Nuclio 모델배포 : custom model

Published on 23-02-1723-02-17

모델배포 : custom model – mmdetection

0. inference demo 확인

main.py의 handler작성을 위해 inference demo확인 ( init_detector, inference_detector, [show_result_pyplot])

demo/image_demo.py —[copy]—>demo_img.py

# ln -s /raid/templates/cvat cvat   (mmdet안에 링크)
# ~/my/git/mmdetection/cvat/        
# ~/my/git/mmdetection/cvat/serverless

# pwd
# ~/my/git/mmdetection/
$ python demo_img.py \\
    demo/demo.jpg \\
    configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \\
    checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \\
    --out-file demo/out/demo_out.jpg \\
    --score-thr 0.4 \\
    --device cuda:6

Config파일 만들기 (work_dirs/faster_rcnn/faster_rcnn.py)
CheckPoint파일 확인 (work_dirs/faster_rcnn/lastest.pth)

1. MakeConfig

$ python sixxtools/makeConfig_sixx.py \\
 --fromconfig configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \\
 --toconfig   sixxconfigs/faster_rcnn_r50_fpn_1x_coco_001.py

https://onesixx.com/mmdet-train/ 참고

Dataset 수집/구성/> Model 선정/checkpoint > config 수정 > Training

2. dockerfile

$ docker build -t base.me  cvat/serverless/sixx/mmdet/faster_rcnn/nuclio/

~/my/git/mmdetection/docker$ docker build -t base.mmdet .

FROM ubuntu:20.04FROM ubuntu:20.04

ENV TZ=Asiz/Seoul
RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone
ENV PATH /opt/conda/bin:$PATH

RUN apt update
RUN apt install -y python3.8 \\
    && ln -s /usr/bin/python3.8 /usr/bin/python3
RUN apt install -y tzdata
RUN apt install -y --no-install-recommends w g e t git ninja-build ca-certificates libglib2.0-0 libsm6 libxrender-dev libxrender1 libxext6 python3.8-dev build-essential ffmpeg
# Install miniconda3
RUN w g e t --quiet https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda.sh && \\
    /bin/bash ~/miniconda.sh -b -p /opt/conda && \\
    rm ~/miniconda.sh && \\
    /opt/conda/bin/conda clean -tipy && \\
    ln -s /opt/conda/etc/profile.d/conda.sh /etc/profile.d/conda.sh && \\
    echo ". /opt/conda/etc/profile.d/conda.sh" >> ~/.bashrc && \\
    echo "conda activate my_env" >> ~/.bashrc

RUN conda create -n mmdet380 python=3.8
SHELL ["conda", "run", "-n", "mmdet380", "/bin/bash", "-c"]
RUN conda install -y pytorch torchvision -c pytorch
RUN conda install -y pip
RUN pip install --upgrade pip setuptools wheel
RUN pip install pycocotools
RUN pip install openmim
RUN mim install mmdet
ENTRYPOINT ["conda", "run", "-n", "mmdet380"]

function.yaml

metadata:
  name: mmdet-faster-rcnn-x
  namespace: cvat
  annotations:
    name: faster-rcnn-x
    type: detector
    framework: pytorch
    spec: |
      [
        { "id": 1, "name": "person" },
        { "id": 2, "name": "bicycle" },
        { "id": 3, "name": "car" },
        { "id": 4, "name": "motorcycle" },
        { "id": 5, "name": "airplane" },
        { "id": 6, "name": "bus" },
        { "id": 7, "name": "train" },
        { "id": 8, "name": "truck" },
        { "id": 9, "name": "boat" },
        { "id":10, "name": "traffic_light" },
        { "id":11, "name": "fire_hydrant" },
        { "id":13, "name": "stop_sign" },
        { "id":14, "name": "parking_meter" },
        { "id":15, "name": "bench" },
        { "id":16, "name": "bird" },
        { "id":17, "name": "cat" },
        { "id":18, "name": "dog" },
        { "id":19, "name": "horse" },
        { "id":20, "name": "sheep" },
        { "id":21, "name": "cow" },
        { "id":22, "name": "elephant" },
        { "id":23, "name": "bear" },
        { "id":24, "name": "zebra" },
        { "id":25, "name": "giraffe" },
        { "id":27, "name": "backpack" },
        { "id":28, "name": "umbrella" },
        { "id":31, "name": "handbag" },
        { "id":32, "name": "tie" },
        { "id":33, "name": "suitcase" },
        { "id":34, "name": "frisbee" },
        { "id":35, "name": "skis" },
        { "id":36, "name": "snowboard" },
        { "id":37, "name": "sports_ball" },
        { "id":38, "name": "kite" },
        { "id":39, "name": "baseball_bat" },
        { "id":40, "name": "baseball_glove" },
        { "id":41, "name": "skateboard" },
        { "id":42, "name": "surfboard" },
        { "id":43, "name": "tennis_racket" },
        { "id":44, "name": "bottle" },
        { "id":46, "name": "wine_glass" },
        { "id":47, "name": "cup" },
        { "id":48, "name": "fork" },
        { "id":49, "name": "knife" },
        { "id":50, "name": "spoon" },
        { "id":51, "name": "bowl" },
        { "id":52, "name": "banana" },
        { "id":53, "name": "apple" },
        { "id":54, "name": "sandwich" },
        { "id":55, "name": "orange" },
        { "id":56, "name": "broccoli" },
        { "id":57, "name": "carrot" },
        { "id":58, "name": "hot_dog" },
        { "id":59, "name": "pizza" },
        { "id":60, "name": "donut" },
        { "id":61, "name": "cake" },
        { "id":62, "name": "chair" },
        { "id":63, "name": "couch" },
        { "id":64, "name": "potted_plant" },
        { "id":65, "name": "bed" },
        { "id":67, "name": "dining_table" },
        { "id":70, "name": "toilet" },
        { "id":72, "name": "tv" },
        { "id":73, "name": "laptop" },
        { "id":74, "name": "mouse" },
        { "id":75, "name": "remote" },
        { "id":76, "name": "keyboard" },
        { "id":77, "name": "cell_phone" },
        { "id":78, "name": "microwave" },
        { "id":79, "name": "oven" },
        { "id":80, "name": "toaster" },
        { "id":81, "name": "sink" },
        { "id":83, "name": "refrigerator" },
        { "id":84, "name": "book" },
        { "id":85, "name": "clock" },
        { "id":86, "name": "vase" },
        { "id":87, "name": "scissors" },
        { "id":88, "name": "teddy_bear" },
        { "id":89, "name": "hair_drier" },
        { "id":90, "name": "toothbrush" }
      ]
 
