Jetson Container 使用紀錄，二

參考 Jetson Container 使用紀錄

因為這次更新到 Jetpack 6.0, 所以測試得很順利

參考 https://jetsonhacks.com/2023/08/07/speech-ai-on-nvidia-jetson-tutorial/

註冊 NGC

生成 API 密鑰

$ ngc config set

$ ngc registry resource download-version nvidia/riva/riva_quickstart_arm64:2.17.0

$ cd riva_quickstart

$ vi config.sh

service_enabled_nlp=false

service_enabled_nmt=false

asr_language_code=("en-US")

#asr_language_code=("zh-CN")

tts_language_code=("en-US")

#tts_language_code=("zh-CN")

$ sudo bash riva_init.sh

$ sudo bash riva_start.sh

安裝 https://github.com/nvidia-riva/python-clients.git

### 文字介面

$ jetson-containers run --env HUGGINGFACE_TOKEN=hf_abc123def \

  $(./autotag nano_llm) \

  python3 -m nano_llm.chat --api=mlc \

    --model meta-llama/Llama-2-7b-chat-hf \

    --quantization q4f16_ft

### llamaspeak

$ jetson-containers run --env HUGGINGFACE_TOKEN=hf_xyz123abc456 \

  $(autotag nano_llm) \

  python3 -m nano_llm.agents.web_chat --api=mlc \

    --model meta-llama/Meta-Llama-3-8B-Instruct \

    --asr=riva --language-code=en-US --tts=piper \

    --voice=en_US-libritts-high --web-port 8050

$ jetson-containers run --env HUGGINGFACE_TOKEN=hf_xyz123abc456 \

  $(autotag nano_llm) \

  python3 -m nano_llm.agents.web_chat --api=mlc \

    --model meta-llama/Meta-Llama-3-8B-Instruct \

    --asr=riva --language-code=zh-CN --tts=piper \

    --voice=zh_CN-huayan-x_low --web-port 8050

參考 https://ithelp.ithome.com.tw/articles/10347557

輸入的中文識別：修改/opt/NanoLLM/nano_llm/plugins/speech/riva_asr.py，將裡面的language_code=從en-US改成zh-CN

輸出的中文語音：修改/opt/NanoLLM/nano_llm/plugins/speech/piper_tts.py，將裡面的en_US-libritts-high模型（有兩處）改成zh_CN-huayan-medium模型。這裡可以選擇的中文模型，可以在/data/models/piper/voices.json裡找到，請用關鍵字zh-CN搜索，會找到zh_CN-huayan-medium與zh_CN-huayan-x_low兩組可用的中文語言

還須設定 ws-port(49000) 的 nat port 轉換

can you describe the image?

##Multimodality

$ jetson-containers run $(autotag nano_llm) \

  python3 -m nano_llm.agents.web_chat --api=mlc \

    --model Efficient-Large-Model/VILA-7b \

    --asr=riva --tts=piper

### Agent Studio

# python3 -m nano_llm.studio

https://IP_ADDRESS:8050

### text-generation-webui

$ jetson-containers run $(autotag text-generation-webui)

http://<IP_ADDRESS>:7860

中文模型

ckip-joint_bloom-3b-zh

參考  https://www.atyun.com/models/info/ckip-joint/bloom-3b-zh.html

taide_Llama3-TAIDE-LX-8B-Chat-Alpha1-4bit

參考  https://taide.tw/index

yentinglin_Llama-3-Taiwan-8B-Instruct

參考  https://medium.com/@simon3458/project-tame-llama-3-taiwan-1b249b88ab67

### ollama (Open WebUI)

$ jetson-containers run --name ollama $(autotag ollama)

