README.md

---
license: mit
language:
- pt
tags:
- CAPIVARA
- Portuguese CLIP
- OpenCLIP
datasets:
- conceptual_captions
- PraCegoVer
- MS_COCO
- Flickr30K
- ImageNet
- ELEVATER
---
# Model Card for CAPIVARA

CAPIVARA is a cost-efficient framework designed to enhance the performance of multilingual CLIP models in low-resource languages. 
This model  holds the state of the art in many zero-shot tasks involving images and Portuguese texts.

## How to use
```python
from utils.capivara_utils import download_pretrained_from_hf

model_path = download_pretrained_from_hf(model_id="hiaac-nlp/CAPIVARA")
model = OpenCLIPWrapper.load_from_checkpoint(model_path, strict=False).model
```

For more details refer to [Github repo](https://github.com/hiaac-nlp/CAPIVARA/).


## Model Details

### Model Description

CAPIVARA is built upon pre-trained [OpenCLIP ViT-B/32 XLM-Roberta Base](https://huggingface.co/laion/CLIP-ViT-B-32-xlm-roberta-base-laion5B-s13B-b90k) and 
fine-tuned with [Conceptual Captions](https://aclanthology.org/P18-1238.pdf) and synthetic captions generated by [BLIP2](https://huggingface.co/Salesforce/blip2-opt-2.7b-coco).
All the captions are translated with Google Translator.


## Uses

### Direct Use

Zero-shot image classification, zero-shot image and text retrieval, etc.

### Downstream Use

Image classification and other image task fine-tuning, linear probe image classification, 
image captioning, image generation guiding and conditioning, etc.


## Ethical considerations

For ethical considerations, please, refer to the Model Cards section in the [paper](https://arxiv.org/abs/2310.13683).

## Training Details

### Training Data
The model was fine-tuned with [Conceptual Captions](https://aclanthology.org/P18-1238.pdf) and synthetic captions generated by [BLIP2](https://huggingface.co/Salesforce/blip2-opt-2.7b-coco).
All the captions are translated with Google Translator.

#### Training Hyperparameters
```
Optimizer: "Adam"
eps: 1e-8
weight_decay: 0.2
betas: [ 0.9, 0.98 ]

LR_scheduler: "CosineWarmupLR"
min_learning_rate: 1e-7
max_learning_rate: 5e-7
warmup_lr: 500

batch_size: 2816
max_steps: 5863 # 10 epochs
```

## Evaluation

+ [Zero-shot image classification](https://github.com/hiaac-nlp/CAPIVARA/blob/main/clip_pt/src/evaluate/capivara_classification.ipynb)
+ [Zero-shot cross-modal retrieval](https://github.com/hiaac-nlp/CAPIVARA/blob/main/clip_pt/src/evaluate/capivara_retrieval.ipynb)


### Testing Data, Factors & Metrics

#### Testing Data

For cross-modal retrieval, we used [PraCegoVer](https://www.mdpi.com/2306-5729/7/2/13), which is composed of images annotated originally 
with Portuguese texts, and our Portuguese-translated versions of [MS COCO](https://link.springer.com/chapter/10.1007/978-3-319-10602-1_48) 
and [Flickr30k](https://openaccess.thecvf.com/content_iccv_2015/papers/Plummer_Flickr30k_Entities_Collecting_ICCV_2015_paper.pdf). 
We also translate the labels from [ImageNet](https://ieeexplore.ieee.org/document/5206848) and the  
[ELEVATER](https://proceedings.neurips.cc/paper_files/paper/2022/hash/3c4688b6a76f25f2311daa0d75a58f1a-Abstract-Datasets_and_Benchmarks.html) 
benchmark datasets for image classification.

### Results

#### Zero-shot Cross-Modal Retrieval

We conducted zero-shot cross-modal retrieval experiments on Flickr30k and MS COCO with captions
translated into Portuguese, and PraCegoVer. We report the average and standard deviation for 3 runs.

