README.md

metadata

library_name: transformers
datasets:
  - laicsiifes/flickr30k-pt-br
language:
  - pt
metrics:
  - bleu
  - rouge
  - meteor
  - bertscore
base_model: laicsiifes/swin-gportuguese-2
pipeline_tag: image-to-text
model-index:
  - name: Swin-GPorTuguese-2
    results:
      - task:
          name: Image Captioning
          type: image-to-text
        dataset:
          name: laicsiifes/flickr30k-pt-br
          type: flickr30k-pt-br
        metrics:
          - name: Cider-D
            type: cider
            value: 64.71%
          - name: BLEU@4
            type: bleu
            value: 23.15%
          - name: ROUGE-L
            type: rouge
            value: 39.39%
          - name: METEOR
            type: meteor
            value: 44.36%
          - name: BERTScore
            type: bertscore
            value: 71.70%

🎉 Swin-GPorTuguese-2 for Brazilian Portuguese Image Captioning

Swin-GPorTuguese-2 model trained for image captioning on Flickr30K Portuguese (translated version using Google Translator API) at resolution 224x224 and max sequence length of 1024 tokens.

🤖 Model Description

The Swin-GPorTuguese-2 is a type of Vision Encoder Decoder which leverage the checkpoints of the Swin Transformer as encoder and the checkpoints of the GPorTuguese-2 as decoder. The encoder checkpoints come from Swin Trasnformer version pre-trained on ImageNet-1k at resolution 224x224.

The code used for training and evaluation is available at: https://github.com/laicsiifes/ved-transformer-caption-ptbr. In this work, Swin-GPorTuguese-2 was trained together with its buddy Swin-DistilBERTimbau.

Other models evaluated didn't achieve performance as high as Swin-DistilBERTimbau and Swin-GPorTuguese-2, namely: DeiT-BERTimbau, DeiT-DistilBERTimbau, DeiT-GPorTuguese-2, Swin-BERTimbau, ViT-BERTimbau, ViT-DistilBERTimbau and ViT-GPorTuguese-2.

🧑‍💻 How to Get Started with the Model

Use the code below to get started with the model.

import requests
from PIL import Image

from transformers import AutoTokenizer, AutoImageProcessor, VisionEncoderDecoderModel

# load a fine-tuned image captioning model and corresponding tokenizer and image processor
model = VisionEncoderDecoderModel.from_pretrained("laicsiifes/swin-gportuguese-2")
tokenizer = AutoTokenizer.from_pretrained("laicsiifes/swin-gportuguese-2")
image_processor = AutoImageProcessor.from_pretrained("laicsiifes/swin-gportuguese-2")

# perform inference on an image
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
pixel_values = image_processor(image, return_tensors="pt").pixel_values

# generate caption
generated_ids = model.generate(pixel_values)
generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(generated_text)

📈 Results

The evaluation metrics Cider-D, BLEU@4, ROUGE-L, METEOR and BERTScore (using BERTimbau) are abbreviated as C, B@4, RL, M and BS, respectively.

Model	Training	Evaluation	C	B@4	RL	M	BS
Swin-DistilBERTimbau	Flickr30K Portuguese	Flickr30K Portuguese	66.73	24.65	39.98	44.71	72.30
Swin-GPorTuguese-2	Flickr30K Portuguese	Flickr30K Portuguese	64.71	23.15	39.39	44.36	71.70

📋 BibTeX entry and citation info

@inproceedings{bromonschenkel2024comparative,
                title = "A Comparative Evaluation of Transformer-Based Vision 
                         Encoder-Decoder Models for Brazilian Portuguese Image Captioning",
               author = "Bromonschenkel, Gabriel and Oliveira, Hil{\'a}rio and 
                         Paix{\~a}o, Thiago M.",
            booktitle = "Proceedings...",
         organization = "Conference on Graphics, Patterns and Images, 37. (SIBGRAPI)",
                 year = "2024"
}