# /bin/ollama run llama3

$ docker run -it --rm --network=host --add-host=host.docker.internal:host-gateway ghcr.io/open-webui/open-webui:main

http://JETSON_IP:8080

### Stable Diffusion

$ jetson-containers run $(autotag stable-diffusion-webui)

http://<IP_ADDRESS>:7860

txt2img: a person sitting at an office desk working

img2img: 先放一張圖

1. 右上 Generate 下方有 Interrogate CLIP 和 Interrogate DeepBooru 兩按鈕，可讀取圖片資訊

2. 可修改讀取出的資訊，按 Generate, 重新產生圖

3. 選 Resize and fill, 並配合不同的 Denising strength, 按 Generate, 重新產生圖

4. 選 Just resize, 並配合不同的 Denising strength, 按 Generate, 重新產生圖

5. Inpaint: 在 prompt 上填入 sunglasses, 將圖片上的眼睛位置塗黑, 按 Generate, 重新產生圖

### Stable Diffusion XL

$ CONTAINERS_DIR=`pwd`

$ MODEL_DIR=$CONTAINERS_DIR/data/models/stable-diffusion/models/Stable-diffusion/

$ sudo chown -R $USER $MODEL_DIR

$ wget -P $MODEL_DIR https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors

$ wget -P $MODEL_DIR https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/resolve/main/sd_xl_refiner_1.0.safetensors

左上角 Stable Diffusion checkpoint 選擇 sd_xl_base_1.0.safetensors

Generation tab 點選 Refiner 選擇 sd_xl_refiner_1.0.safetensors

改 width/height 到 1024

txt2img: photograph of a friendly robot alongside a person climbing a mountain 

### Jetson Copilot

git clone https://github.com/NVIDIA-AI-IOT/jetson-copilot/

cd jetson-copilot

./setup_environment.sh

./launch_jetson_copilot.sh

http://JETSON_IP:8501

$ virtualenv --system-site-packages tensorrt

$ docker ps -a

$ docker kill wqerqwer

模型優化之指派問題(linear sum assignment problem)

參考 匈牙利算法

python 使用函數

scipy.optimize.linear_sum_assignment

lap.lapjv

MMTracking 學習紀錄

參考 https://mmtracking.readthedocs.io/en/latest/index.html

參考 https://github.com/open-mmlab/mmtracking

$ git clone https://github.com/open-mmlab/mmtracking.git

$ docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:21.05-py3

$ docker run --gpus all -it --name MMTracking nvcr.io/nvidia/pytorch:21.05-py3

$ docker start MMTracking

$ docker attach MMTracking

# <ctrl+p><ctrl+q>

$ docker attach MMTracking

$ docker stop MMTracking

$ docker rm MMTracking

$ echo $DISPLAY

$ export DISPLAY=:0

$ xhost +

$ docker run --gpus all -it --name MMTracking --shm-size=8G \

  --net=host -e DISPLAY=$DISPLAY -e XAUTHORITY=/tmp/xauth \

  -e QT_X11_NO_MITSHM=1 \

  -v /tmp/.X11-unix/:/tmp/.X11-unix \

  -v ~/.Xauthority:/tmp/xauth \

  -v /etc/localtime:/etc/localtime \

  -v /mnt/Data/MMTracking/mmtracking:/workspace/mmtracking \

  nvcr.io/nvidia/pytorch:21.05-py3

# pip install git+https://github.com/votchallenge/toolkit.git

# nvidia-smi

CUDA Version: 12.2

# pip list

opencv-python                    4.10.0.84

torch                            1.9.0a0+2ecb2c7

# pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu122/torch1.9.0/index.html

#### 出現下列錯誤

####    cv.gapi.wip.GStreamerPipeline = cv.gapi_wip_gst_GStreamerPipeline

####AttributeError: partially initialized module 'cv2' has no attribute 'gapi_wip_gst_GStreamerPipeline' (most likely due to a circular import)

#### 解決

# pip install opencv-python==4.5.1.48

# pip install mmdet==2.28.2

# pip install mmengine==0.10.4

# pip list

mmcv-full                        1.7.2

mmdet                            2.28.2

mmengine                         0.10.4

pillow                           10.4.0

# cd /workspace/mmtracking/

# pip install -r requirements/build.txt

# pip install -v -e .

# pip install git+https://github.com/JonathonLuiten/TrackEval.git

# pip install git+https://github.com/lvis-dataset/lvis-api.git

# pip install git+https://github.com/TAO-Dataset/tao.git

# python demo/demo_mot_vis.py configs/mot/deepsort/deepsort_faster-rcnn_fpn_4e_mot17-private-half.py \

  --input demo/demo.mp4 --output mot.mp4 --show

#### 出現下列錯誤

####    from .cv2 import *

####ImportError: libGL.so.1: cannot open shared object file: No such file or directory

