【Llama-3-EvoVLM-JP-v2】SakanaAI発の複数の画像対応VLMに様々な食材画像を読ませてレシピ考えさせてみた

import requests
from PIL import Image

import torch
from mantis.models.conversation import Conversation, SeparatorStyle
from mantis.models.mllava import chat_mllava, LlavaForConditionalGeneration, MLlavaProcessor
from mantis.models.mllava.utils import conv_templates
from transformers import AutoTokenizer

# 1. Set the system prompt
conv_llama_3_elyza = Conversation(
    system="<|start_header_id|>system<|end_header_id|>\n\nあなたは誠実で優秀な日本人のアシスタントです。特に指示が無い場合は、常に日本語で回答してください。",
    roles=("user", "assistant"),
    messages=(),
    offset=0,
    sep_style=SeparatorStyle.LLAMA_3,
    sep="<|eot_id|>",
)
conv_templates["llama_3"] = conv_llama_3_elyza

# 2. Load model
device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "SakanaAI/Llama-3-EvoVLM-JP-v2"

processor = MLlavaProcessor.from_pretrained("TIGER-Lab/Mantis-8B-siglip-llama3")
processor.tokenizer.pad_token = processor.tokenizer.eos_token

model = LlavaForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.float16, device_map=device).eval()

# 3. Prepare a generate config
generation_kwargs = {
    "max_new_tokens": 128,
    "num_beams": 1,
    "do_sample": False,
    "no_repeat_ngram_size": 3,
}

# 4. Generate
text = "<image>の信号は何色ですか？"
url_list = [
    "https://images.unsplash.com/photo-1694831404826-3400c48c188d?q=80&w=2070&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D",
    "https://images.unsplash.com/photo-1693240876439-473af88b4ed7?q=80&w=1974&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
]
images = [
   Image.open(requests.get(url_list[0], stream=True).raw).convert("RGB")
]

response, history = chat_mllava(text, images, model, processor, **generation_kwargs)

print(response)
# 信号の色は、青色です。

# 5. Multi-turn conversation
text = "では、<image>の信号は？"
images += [
   Image.open(requests.get(url_list[1], stream=True).raw).convert("RGB")
]
response, history = chat_mllava(text, images, model, processor, history=history, **generation_kwargs)

print(response)
# 赤色