generate-it: data and pre-trained models

Data and pre-trained models for the code repository researchmm/generate-it, which contains two ACM MM 2021 papers on bi-directional image and text generation:

• UMT-BITG — Unifying Multimodal Transformer for Bi-directional Image and Text Generation, Yupan Huang, Bei Liu, Yutong Lu. ACM MM 2021 (Industrial Track).
• UMT-DBITG — A Picture is Worth a Thousand Words: A Unified System for Diverse Captions and Rich Images Generation, Yupan Huang, Bei Liu, Jianlong Fu, Yutong Lu. ACM MM 2021 (Video and Demo Track).

This repository is a complete, unmodified mirror of the release folder that was previously distributed through a OneDrive link. The directory layout matches the paths expected by config.py in the code repository, so you can download it as-is:

pip install -U "huggingface_hub[hf_xet]"
hf download HYPJUDY/generate-it --repo-type dataset --local-dir /path/to/generate-it
# then set root_path in config.py to /path/to/generate-it

Contents

generate-it
├── annotations
│   ├── dataset_coco.json                  # MS-COCO captions, Karpathy split
│   ├── id2captions_test.json, test_samples.json, test_gts_dict.json
│   ├── karpathy_test_result_example.json
│   ├── captions_1sample.txt, captions_5sample.txt   # demo captions for text-to-image
│   └── rprec_hard_samples/{color,noun,number,verb}.pkl
├── assets
│   ├── ACMMM2021_itp3356_poster.pdf
│   └── ACMMM2021_itp3356_slides.pdf
├── data
│   ├── clustering/                        # k-means centroids and image-id -> cluster-id maps (10k clusters, grid8)
│   ├── grid_features/readme.txt.txt       # grid features are downloaded from X-LXMERT, see below
│   └── samples/{train,trainrestval}_samples.pth
├── images
│   └── imgs_karpathy.zip                  # MS-COCO val2014 images of the Karpathy test split (for FID)
└── models
    ├── ours
    │   ├── it_generator_stage1.pth        # UMT-BITG, stage 1 (token-level training)
    │   ├── it_generator.pth               # UMT-BITG, stage 2 (sequence-level training), final model
    │   └── diverse_it_generator.pth       # UMT-DBITG
    └── pretrained
        ├── Epoch20_LXRT.pth               # X-LXMERT pre-trained transformer
        ├── G_60.pth                       # X-LXMERT image generator
        ├── bert.pth                       # BERT-base + Faster R-CNN fc6, packaged by LaBERT
        └── readme.txt

Grid features (data/grid_features/*.h5) are not included; download them from X-LXMERT as described in the code README.

Third-party files

The following files are redistributed unchanged for reproducibility and remain under their original terms:

File	Source
models/pretrained/Epoch20_LXRT.pth, models/pretrained/G_60.pth	allenai/x-lxmert
models/pretrained/bert.pth	bearcatt/LaBERT (BERT-base, Apache-2.0, plus the Faster R-CNN fc6 layer)
annotations/dataset_coco.json	Karpathy split of the MS-COCO captions (CC BY 4.0)
images/imgs_karpathy.zip	MS-COCO val2014 images; images are subject to the Flickr terms of use
annotations/id2captions_test.json, annotations/test_samples.json	generated with LaBERT's dataset code

Our own code and model weights (models/ours/*, data/clustering, data/samples, the remaining annotation files) are released under the MIT license.

Citation

@inproceedings{huang2021unifying,
  author    = {Yupan Huang and Bei Liu and Yutong Lu},
  title     = {Unifying Multimodal Transformer for Bi-directional Image and Text Generation},
  booktitle = {Proceedings of the 29th ACM International Conference on Multimedia},
  year      = {2021}
}

@inproceedings{huang2021diverse,
  author    = {Yupan Huang and Bei Liu and Jianlong Fu and Yutong Lu},
  title     = {A Picture is Worth a Thousand Words: A Unified System for Diverse Captions and Rich Images Generation},
  booktitle = {Proceedings of the 29th ACM International Conference on Multimedia},
  year      = {2021}
}