【生成AIに特化】Raspberry Pi AI HAT+ 2レビュー

sudo tee /etc/apt/sources.list.d/hailo.sources <<EOF
Types: deb
URIs: https://hailo:chahy5Zo@extranet.raspberrypi.org/hailo
Suites: trixie
Components: main
Signed-By: /usr/share/keyrings/raspberrypi-archive-keyring.pgp
EOF

sudo apt update

sudo apt full-upgrade -y

sudo reboot

sudo apt remove $(dpkg --get-selections docker.io docker-compose docker-doc podman-docker containerd runc | cut -f1)

sudo apt install ca-certificates curl

sudo install -m 0755 -d /etc/apt/keyrings

sudo curl -fsSL https://download.docker.com/linux/debian/gpg -o /etc/apt/keyrings/docker.asc

sudo chmod a+r /etc/apt/keyrings/docker.asc

sudo tee /etc/apt/sources.list.d/docker.sources <<EOF
Types: deb
URIs: https://download.docker.com/linux/debian
Suites: $(. /etc/os-release && echo "$VERSION_CODENAME")
Components: stable
Signed-By: /etc/apt/keyrings/docker.asc
EOF

sudo apt update

sudo apt install -y docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin

sudo systemctl start docker

sudo docker run hello-world

sudo apt install dkms

sudo apt install -y hailo-h10-all

sudo reboot

sudo apt install -y libssl-dev
git clone https://github.com/hailo-ai/hailo_model_zoo_genai.git
cd /home/pi/hailo_model_zoo_genai
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
cmake --build .
mkdir -p ~/.local/bin
cp ./src/apps/server/hailo-ollama ~/.local/bin/
mkdir -p ~/.config/hailo-ollama/
cp ../config/hailo-ollama.json ~/.config/hailo-ollama/
mkdir -p ~/.local/share/hailo-ollama
cp -r ../models/ ~/.local/share/hailo-ollama
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

hailo-ollama

curl --silent http://localhost:8000/hailo/v1/list

curl --silent http://localhost:8000/api/pull \
     -H 'Content-Type: application/json' \
     -d '{ "model": "qwen2:1.5b", "stream" : true }'

curl --silent http://localhost:8000/api/chat \
     -H 'Content-Type: application/json' \
     -d '{"model": "qwen2:1.5b", "messages": [{"role": "user", "content": "Translate to French: The cat is on the table."}]}'

sudo docker pull ghcr.io/open-webui/open-webui:main

sudo docker run -d \
  -e OLLAMA_BASE_URL=http://127.0.0.1:8000 \
  -v open-webui:/app/backend/data \
  --name open-webui \
  --network=host \
  --restart always \
  ghcr.io/open-webui/open-webui:main

sudo docker logs open-webui -f

http://127.0.0.1:8080

import requests
import json
import time

# ローカルで起動しているLLMのAPIエンドポイント
url = "http://localhost:8000/api/chat"

# ユーザーの質問
question = "おすすめの気分転換の方法を教えてください。"

payload = {
    # 使用するLLMモデルを指定
    "model": "qwen2:1.5b",
    # "model": "qwen2.5:1.5b",
    # "model": "llama3.2:1b",
    # "model": "deepseek_r1:1.5b",

    # LLMに送る会話内容を指定
    "messages": [
        {"role": "user", "content": question}
    ]
}

# 質問内容を表示
print(f"質問: {question}\n")

# 思考開始時刻を記録
think_start_time = time.time()
print("思考開始\n")

# LLM APIにPOSTリクエストを送信（ストリーミング）
res = requests.post(url, json=payload, stream=True)

answer_started = False        # 回答開始検知フラグ
final_answer = ""             # 回答全文を格納する変数

for line in res.iter_lines():
    if not line:
        continue

    data = json.loads(line.decode("utf-8"))
    content = data.get("message", {}).get("content", "")

    # 回答が初めて返ってきた瞬間
    if content and not answer_started:
        print("回答開始\n")
        answer_started = True

    # 回答内容を逐次表示しつつ、変数に蓄積
    if content:
        print(content, end="", flush=True)
        final_answer += content

    # 回答終了を検知
    if data.get("done"):
        elapsed = time.time() - think_start_time
        print(f"\n\n処理時間: {elapsed:.2f}秒")
        break

hailortcli scan

git clone https://github.com/hailo-ai/hailo-apps.git

cd hailo-apps

python3 -m venv venv --system-site-packages

source venv/bin/activate

sudo apt update
sudo apt install -y portaudio19-dev

pip install -e ".[gen-ai]"