#### 解決

# apt-get update

# apt-get install libgl1

#### 出現下列錯誤

####    NumpyArray = npt.NDArray[Any]

####AttributeError: module 'numpy.typing' has no attribute 'NDArray'

#### 解決

# pip install Pillow==9.5.0

#### 出現下列錯誤

#### Could not load the Qt platform plugin "xcb" in "/opt/conda/lib/python3.8/site-packages/cv2/qt/plugins" even though it was found.

####This application failed to start because no Qt platform plugin could be initialized. Reinstalling the application may fix this problem.

#### 查詢錯誤所在

# export QT_DEBUG_PLUGINS=1

####Cannot load library /opt/conda/lib/python3.8/site-packages/cv2/qt/plugins/platforms/libqxcb.so: (libSM.so.6: cannot open shared object file: No such file or directory)

#### 解決

# apt-get install -y libsm6 libxext6 libxrender-dev

# pip install opencv-contrib-python==4.5.1.48

# wget https://download.openmmlab.com/mmtracking/mot/ocsort/mot_dataset/ocsort_yolox_x_crowdhuman_mot17-private-half_20220813_101618-fe150582.pth

# python demo/demo_mot_vis.py configs/mot/ocsort/ocsort_yolox_x_crowdhuman_mot17-private-half.py \

  --checkpoint ocsort_yolox_x_crowdhuman_mot17-private-half_20220813_101618-fe150582.pth \

  --input demo/demo.mp4 --output mot.mp4 --show

# wget https://download.openmmlab.com/mmtracking/vid/selsa/selsa_faster_rcnn_r101_dc5_1x_imagenetvid/selsa_faster_rcnn_r101_dc5_1x_imagenetvid_20201218_172724-aa961bcc.pth

# python demo/demo_vid.py configs/vid/selsa/selsa_faster_rcnn_r101_dc5_1x_imagenetvid.py \

  --checkpoint selsa_faster_rcnn_r101_dc5_1x_imagenetvid_20201218_172724-aa961bcc.pth \

  --input demo/demo.mp4 --show

# python demo/demo_sot.py configs/sot/siamese_rpn/siamese_rpn_r50_20e_lasot.py \

  --input demo/demo.mp4 --output sot.mp4 --show

#### 用滑鼠框出要追蹤的物件，按空白鍵

MMOCR 學習紀錄

參考 https://mmocr.readthedocs.io/en/dev-1.x/

參考 https://github.com/open-mmlab/mmocr

$ docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:21.05-py3

$ docker run --gpus all -it --name MMOCR nvcr.io/nvidia/pytorch:21.05-py3

$ docker start MMOCR

$ docker attach MMOCR

# <ctrl+p><ctrl+q>

$ docker attach MMOCR

$ docker stop MMOCR

$ docker rm MMOCR

$ docker run --gpus all -it --name MMOCR --shm-size=8G \

  -v /mnt/Data/MMOCR/mmocr:/workspace/mmocr \

  -v /mnt/QNAP_A/ImageData/ICDAR:/mmocr/data \

  nvcr.io/nvidia/pytorch:21.05-py3

# pip install -U openmim

#### 出現下列錯誤

####ERROR: After October 2020 you may experience errors when installing or updating packages. This is because pip will change the way that it resolves dependency conflicts.

####We recommend you use --use-feature=2020-resolver to test your packages with the new resolver before it becomes the default.

