ink-usc / viscoll

Code and data for the project "Visually grounded continual learning of compositional semantics"

VisCOLL

Code and data for the paper "Visually grounded continual learning of compositional phrases, EMNLP 2020". Paper

Checkout our Project website for data explorers and the leaderboard!

Installation

conda create -n viscoll-env python==3.7.5
conda activate viscoll-env
pip install -r requirements.txt

Overview

VisCOLL proposes a problem setup and studies algorithms for continual learning and compositionality over visual-linguistic data. In VisCOLL, the model visits a stream of examples with an evolving data distribution over time and learn to perform masked phrases prediction. We create COCO-shift and Flickr-shift (based on COCO-captions and Flickr30k-entities) for study.

Training and evaluation

This repo include code for running and evaluating ER/MIR/AGEM continual learning algorithms on VisCOLL datasets (COCO-shift and Flickr-shift), with VLBERT or LXMERT models.

For example, to train a VLBERT model with a memory of 10,000 examples on coco using Experience Replay (ER) algorithm, run:

python train_mlm.py --name debug --config configs/mlmcaptioning/er.yaml --seed 0 --cfg MLMCAPTION.BASE=vlbert OUTPUT_DIR=runs/

You may check walkthrough.ipynb in this repo for a detailed walkthrough of training, inference, and evaluation.

Data

We release the constructed data streams and scripts for visualization in the datasets folder.

COCO shift

COCO-shift is under datasets/coco/coco_buffer, with the name formats: task_buffer_real_split_{1}split{2}novel_comps{3}_task_partition_any_objects.pkl

1. The official data split where examples are drawn from
2. How the dataset is applied during training. We separate out 5,000 examples from original train split as the validation examples. The official val split is used as the test split
3. Whether the data is the novel-composition split of 24 held-out concept pairs

The pickled file is python list where each item is a dict:

{'annotation': {'image_id': 374041, 'id': 31929, 'caption': 'Someone sitting on an elephant standing by a fence.  '}, 'task': 'elephant', 'mask': (3, 5)}

The mask is the text span that should be masked for prediction. image_id can be used for locating the image in the dataset.

Image features for COCO

If you use the provided data stream above (i.e. do not create the non-stationary data stream itself, which requires additional extra steps such as phrase extraction), the only extra files required are,

1. Extracted image features for COCO under datasets/coco-features. We use code from this repo to extract features. We will upload our extracted features soon;

2. A json dictionary mapping of image id to image features in the file above. Included in this repo.

Measuring compositional generalization

We use data from this repo to perform evaluation of compositional generalization. Please put compositional-image-captioning/data/dataset_splits/ under datasets/novel_comps and compositional-image-captioning/data/occurrences/ under datasets/occurrences.

Flickr shift

Flickr-shift is under datasets/flickr/flickr_buffer. We use official split for Flickr30k Entities. Below is an example from the Flickr shift dataset.

{'tokens': ['a', 'motorcyclist', '"', 'pops', 'a', 'wheelie', '"', 'in', 'a', 'grassy', 'field', 'framed', 'by', 'rolling', 'hills', '.'], 'task_id': 518, 'image_id': '3353962769', 'instance_id': 8714, 'phrase_offset': (0, 2)}

The phrase offset is the text span that should be masked for prediction. image_id can be used for retrieving the image in the dataset.

Image features for Flickr

Similar to COCO, unless you want to create the datastream yourself, the only files required are,

1. Extracted image features for Flickr under datasets/flickr-features. Simiarly, we use code from this repo to perfrom extraction. The extracted features will be uploaded soon.

2. A json dictionary mapping of image id to image features in the file above. Included in this repo.

Visualizing the stream

We provide visual_stream.ipynb to visualize task distributions over time in Flickr-shift dataset.

Citation

@inproceedings{Jin2020VisuallyGC,
    title={Visually Grounded Continual Learning of Compositional Phrases},
    author={Xisen Jin and Junyi Du and Arka Sadhu and R. Nevatia and X. Ren},
    booktitle={EMNLP},
    year={2020}
}

viscoll's Issues

label not matching

{'annotation': {'image_id': 374041, 'id': 31929, 'caption': 'Someone sitting on an elephant standing by a fence. '}, 'task': 'elephant', 'mask': (3, 5)}

the image_id and the id (annatation_id) is not matching.

1. the image is an elephant.
2. but the id is annatation for an computer image.
   Not sure for other items in pickle.

setuptools == 42.0.2.post20191203

When I install each pacakge in requirements, I encountered a issue involoving "
setuptools == 42.0.2.post20191203". The tip of error is "ERROR: No matching distribution found for setuptools==42.0.2.post20191203". However, I can successfully perform the commond line "pip install setuptools==42.0.2", but the version number don't compy with 42.0.2.post20191203. Please tell me what should I do next?

Looking forward to your answer, thanks!