sudo mkdir -p /usr/local/hailo

sudo chown -R pi:pi /usr/local/hailo

python -m hailo_apps.python.gen_ai_apps.simple_vlm_chat.simple_vlm_chat

python -m hailo_apps.python.gen_ai_apps.vlm_chat.vlm_chat --input rpi

import subprocess
import time

# エラー回避のための強制開放（前回プロセスの残骸対策）
subprocess.run(
    "sudo fuser -k /dev/video* /dev/media* /dev/v4l-subdev*",
    shell=True,
    stdout=subprocess.DEVNULL,
    stderr=subprocess.DEVNULL
)
time.sleep(1)

import sys
import time
import signal
import os
import cv2
from typing import Optional

from hailo_apps.python.gen_ai_apps.vlm_chat.backend import Backend
from hailo_apps.python.core.common.core import (
    get_standalone_parser,
    handle_list_models_flag,
    resolve_hef_path,
)
from hailo_apps.python.core.common.defines import (
    VLM_CHAT_APP,
    HAILO10H_ARCH,
)
from hailo_platform import VDevice
from hailo_platform.genai import LLM

def translate_to_japanese(text: str) -> str:
    vdevice_llm = VDevice()
    llm = LLM(vdevice_llm, "/home/pi/models/Qwen2-1.5B-Instruct.hef")
    
    prompt = f"以下の英文を自然で正しい日本語の表現で翻訳してください。難しい漢字を使わないこと\n\n{text}"    
    #print(prompt)

    messages = [
        {"role": "user", "content": prompt}
    ]

    result = ""
    for chunk in llm.generate(messages):
        result += chunk

    llm.release()
    return result.replace("<|im_end|>", "").strip()

# =========================
# 設定
# =========================
WAIT_SECONDS = 3  # 撮影までの待機時間

QUESTION = """
Look at the camera image carefully.
Describe what is happening in the image.
Be concise and clear.
Mention the main subject and important details.
"""
MAX_TOKENS = 30
TEMPERATURE = 0.1
SEED = 42

# 画像説明用のシステムプロンプト
SYSTEM_PROMPT = """
You see images from a camera.
You describe the image clearly and objectively.
Focus on what is visible.
"""
INFERENCE_TIMEOUT = 60
WINDOW_NAME = "Captured Frame"

def extract_model_name(hef_path: str) -> str:
    # hefファイル名からモデル名を取得
    return os.path.splitext(os.path.basename(hef_path))[0]

class OneShotVLMApp:
    def __init__(self, hef_path: str):
        self.hef_path = hef_path
        self.backend: Optional[Backend] = None
        self.running = True
        signal.signal(signal.SIGINT, self._signal_handler)

    def _signal_handler(self, sig, frame):
        self.running = False

    def _init_camera(self):
        from picamera2 import Picamera2
        from libcamera import controls

        picam2 = Picamera2()
        config = picam2.create_preview_configuration(
            main={"size": (320, 240), "format": "RGB888"}
        )
        picam2.configure(config)
        picam2.start()
        
        # オートフォーカス設定
        picam2.set_controls({
            "AfMode": controls.AfModeEnum.Continuous,
            "AfSpeed": controls.AfSpeedEnum.Fast,
        })
        return picam2

    def _init_backend(self):
        # Hailo VLMバックエンド初期化
        self.backend = Backend(
            hef_path=self.hef_path,
            max_tokens=MAX_TOKENS,
            temperature=TEMPERATURE,
            seed=SEED,
            system_prompt=SYSTEM_PROMPT,
        )

    def run(self):
        
        picam2 = self._init_camera()
        self._init_backend()

        cv2.namedWindow(WINDOW_NAME, cv2.WINDOW_AUTOSIZE)

        # ---------- ライブ表示 ----------
        while self.running:

            start_time = time.time()
            captured_frame = None
            
            while self.running and captured_frame is None:
                frame = picam2.capture_array()
                rgb_frame = Backend.convert_resize_image(frame)
                bgr_frame = cv2.cvtColor(rgb_frame, cv2.COLOR_RGB2BGR)

                cv2.imshow(WINDOW_NAME, bgr_frame)
                
                # 指定時間経過で撮影
                if (time.time() - start_time) >= WAIT_SECONDS:
                    captured_frame = cv2.cvtColor(rgb_frame, cv2.COLOR_RGB2BGR)
                    break

                if cv2.waitKey(1) & 0xFF == ord("q"):
                    self.running = False
                    break
                
            if not self.running:
                break
        
            loop_start = time.perf_counter()
            t0 = time.perf_counter()
                    
            # VLM推論を実行（画像→テキスト生成）
            result = self.backend.vlm_inference(
                captured_frame,
                QUESTION,
                INFERENCE_TIMEOUT,
            )
            t1 = time.perf_counter()
            inference_time = t1 - t0
            
            answer_text = result.get("answer", "回答を取得できませんでした")

            model_name = extract_model_name(self.hef_path)

            print("\n=====================================\n")
            print("---------- 実行情報 ----------")
            print(f"モデル名   ：{model_name}")
            print(f"VLM推論時間   ：{inference_time:.2f}秒")
            print("------------------------------\n")
                    
