---
license: apache-2.0
datasets:
- michelecafagna26/hl
language:
- en
metrics:
- sacrebleu
- rouge
- meteor
- spice
- cider

library_name: pytorch
tags:
- pytorch
- image-to-text
---

# Model Card: VinVL for Captioning 🖼️

[Microsoft's VinVL](https://github.com/microsoft/Oscar) base fine-tuned on [HL dataset](https://arxiv.org/abs/2302.12189?context=cs.CL)  for **scene description generation** downstream task.

# Model fine-tuning 🏋️‍

The model has been finetuned for 10 epochs on the scenes captions of the [HL dataset](https://arxiv.org/abs/2302.12189?context=cs.CL) (available on 🤗 HUB: [michelecafagna26/hl](https://huggingface.co/datasets/michelecafagna26/hl))

# Test set metrics 📈

Obtained with beam size 5 and max length 20

| Bleu-1 | Bleu-2 | Bleu-3 | Bleu-4 | METEOR | ROUGE-L | CIDEr | SPICE |
|--------|--------|--------|--------|--------|---------|-------|-------|
|  0.68  |  0.55  |  0.45  |  0.36  |  0.36  |  0.63   |  1.42 |  0.40 |


# Usage and Installation:

More info about how to install and use this model can be found here: [michelecafagna26/VinVL
](https://github.com/michelecafagna26/VinVL)

# Feature extraction ⛏️

This model has a separate Visualbackbone used to extract features.
More info about:
- the model: [michelecafagna26/vinvl_vg_x152c4](https://huggingface.co/michelecafagna26/vinvl_vg_x152c4)
- the usage: [michelecafagna26/vinvl-visualbackbone](https://github.com/michelecafagna26/vinvl-visualbackbone)

# Quick start: 🚀

```python
from transformers.pytorch_transformers import BertConfig, BertTokenizer
from oscar.modeling.modeling_bert import BertForImageCaptioning
from oscar.wrappers import OscarTensorizer

ckpt = "path/to/the/checkpoint"
device = "cuda" if torch.cuda.is_available() else "cpu"

# original code
config = BertConfig.from_pretrained(ckpt)
tokenizer = BertTokenizer.from_pretrained(ckpt)
model = BertForImageCaptioning.from_pretrained(ckpt, config=config).to(device)

# This takes care of the preprocessing
tensorizer = OscarTensorizer(tokenizer=tokenizer, device=device)

# numpy-arrays with shape (1, num_boxes, feat_size)
# feat_size is 2054 by default in VinVL
visual_features = torch.from_numpy(feat_obj).to(device).unsqueeze(0)

# labels are usually extracted by the features extractor
labels = [['boat', 'boat', 'boat', 'bottom', 'bush', 'coat', 'deck', 'deck', 'deck', 'dock', 'hair', 'jacket']]

inputs = tensorizer.encode(visual_features, labels=labels)
outputs = model(**inputs)

pred = tensorizer.decode(outputs)

# the output looks like this:
# pred = {0: [{'caption': 'in a library', 'conf': 0.7070220112800598]}
```

# Citations 🧾

VinVL model finetuned on scenes descriptions:

```BibTeX
@inproceedings{cafagna-etal-2022-understanding,
    title = "Understanding Cross-modal Interactions in {V}{\&}{L} Models that Generate Scene Descriptions",
    author = "Cafagna, Michele  and
      Deemter, Kees van  and
      Gatt, Albert",
    booktitle = "Proceedings of the Workshop on Unimodal and Multimodal Induction of Linguistic Structures (UM-IoS)",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates (Hybrid)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.umios-1.6",
    pages = "56--72",
    abstract = "Image captioning models tend to describe images in an object-centric way, emphasising visible objects. But image descriptions can also abstract away from objects and describe the type of scene depicted. In this paper, we explore the potential of a state of the art Vision and Language model, VinVL, to caption images at the scene level using (1) a novel dataset which pairs images with both object-centric and scene descriptions. Through (2) an in-depth analysis of the effect of the fine-tuning, we show (3) that a small amount of curated data suffices to generate scene descriptions without losing the capability to identify object-level concepts in the scene; the model acquires a more holistic view of the image compared to when object-centric descriptions are generated. We discuss the parallels between these results and insights from computational and cognitive science research on scene perception.",
}
```

HL Dataset paper:

```BibTeX
@inproceedings{cafagna2023hl,
  title={{HL} {D}ataset: {V}isually-grounded {D}escription of {S}cenes, {A}ctions and
{R}ationales},
  author={Cafagna, Michele and van Deemter, Kees and Gatt, Albert},
  booktitle={Proceedings of the 16th International Natural Language Generation Conference (INLG'23)},
address = {Prague, Czech Republic},
  year={2023}
}
```

Please consider citing the original project and the VinVL paper

```BibTeX

@misc{han2021image,
      title={Image Scene Graph Generation (SGG) Benchmark}, 
      author={Xiaotian Han and Jianwei Yang and Houdong Hu and Lei Zhang and Jianfeng Gao and Pengchuan Zhang},
      year={2021},
      eprint={2107.12604},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@inproceedings{zhang2021vinvl,
  title={Vinvl: Revisiting visual representations in vision-language models},
  author={Zhang, Pengchuan and Li, Xiujun and Hu, Xiaowei and Yang, Jianwei and Zhang, Lei and Wang, Lijuan and Choi, Yejin and Gao, Jianfeng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5579--5588},
  year={2021}
}
```