File size: 2,911 Bytes
d52b101
 
 
e602494
5549641
80355e0
 
e602494
 
 
 
80355e0
e602494
a96038b
e602494
 
f9c8e7a
e602494
 
 
f963786
e602494
 
ca2ea37
 
e602494
 
980ce70
e602494
 
 
 
 
 
 
ca2ea37
e602494
 
ca2ea37
 
e602494
 
 
ca2ea37
b9703c0
ca2ea37
89faeab
 
 
 
 
 
 
ca2ea37
 
 
 
 
b9703c0
ca2ea37
 
 
e602494
 
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
---
license: apache-2.0
---

# OFA-large

## Introduction
This is the **large** version of OFA pretrained model. 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.
```bash
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
```
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))
```