<table>
<thead>
  <tr>
    <th>Models</th>
    <th colspan="2">Flickr30k</th>
    <th colspan="2"> MS COCO</th>
    <th colspan="2">PraCegoVer</th>
  </tr>
</thead>
<tbody>
  <tr>
    <td></td>
    <td>text-to-image</td>
    <td> image-to-text</td>
    <td>text-to-image</td>
    <td> image-to-text</td>
    <td>text-to-image</td>
    <td> image-to-text</td>
  </tr>
  <tr>
    <td>OpenCLIP ViT-B/32 XLM-Roberta Base (Baseline)</td>
    <td>76.23</td>
    <td>87.93</td>
    <td>52.62</td>
    <td>66.55</td>
    <td>65.36</td>
    <td><b>69.43</b></td>
  </tr>
  <tr>
    <td>CAPIVARA</td>
    <td><b>79.56 ± 0.01</b></td>
    <td><b>89.95 ± 0.04</b></td>
    <td><b>56.27 ± 0.01</b></td>
    <td><b>71.24 ± 0.01</b></td>
    <td><b>66.40 ± 0.01</b></td>
    <td>64.75 ± 0.01</td>
  </tr>
</tbody>
</table>

#### Zero-shot image classification

|            Models             | **Caltech-101** | **CIFAR-10** | **CIFAR-100** | **Country-211** | **DTD**      | **EuroSAT**  | **FER-2013** | **FGVC-Aircraft** | **Food-101** | **GTSRB**    | **Hateful-Memes** | **KITTI-Distance** | **MNIST**    | **Oxford Flowers-102** | **Oxford-IIIT Pets** | **PatchCamelyon** | **Rendered-SST2** | **RESISC-45** | **Stanford-Cars** | **PASCAL VOC-2007** | **Average**  | **ImageNet-1k** |
|:-----------------------:|:---------------:|:------------:|:-------------:|:---------------:|:------------:|:------------:|:------------:|:-----------------:|:------------:|:------------:|:-----------------:|:------------------:|:------------:|:----------------------:|:--------------------:|:-----------------:|:-----------------:|:-------------:|:-----------------:|:-------------------:|:------------:|:---------------:|
| OpenCLIP ViT-B/32 XLM-Roberta Base (Baseline) | 84.53 ± 0.00    | 93.99 ± 0.00 | 68.44 ± 0.00  | 17.82 ± 0.00    | 41.17 ± 0.00 | 47.16 ± 0.00 | 48.65 ± 0.00 | 26.30 ± 0.00      | 65.06 ± 0.00 | 43.27 ± 0.00 | 56.50 ± 0.00      | 28.41 ± 0.00       | 54.99 ± 0.00 | 50.88 ± 0.00           | 81.56 ± 0.00         | 50.96 ± 0.00      | 54.20 ± 0.00      | 58.51 ± 0.00  | 84.93 ± 0.00      | 82.09 ± 0.00        | 56.97 ± 0.00 | 45.84 ± 0.00    |
|      CAPIVARA       | 82.97 ± 0.03    | 93.85 ± 0.00 | 69.37 ± 0.01  | 17.61 ± 0.00    | 42.34 ± 0.04 | 47.77 ± 0.02 | 46.68 ± 0.05 | 25.49 ± 0.01      | 64.58 ± 0.01 | 46.34 ± 0.01 | 56.17 ± 0.00      | 33.94 ± 0.13       | 60.14 ± 0.04 | 49.93 ± 0.02           | 79.37 ± 0.00         | 51.71 ± 0.01      | 54.82 ± 0.03      | 59.71 ± 0.01  | 85.10 ± 0.02      | 82.29 ± 0.00        | **57.51 ± 0.02** | **46.06 ± 0.01**    |

## Environmental Impact

- **GPU:** 1 x Quadro RTX 8000 (48 GB)
- **Hours used:** 31 hours
- **Compute Region:** Brazil
- **Carbon footprint:** 0.5 Kg
- **Energy**: 6.49 kW


## Citation

```bibtex
@inproceedings{santos2023capivara,
  title={CAPIVARA: Cost-Efficient Approach for Improving Multilingual CLIP Performance on Low-Resource Languages},
  author={Santos, Gabriel O. dos and Moreira, Diego A. B. and Ferreira, Alef I. and Silva, Jhessica and Pereira, Luiz and Bueno, Pedro and Sousa, Thiago and Maia, Helena and da Silva, N{\'a}dia and Colombini, Esther and Pedrini, Helio and Avila, Sandra},
  booktitle = "Workshop on Multi-lingual Representation Learning (MRL), Conference on Empirical Methods in Natural Language Processing (EMNLP)",
  year = "2023"
}
```