README.md

metadata

library_name: transformers
datasets:
  - laicsiifes/flickr30k-pt-br
language:
  - pt
metrics:
  - bleu
  - rouge
  - meteor
  - bertscore
base_model: laicsiifes/swin-distilbertimbau
pipeline_tag: image-to-text
model-index:
  - name: Swin-DistilBERTimbau
    results:
      - task:
          name: Image Captioning
          type: image-to-text
        dataset:
          name: laicsiifes/flickr30k-pt-br
          type: flickr30k-pt-br
          split: test
        metrics:
          - name: Cider-D
            type: cider
            value: 66.73
          - name: BLEU@4
            type: bleu
            value: 24.65
          - name: ROUGE-L
            type: rouge
            value: 39.98
          - name: METEOR
            type: meteor
            value: 44.71
          - name: BERTScore
            type: bertscore
            value: 72.3

🎉 Swin-DistilBERTimbau for Brazilian Portuguese Image Captioning

Swin-DistilBERTimbau model trained for image captioning on Flickr30K Portuguese (translated version using Google Translator API) at resolution 224x224 and max sequence length of 512 tokens.

🤖 Model Description

The Swin-DistilBERTimbau is a type of Vision Encoder Decoder which leverage the checkpoints of the Swin Transformer as encoder and the checkpoints of the DistilBERTimbau 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-DistilBERTimbau was trained together with its buddy Swin-GPorTuguese-2.

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-distilbertimbau")
tokenizer = AutoTokenizer.from_pretrained("laicsiifes/swin-distilbertimbau")
image_processor = AutoImageProcessor.from_pretrained("laicsiifes/swin-distilbertimbau")

# 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"
}