File size: 3,862 Bytes
2e84f1d 75b5b6a 2e84f1d 9e15c84 2e84f1d 9e15c84 2e84f1d 5b01b86 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 |
---
language:
- en
tags:
- llava
- multimodal
- qwen
license: apache-2.0
pipeline_tag: image-text-to-text
---
# nanoLLaVA-1.5 - Improved sub 1B Vision-Language Model
<p align="center">
<img src="https://i.postimg.cc/d15k3YNG/nanollava.webp" alt="Logo" width="350">
</p>
## Description
nanoLLaVA-1.5 is a "small but mighty" 1B vision-language model designed to run efficiently on edge devices. This is an update from the v1.0 version [qnguyen3/nanoLLaVA](https://huggingface.co/qnguyen3/nanoLLaVA)
- **Base LLM**: [Quyen-SE-v0.1](https://huggingface.co/vilm/Quyen-SE-v0.1) (Qwen1.5-0.5B)
- **Vision Encoder**: [google/siglip-so400m-patch14-384](https://huggingface.co/google/siglip-so400m-patch14-384)
| Model | **VQA v2** | **TextVQA** | **ScienceQA** | **POPE** | **MMMU (Test)** | **MMMU (Eval)** | **GQA** | **MM-VET** |
|---------|--------|---------|-----------|------|-------------|-------------|------|--------|
| nanoLLavA-1.0 | 70.84 | 46.71 | 58.97 | 84.1 | 28.6 | 30.4 | 54.79| 23.9 |
| nanoLLavA-1.5 | TBD | TBD | TBD | TBD | TBD | TBD | TBD| TBD |
## Training Data
Training Data will be released later as I am still writing a paper on this. Expect the final final to be much more powerful than the current one.
## Finetuning Code
Coming Soon!!!
## Usage
You can use with `transformers` with the following script:
```bash
pip install -U transformers accelerate flash_attn
```
```python
import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import warnings
# disable some warnings
transformers.logging.set_verbosity_error()
transformers.logging.disable_progress_bar()
warnings.filterwarnings('ignore')
# set device
torch.set_default_device('cuda') # or 'cpu'
model_name = 'qnguyen3/nanoLLaVA-1.5'
# create model
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,
device_map='auto',
trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
model_name,
trust_remote_code=True)
# text prompt
prompt = 'Describe this image in detail'
messages = [
{"role": "user", "content": f'<image>\n{prompt}'}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
print(text)
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, sample images can be found in images folder
image = Image.open('/path/to/image.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=2048,
use_cache=True)[0]
print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())
```
## Prompt Format
The model follow the ChatML standard, however, without `\n` at the end of `<|im_end|>`:
```
<|im_start|>system
Answer the question<|im_end|><|im_start|>user
<image>
What is the picture about?<|im_end|><|im_start|>assistant
```
<!-- ---
| Image | Example |
|--------------------------------------|---------------------------------------------------------------------------------------------|
|  | **What is the text saying?** <br> "Small but mighty". <br>**How does the text correlate to the context of the image?** <br> The text seems to be a playful or humorous representation of a small but mighty figure, possibly a mouse or a mouse toy, holding a weightlifting bar. |
--- -->
Model is trained using a modified version from [Bunny](https://github.com/BAAI-DCAI/Bunny/tree/main/bunny) |