔ | ||||
3_lstm_roberta |
RoBERTa | LSTM | ✔ | ||||
4_no_image |
RoBERTa | Transformer | |||||
5_transformer_roberta |
RoBERTa | Transformer | ✔ | ||||
6_transformer_weighted_roberta |
RoBERTa | Transformer | ✔ | ✔ | |||
7_trasnformer_location_aware |
RoBERTa | Transformer | ✔ | ✔ | ✔ | ||
8_transformer_faces |
RoBERTa | Transformer | ✔ | ✔ | ✔ | ✔ | |
9_transformer_objects |
RoBERTa | Transformer | ✔ | ✔ | ✔ | ✔ | ✔ |
-
The training and evaluation workflow is based on the AllenNLP framework.
-
The Dynamic Convolution architecture is built upon Facebook's fairseq library.
-
The ZeroMQ implementation of the demo backend server is based on bert-as-service.
-
The front-end of the demo server is created with create-react-app
-
ResNet code is adapted from the Pytorch implementation.
-
We use Ultralytics' YOLOv3 implementation.
-
FaceNet and MTCNN implementations come from here.