---
license: cc-by-nc-4.0
datasets:
- turing-motors/LLaVA-Pretrain-JA
- turing-motors/LLaVA-v1.5-Instruct-620K-JA
language:
- ja
pipeline_tag: image-to-text
tags:
- vision
- image-captioning
- VQA
---

# LLaVA-JP Model Card

## Model detail

**Model type:**

LLaVA-JP is a vision-language model that can converse about input images.<br>
This model is an LVLM model trained using [google/siglip-so400m-patch14-384](https://huggingface.co/google/siglip-so400m-patch14-384) as the image encoder and [llm-jp/llm-jp-1.3b-v1.0](https://huggingface.co/llm-jp/llm-jp-1.3b-v1.0) as the text decoder. supports the input of 768 x 768 high resolution images by scaling_on_scales method. 

**Training:**

This model was initially trained with the Vision Projector using LLaVA-Pretrain-JA.<br>
In the second phase, it was fine-tuned with LLaVA-v1.5-Instruct-620K-JA.

resources for more information: https://github.com/tosiyuki/LLaVA-JP/tree/main

**Comparing VLMs**
|Model|JA-VG-VQA-500<br>(ROUGE-L)|JA-VLM-Bench-In-the-Wild<br>(ROUGE-L)|Heron-Bench(Detail)|Heron-Bench(Conv)|Heron-Bench(Complex)|Heron-Bench(Average)
|-|-|-|-|-|-|-|
|[Japanese Stable VLM](https://huggingface.co/stabilityai/japanese-stable-vlm)|-|40.50|25.15|51.23|37.84|38.07|
|[EvoVLM-JP-v1-7B](https://huggingface.co/SakanaAI/EvoVLM-JP-v1-7B)|**19.70**|**51.25**|50.31|44.42|40.47|45.07|
|[Heron BLIP Japanese StableLM Base 7B llava-620k](https://huggingface.co/turing-motors/heron-chat-blip-ja-stablelm-base-7b-v1-llava-620k)|14.51|33.26|49.09|41.51|45.72|45.44|
|[Heron GIT Japanese StableLM Base 7B](https://huggingface.co/turing-motors/heron-chat-git-ja-stablelm-base-7b-v1)|15.18|37.82|42.77|**54.20**|43.53|46.83|
|[llava-jp-1.3b-v1.0-620k](https://huggingface.co/toshi456/llava-jp-1.3b-v1.0-620k)|12.69|44.58|**51.21**|41.05|45.95|44.84|
|[llava-jp-1.3b-v1.1](https://huggingface.co/toshi456/llava-jp-1.3b-v1.1)|13.33|44.40|50.00|51.83|**48.98**|**50.39**|

![image/png](https://cdn-uploads.huggingface.co/production/uploads/630af71ffaaea618ebc973db/rnzCN-LFpK4iDL5RZ9oyI.png)

## How to use the model
**1. Download dependencies**
```
git clone https://github.com/tosiyuki/LLaVA-JP.git
```

**2. Inference**
```python
import requests
import torch
import transformers
from PIL import Image

from transformers.generation.streamers import TextStreamer
from llava.constants import DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_INDEX
from llava.conversation import conv_templates, SeparatorStyle
from llava.model.llava_gpt2 import LlavaGpt2ForCausalLM
from llava.train.arguments_dataclass import ModelArguments, DataArguments, TrainingArguments
from llava.train.dataset import tokenizer_image_token


if __name__ == "__main__":
    model_path = 'toshi456/llava-jp-1.3b-v1.1'
    device = "cuda" if torch.cuda.is_available() else "cpu"
    torch_dtype = torch.bfloat16 if device=="cuda" else torch.float32

    model = LlavaGpt2ForCausalLM.from_pretrained(
        model_path, 
        low_cpu_mem_usage=True,
        use_safetensors=True,
        torch_dtype=torch_dtype,
        device_map=device,
    )
    tokenizer = transformers.AutoTokenizer.from_pretrained(
        model_path,
        model_max_length=1532,
        padding_side="right",
        use_fast=False,
    )
    model.eval()

    conv_mode = "v1"
    conv = conv_templates[conv_mode].copy()

    # image pre-process
    image_url = "https://huggingface.co/rinna/bilingual-gpt-neox-4b-minigpt4/resolve/main/sample.jpg"
    image = Image.open(requests.get(image_url, stream=True).raw).convert('RGB')
    
    image_size = model.get_model().vision_tower.image_processor.size["height"]
    if model.get_model().vision_tower.scales is not None:
        image_size = model.get_model().vision_tower.image_processor.size["height"] * len(model.get_model().vision_tower.scales)
    
    if device == "cuda":
        image_tensor = model.get_model().vision_tower.image_processor(
            image, 
            return_tensors='pt', 
            size={"height": image_size, "width": image_size}
        )['pixel_values'].half().cuda().to(torch_dtype)
    else:
        image_tensor = model.get_model().vision_tower.image_processor(
            image, 
            return_tensors='pt', 
            size={"height": image_size, "width": image_size}
        )['pixel_values'].to(torch_dtype)

    # create prompt
    # ユーザー: <image>\n{prompt}
    prompt = "猫の隣には何がありますか？"
    inp = DEFAULT_IMAGE_TOKEN + '\n' + prompt
    conv.append_message(conv.roles[0], inp)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()

    input_ids = tokenizer_image_token(
        prompt, 
        tokenizer, 
        IMAGE_TOKEN_INDEX, 
        return_tensors='pt'
    ).unsqueeze(0)
    if device == "cuda":
        input_ids = input_ids.to(device)

    input_ids = input_ids[:, :-1] # </sep>がinputの最後に入るので削除する
    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
    keywords = [stop_str]
    streamer = TextStreamer(tokenizer, skip_prompt=True, timeout=20.0)

    # predict
    with torch.inference_mode():
        model.generate(
            inputs=input_ids,
            images=image_tensor,
            do_sample=True,
            temperature=0.1,
            top_p=1.0,
            max_new_tokens=256,
            streamer=streamer,
            use_cache=True,
        )
    """猫の隣にはノートパソコンがあります。"""

```

## Training dataset
**Stage1 Pretrain**
- [LLaVA-Pretrain-JA](https://huggingface.co/datasets/turing-motors/LLaVA-Pretrain-JA)

**Stage2 Fine-tuning**
- [LLaVA-v1.5-Instruct-620K-JA](https://huggingface.co/datasets/turing-motors/LLaVA-v1.5-Instruct-620K-JA)

## Acknowledgement
- [LLaVA](https://llava-vl.github.io/)
- [LLM-jp](https://llm-jp.nii.ac.jp/)
- [scaling_on_scales](https://github.com/bfshi/scaling_on_scales/tree/master)

## License
cc-by-nc-4.0