README.md

---
license: apache-2.0
---

# OFA-tiny

## Introduction
This is the **tiny** version of OFA pretrained model finetuned on CLEVR and a custom block stack dataset. 

The directory includes 4 files, namely `config.json` which consists of model configuration, `vocab.json` and `merge.txt` for our OFA tokenizer, and lastly `pytorch_model.bin` which consists of model weights.


## How to use
Download the models as shown below.
```bash
git clone https://github.com/sohananisetty/OFA_VQA.git
git clone https://huggingface.co/SohanAnisetty/ofa-vqa-base
```

After, refer the path to ofa-vqa-base to `ckpt_dir`, and prepare an image for the testing example below. 

```python
from PIL import Image
from torchvision import transforms
from transformers import OFATokenizer, OFAModelForVQA

mean, std = [0.5, 0.5, 0.5], [0.5, 0.5, 0.5]
resolution = 480
patch_resize_transform = transforms.Compose([
        lambda image: image.convert("RGB"),
        transforms.Resize((resolution, resolution), interpolation=Image.BICUBIC),
        transforms.ToTensor(), 
        transforms.Normalize(mean=mean, std=std)
    ])


tokenizer = OFATokenizer.from_pretrained(ckpt_dir)

txt = " what does the image describe?"
inputs = tokenizer([txt], return_tensors="pt").input_ids
inputs = inputs.cuda()
img = Image.open(path_to_image)
patch_img = patch_resize_transform(img).unsqueeze(0).cuda()


model = OFAModel.from_pretrained(ckpt_dir, use_cache=False).cuda()
gen = model.generate(inputs, patch_images=patch_img, num_beams=5, no_repeat_ngram_size=3) 

print(tokenizer.batch_decode(gen skip_special_tokens=True))
```