#### 不用擔心

確認 https://mmocr.readthedocs.io/en/dev-1.x/get_started/install.html 底部的版本資訊

安裝正確版本套件

# mim list

# mim install mmengine==

# mim install mmengine

# mim install mmcv==2.0.1

#### 出現下列錯誤

####    cv.gapi.wip.GStreamerPipeline = cv.gapi_wip_gst_GStreamerPipeline

####AttributeError: partially initialized module 'cv2' has no attribute 'gapi_wip_gst_GStreamerPipeline' (most likely due to a circular import)

#### 解決

#### 因為目前 opencv-python 版本為 4.10.0.84, 降版本

# pip install opencv-python==4.5.1.48

# mim install mmcv==2.0.1

# mim install mmdet==3.1.0

# cd /workspace/mmocr/

# pip install -v -e .

# pip install opencv-python-headless==4.5.1.48

# pip install -r requirements/albu.txt

# pip install -r requirements.txt

# python tools/infer.py demo/images --det DBNet --rec CRNN --print-result \

  --save_pred --save_vis --out-dir='results/' --batch-size=2

SwinTransformer 學習紀錄

參考 https://github.com/microsoft/Swin-Transformer?tab=readme-ov-file

參考 https://github.com/SwinTransformer/Swin-Transformer-Object-Detection

參考 https://mmdetection.readthedocs.io/en/latest/get_started.html

參考 https://github.com/open-mmlab/mmdetection/blob/master/docs/en/get_started.md

參考 https://github.com/open-mmlab/mmcv

參考 https://github.com/open-mmlab/mim?tab=readme-ov-file

Swin-Transformer 是原始版本，主要針對圖片分類

Swin-Transformer-Object-Detection 主要針對物件偵測(基於 mmdetection)

mmdetection 主要針對物件偵測，其中不只是 Swin-Transformer, 包含各式各樣先進的演算法

mmdetection 可在 mmdet/__init__.py 可以查詢到 mmcv 所需要的版本

Swin-Transformer-Object-Detection 要求使用 mmcv(1.4.0)

但 mmdetection 最後對應到 mmcv(1.4.0) 的版本為(v2.18.1)

所以以下放棄使用 Swin-Transformer-Object-Detection, 直接使用 mmdetection 的最新版本

$ docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:21.05-py3

$ docker run --gpus all -it --name SwinTransformer nvcr.io/nvidia/pytorch:21.05-py3

$ docker start SwinTransformer

$ docker attach SwinTransformer

# <ctrl+p><ctrl+q>

$ docker attach SwinTransformer

$ docker stop SwinTransformer

$ docker rm SwinTransformer

$ git clone https://github.com/microsoft/Swin-Transformer.git

$ git clone https://github.com/SwinTransformer/Swin-Transformer-Object-Detection.git

$ git clone https://github.com/open-mmlab/mmdetection.git

$ docker run --gpus all -it --name SwinTransformer --shm-size=8G \

  -v /mnt/Data/SwinTransformer/Swin-Transformer:/workspace/Swin-Transformer \

  -v /mnt/Data/SwinTransformer/mmdetection:/workspace/mmdetection \

  -v /mnt/Data/SwinTransformer/Swin-Transformer-Object-Detection:/workspace/Swin-Transformer-Object-Detection \

  -v /mnt/QNAP_A/ImageData/ImageNet:/workspace/ImageNet \

  nvcr.io/nvidia/pytorch:21.05-py3

# pip install timm==0.4.12

# pip install opencv-python==4.4.0.46 termcolor==1.1.0 yacs==0.1.8 pyyaml scipy

# cd /workspace/Swin-Transformer/kernels/window_process/

# python setup.py install

# cd /workspace/Swin-Transformer/

# wget https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth

# python -m torch.distributed.launch --nproc_per_node 1 --master_port 12345 main.py \

  --eval --cfg configs/swin/swin_tiny_patch4_window7_224.yaml \

  --resume swin_tiny_patch4_window7_224.pth --data-path /workspace/ImageNet \

  --batch_size=64

#### 出現下列錯誤

####  File "/opt/conda/lib/python3.8/site-packages/PIL/_typing.py", line 10, in <module>

####    NumpyArray = npt.NDArray[Any]

####AttributeError: module 'numpy.typing' has no attribute 'NDArray'

#### 解決

# pip install Pillow==9.5.0

#### 出現下列錯誤

####RuntimeError: Found 0 files in subfolders of: /workspace/ImageNet/val

####Supported extensions are: .jpg,.jpeg,.png,.ppm,.bmp,.pgm,.tif,.tiff,.webp

#### 解決

$ cd /mnt/QNAP_A/ImageData/ImageNet/

$ mv val val_a ;mkdir val; mv val_a val

#### 出現下列錯誤

ERROR: Unexpected bus error encountered in worker. This might be caused by insufficient shared memory (shm).