            # Hailoデバイス解放
            self.backend.close()
                             
            # 翻訳 -----------------------------------------------
            t2 = time.perf_counter()

            translated_text = translate_to_japanese(answer_text)
            print(translated_text)
            
            t3 = time.perf_counter()
            llm_time = t3 - t2            
                    
            total_time = time.perf_counter() - loop_start # トータル時間計算
                    
            print("\n---------- 処理時間 ----------")
            print(f"LLM推論時間   ：{llm_time:.2f}秒")
            print(f"トータル時間 ：{total_time:.2f}秒")
            print("------------------------------\n")

            self._init_backend()
            time.sleep(1)

            if cv2.waitKey(30) & 0xFF == ord("q"):
                break

if __name__ == "__main__":
    parser = get_standalone_parser()
    handle_list_models_flag(parser, VLM_CHAT_APP)
    options = parser.parse_args()

    hef_path = resolve_hef_path(
        options.hef_path if hasattr(options, "hef_path") else None,
        app_name=VLM_CHAT_APP,
        arch=HAILO10H_ARCH,
    )

    if hef_path is None:
        sys.exit(1)

    OneShotVLMApp(str(hef_path)).run()
    sys.exit(0)

git clone https://github.com/PiSugar/Whisplay.git --depth 1
cd Whisplay/Driver
sudo bash install_wm8960_drive.sh
sudo reboot

hailo-audio-troubleshoot --install-tts

python -c "from hailo_apps.python.gen_ai_apps.gen_ai_utils.voice_processing.text_to_speech import TextToSpeechProcessor as T; t=T(); t.queue_text('Hello, how are you today?'); t.wait_for_completion()"

sudo apt install -y open-jtalk open-jtalk-mecab-naist-jdic hts-voice-nitech-jp-atr503-m001

sudo apt install -y open-jtalk-mecab-naist-jdic hts-voice-nitech-jp-atr503-m001

echo "こんにちは" | open_jtalk -x /var/lib/mecab/dic/open-jtalk/naist-jdic -m /usr/share/hts-voice/nitech-jp-atr503-m001/nitech_jp_atr503_m001.htsvoice -ow /tmp/test.wav && aplay /tmp/test.wav

wget https://downloads.sourceforge.net/project/mmdagent/MMDAgent_Example/MMDAgent_Example-1.8/MMDAgent_Example-1.8.zip

unzip MMDAgent_Example-1.8.zip

sudo cp -r MMDAgent_Example-1.8/Voice/mei /usr/share/hts-voice/

import subprocess

def speak_japanese(text, voice="normal", speed=0.8):
    # voiceで選択できる種類
    # "normal"   : 標準
    # "happy"    : 明るい
    # "angry"    : 怒り
    # "sad"      : 悲しい
    # "bashful"  : 照れ
    voice_path = f"/usr/share/hts-voice/mei/mei_{voice}.htsvoice"

    subprocess.run(
        [
            "open_jtalk",
            "-x", "/var/lib/mecab/dic/open-jtalk/naist-jdic",
            "-m", voice_path,
            "-r", str(speed),
            "-ow", "/tmp/test.wav"
        ],
        input=text,
        text=True
    )

    subprocess.run(["aplay", "/tmp/test.wav"])

speak_japanese("こんにちは", "happy")

cd hailo-apps && source venv/bin/activate

python -m hailo_apps.python.gen_ai_apps.voice_assistant.voice_assistant

mkdir -p /home/pi/models

mv /home/pi/ダウンロード/Qwen2-1.5B-Instruct.hef /home/pi/models/

LLM_PROMPT_PREFIX = "日本語で答えてください。\n"

nano ~/hailo-apps/hailo_apps/python/gen_ai_apps/voice_assistant/tts_openjtalk.py

import subprocess

def speak_japanese(text, voice="normal", speed=0.8):
    voice_path = f"/usr/share/hts-voice/mei/mei_{voice}.htsvoice"

    subprocess.run(
        [
            "open_jtalk",
            "-x", "/var/lib/mecab/dic/open-jtalk/naist-jdic",
            "-m", voice_path,
            "-r", str(speed),
            "-ow", "/tmp/test.wav"
        ],
        input=text,
        text=True
    )

    subprocess.run(
        ["aplay", "/tmp/test.wav"],
        stdout=subprocess.DEVNULL,
        stderr=subprocess.DEVNULL
    )

from .tts_openjtalk import speak_japanese

def collect_callback(chunk: str):
    nonlocal response_text
    response_text += chunk

python -m hailo_apps.python.gen_ai_apps.voice_assistant.voice_assistant --no-tts