File size: 2,956 Bytes
2f13d6c
 
 
 
 
fe49e57
 
2f13d6c
 
 
 
fe49e57
 
2f13d6c
 
 
 
 
 
 
 
fe49e57
2f13d6c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
---
license: apache-2.0
---

# OFA-large-caption

## Introduction
This is the **large** version of OFA model finetuned for **image captioning**. OFA is a unified multimodal pretrained model that unifies modalities (i.e., cross-modality, vision, language) and tasks (e.g., image generation, visual grounding, image captioning, image classification, text generation, etc.) to a simple sequence-to-sequence learning framework.

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. There is no need to worry about the mismatch between Fairseq and transformers, since we have addressed the issue yet. 


## How to use
To use it in transformers, please refer to https://github.com/OFA-Sys/OFA/tree/feature/add_transformers. Install the transformers and download the models as shown below.
```
git clone --single-branch --branch feature/add_transformers https://github.com/OFA-Sys/OFA.git
pip install OFA/transformers/
git clone https://huggingface.co/OFA-Sys/OFA-large-caption
```
After, refer the path to OFA-large to `ckpt_dir`, and prepare an image for the testing example below. Also, ensure that you have pillow and torchvision in your environment. 

```python
>>> from PIL import Image
>>> from torchvision import transforms
>>> from transformers import OFATokenizer, OFAModel
>>> from generate import sequence_generator

>>> 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
>>> img = Image.open(path_to_image)
>>> patch_img = patch_resize_transform(img).unsqueeze(0)


>>> # using the generator of fairseq version
>>> model = OFAModel.from_pretrained(ckpt_dir, use_cache=True)
>>> generator = sequence_generator.SequenceGenerator(
                    tokenizer=tokenizer,
                    beam_size=5,
                    max_len_b=16, 
                    min_len=0,
                    no_repeat_ngram_size=3,
                )
>>> data = {}
>>> data["net_input"] = {"input_ids": inputs, 'patch_images': patch_img, 'patch_masks':torch.tensor([True])}
>>> gen_output = generator.generate([model], data)
>>> gen = [gen_output[i][0]["tokens"] for i in range(len(gen_output))]

>>> # using the generator of huggingface version
>>> model = OFAModel.from_pretrained(ckpt_dir, use_cache=False)
>>> 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))
```