#### 解決 docker run 時須加上 --shm-size=8G

# cd /workspace/mmdetection

# pip install -U openmim

#### 出現下列錯誤

####ERROR: After October 2020 you may experience errors when installing or updating packages. This is because pip will change the way that it resolves dependency conflicts.

####We recommend you use --use-feature=2020-resolver to test your packages with the new resolver before it becomes the default.

#### 不用擔心

# mim list

# mim install mmengine

# mim install "mmcv>=2.0.0"

# pip install -v -e .

    # "-v" means verbose, or more output

    # "-e" means installing a project in editable mode,

    # thus any local modifications made to the code will take effect without reinstallation.

# mim download mmdet --config yolov3_mobilenetv2_8xb24-320-300e_coco --dest .

# mim list

# ls

# python demo/image_demo.py demo/demo.jpg yolov3_mobilenetv2_8xb24-320-300e_coco.py \

  --weights yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth

#### 出現下列錯誤

####    import cv2

####  File "/opt/conda/lib/python3.8/site-packages/cv2/__init__.py", line 5, in <module>

####    from .cv2 import *

####ImportError: libGL.so.1: cannot open shared object file: No such file or directory

#### 出現下列錯誤

####    cv.gapi.wip.GStreamerPipeline = cv.gapi_wip_gst_GStreamerPipeline

####AttributeError: partially initialized module 'cv2' has no attribute 'gapi_wip_gst_GStreamerPipeline' (most likely due to a circular import)

#### 解決

# pip install opencv-python-headless==4.4.0.46

#### 出現下列錯誤

####    assert (mmcv_version >= digit_version(mmcv_minimum_version)

####AssertionError: MMCV==2.2.0 is used but incompatible. Please install mmcv>=2.0.0rc4, <2.2.0.

#### 解決

# mim install "mmcv==2.0.0rc4"

# mim download mmdet --config mask-rcnn_swin-t-p4-w7_fpn_1x_coco --dest .

# python demo/image_demo.py demo/demo.jpg mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py \

  --weights mask_rcnn_swin-t-p4-w7_fpn_1x_coco_20210902_120937-9d6b7cfa.pth

Jetson Container 使用紀錄

mic-733ao@ubuntu:~/Data/AgentStudio/jetson-containers$ ./run.sh --workdir=/opt/text-generation-webui $(./autotag text-generation-webui) python3 server.py --model-dir=/data/models/text-generation-webui --listen --verbose --trust-remote-code

mic-733ao@ubuntu:~/Data/AgentStudio/jetson-containers$ ./run.sh --workdir=/opt/text-generation-webui $(./autotag text-generation-webui) /bin/bash -c 'python3 download-model.py --output=/data/models/text-generation-webui/THUDM_cogvlm2-llama3-chinese-chat-19B-int4 THUDM/cogvlm2-llama3-chinese-chat-19B-int4'

產生授權錯誤

raise HTTPError(http_error_msg, response=self)

requests.exceptions.HTTPError: 401 Client Error: Unauthorized for url: https://huggingface.co/taide/TAIDE-LX-7B/resolve/main/README_en.md

到 https://huggingface.co/ 網站 右上角使用者/Settings/Access Tokens

建立 Fine-grained token, 打勾所有權限

$ export HUGGINGFACE_TOKEN=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

chrome://flags/#unsafely-treat-insecure-origin-as-secure

有用的 docker 命令

$ sudo docker exec -i -t container_name /bin/bash

$ docker cp container_id:/path /host/path