Hugging Face's logo

Tensoic
/
Cerule-v0.1

Image-Text-to-Text
Transformers
Safetensors
English
cerule-gemma
text-generation
custom_code
Model card Files Community
4
Cerule-v0.1
File size: 5,026 Bytes 
357f72e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
 
---
license: apache-2.0
language:
- en
pipeline_tag: image-text-to-text
---

# Cerule - A <span style="color: #4285F4;">Tiny</span> <span style="color: #DB4437;">Mighty</span> <span style="color: #F4B400;">Vision</span> <span style="color: #0F9D58;">Model</span>
### Based on Google's - <span style="color: #D56c76;">Gemma-2b + SigLIP</span>



```


 ██████╗███████╗██████╗ ██╗   ██╗██╗     ███████╗
██╔════╝██╔════╝██╔══██╗██║   ██║██║     ██╔════╝
██║     █████╗  ██████╔╝██║   ██║██║     █████╗  
██║     ██╔══╝  ██╔══██╗██║   ██║██║     ██╔══╝  
╚██████╗███████╗██║  ██║╚██████╔╝███████╗███████╗
 ╚═════╝╚══════╝╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚══════╝
                                                 


                                                                
                                                                
 ```                                                               


                        


We train and release "Cerule", a tiny yet powerful Vision Lanuage Model based on the newly released Google's [Gemma-2b](https://huggingface.co/google/gemma-2b) and Google's [SigLIP](https://huggingface.co/google/siglip-so400m-patch14-384). 

We utilise highly efficient data selection techniques with:
```
- Pretraining stage : 650K images (A LAION 2M Subset)
- Finetuning stage : 695K images (SVIT-mix-665K modified for finetuning(Dataset SOON!))
```
The training setup was `4xA100's 80GB` and took ~6 hours to pretrain and ~13 hours to finetune. We modify and adapt the training code from [LLaVA](https://github.com/haotian-liu/LLaVA). 

🚨 Training code, Data and more details to release soon!


---
| Image | Example |
|-------|---------|
| ![astronaut](examples/astronaut.png) | **Describe the image**<br>The image is a playful and surreal depiction of a man in a space suit, sitting on a chair and holding a green beer bottle. The man is wearing a white space suit, complete with a helmet and gloves. His feet are clad in black and white shoes, and he is placed on a sandy surface. The background features a large, blue planet, with a moon and a star visible in the sky. |
| ![mario](examples/mario.png) | **Who are the characters in the image?**<br>The image features three characters, two of them are Mario and Luigi, and the third one is Yoshi.<br><br>**Describe the actions of the characters**<br>The Mario and Luigi characters are holding their arms out, as if they are waving. Yoshi is standing on its own, with its arms folded. |
| ![extreme_ironing](examples/extreme_ironing.jpg) | **What's funny about this image?**<br>The image is quite humorous as it depicts a man ironing clothes on the back of a yellow taxi cab. This is not a typical sight you'd expect to see in everyday life. |
---


## Training and Inference:
We will release the training code in some time.

### Inference:
**Please note that running the inference code at this stage may result in errors**. The proper code for training and inference shall be released soon!
Before running the snippet, you need to install the following dependencies:

```shell
pip install torch transformers accelerate pillow
```

```python
import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import warnings

transformers.logging.set_verbosity_error()
transformers.logging.disable_progress_bar()
warnings.filterwarnings('ignore')

torch.set_default_device('cuda')  # or 'cpu'

model = AutoModelForCausalLM.from_pretrained(
    'Tensoic/Cerule',
    torch_dtype=torch.float16,
    device_map='auto',
    trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    'Tensoic/Cerule',
    trust_remote_code=True)

# text prompt
prompt = 'Who are these charecters?'
text = f"A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\n{prompt} ASSISTANT:"
text_chunks = [tokenizer(chunk).input_ids for chunk in text.split('<image>')]
input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1], dtype=torch.long).unsqueeze(0)

image = Image.open('examples/mario.png')
image_tensor = model.process_images([image], model.config).to(dtype=model.dtype)

# generate
output_ids = model.generate(
    input_ids,
    images=image_tensor,
    max_new_tokens=100,
    use_cache=False)[0] #keep use_cache=False or else it might run into some torch dim error

print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=False).strip())
```

## License
Apache 2.0? Maybe... idk