ConvLLaVA-JP is a vision-language model that can converse about input images.
This model is an LVLM model trained using laion/CLIP-convnext_large_d_320.laion2B-s29B-b131K-ft as the image encoder and llm-jp/llm-jp-1.3b-v1.0 as the text decoder. Input of 1280 x 1280 high resolution.

Training:

This model was initially trained with Vision Projector and Stage 5 using LLaVA-Pretrain-JA.
In the second phase, it was trained Image Encoder, Vision Projector, Stage 5 and LLM using LLaVA-Pretrain-JA.
In the third phase, it was fine-tuned with Vision Projector and LLM using 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 (ROUGE-L)	JA-VLM-Bench-In-the-Wild (ROUGE-L)	Heron-Bench(Detail)	Heron-Bench(Conv)	Heron-Bench(Complex)	Heron-Bench(Average)
Japanese Stable VLM	-	40.50	25.15	51.23	37.84	38.07
EvoVLM-JP-v1-7B	19.70	51.25	50.31	44.42	40.47	45.07
Heron BLIP Japanese StableLM Base 7B llava-620k	14.51	33.26	49.09	41.51	45.72	45.44
Heron GIT Japanese StableLM Base 7B	15.18	37.82	42.77	54.20	43.53	46.83
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	13.33	44.40	50.00	51.83	48.98	50.39
ConvLLaVA-JP-1.3b-768	12.05	42.80	44.24	40.00	48.16	44.96
ConvLLaVA-JP-1.3b-1280	11.88	43.64	38.95	44.79	41.24	42.31

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.dataset import tokenizer_image_token


if __name__ == "__main__":
    model_path = 'toshi456/ConvLLaVA-JP-1.3b-1280'
    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')
    
    if device == "cuda":
        image_tensor = model.get_model().vision_tower.image_processor(image).unsqueeze(0).half().cuda().to(torch_dtype)
    else:
        image_tensor = model.get_model().vision_tower.image_processor(image).unsqueeze(0).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():
        output_id = model.generate(
            inputs=input_ids,
            images=image_tensor,
            do_sample=False,
            temperature=1.0,
            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 and Stage2 Pretrain

LLaVA-Pretrain-JA

Stage3 Fine-tuning

LLaVA-v1.5-Instruct-620K-JA

Acknowledgement

License

cc-by-nc-4.0