spec:
  description: faster-rcnn-x
  runtime: "python:3.8"
  handler: main:handler
  eventTimeout: 30s
  build:
    image: cvat/sixx.mm.fast
    baseImage: makeimg #competent_dubinsky #mytag  #base.bkseo
 
    directives:
      preCopy:
        - kind: USER
          value: root
        - kind: WORKDIR
          value: /opt/nuclio
        - kind: ENV
          value: PATH="/root/miniconda3/bin:${PATH}"
        - kind: ARG
          value: PATH="/root/miniconda3/bin:${PATH}"
        - kind: ENTRYPOINT
          value: '["conda", "run", "-n", "mmdet380"]'
  triggers:
    myHttpTrigger:
      maxWorkers: 2
      kind: "http"
      workerAvailabilityTimeoutMilliseconds: 10000
      attributes:
        maxRequestBodySize: 33554432 # 32MB
 
  platform:
    attributes:
      restartPolicy:
        name: always
        maximumRetryCount: 3
      mountMode: volume

외부에서 해당 모델을 통해 접근을 위한 설정

nuclio function의 포트번호와 cvat network를 지정

spec.triggers.myHttpTrigger.attributes에 특정 port를 추가하고,
spec.platform.attributes 에 도커 컴포즈를 실행했을 때 생성된 도커 네트워크(cvat_cvat)를 추가.

  triggers:
    myHttpTrigger:
      maxWorkers: 2
      kind: 'http'
      workerAvailabilityTimeoutMilliseconds: 10000
      attributes:
        maxRequestBodySize: 33554432
        port: 33000 # 추가포트 

  platform:
    attributes:
      restartPolicy:
        name: always
        maximumRetryCount: 3
      mountMode: volume
      network: cvat_cvat # 추가 도커네트웍

Nuclio에 적용

2. 로컬에서 DL모델을 실행하기 위한 소스코드를 Nuclio 플랫폼에 적용

2-1 모델을 메모리에 로딩 (init_context(context)함수를 사용하여)

https://nuclio.io/docs/latest/concepts/best-practices-and-common-pitfalls/#use-init_context-instead-of-global-variable-declarations-or-function-calls

def init_context(context):
    # Create the DB connection under "context.user_data"
    # setattr(context.user_data, 'my_db_connection', my_db.create_connection())

    cfg = get_config('COCO-Detection/retinanet_R_101_FPN_3x.yaml')
    cfg.merge_from_list(CONFIG_OPTS)
    
    cfg.MODEL.RETINANET.SCORE_THRESH_TEST = CONFIDENCE_THRESHOLD
    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = CONFIDENCE_THRESHOLD
    cfg.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = CONFIDENCE_THRESHOLD
    
    cfg.freeze()
    predictor = DefaultPredictor(cfg)
    
    context.user_data.model_handler = predictor
    
def handler(context, event):
    # context.user_data.my_db_connection.query(...)

main.py

from inference demo

# Copyright onesixx. All rights reserved.
import os 
CONFIG = [os.path.join('./param', filenm) for filenm in os.listdir('./param') if filenm.endswith('.py')][0]
CHKPNT = [os.path.join('./param', filenm) for filenm in os.listdir('./param') if filenm.endswith('.pth')][0] 
SCORE_THRESHOLD = 0.9
DEVICE  = "cpu" #"cuda:6"
PALETTE = "coco"

import json
with open('./annotations/instances_train2017.json', 'r') as f:
    data = json.load(f)
categories = data['categories']

category_list = []
for category in categories:
    category_list.append({'id':category['id'], 'name':category['name']})

classes = category_list
# classes = [
#     { "id": 1, "name": "person" },
#     { "id": 2, "name": "bicycle" },
# ...
#     { "id":90, "name": "toothbrush" }
# ]
import mmcv
from mmdet.apis import init_detector
from mmdet.apis import inference_detector

import io
import base64
from PIL import Image
import numpy as np

def init_context(context):
    context.logger.info("Init context... 0%") # --------------------------------
    model = init_detector(CONFIG, CHECKPOINT, device=DEVICE)
    context.user_data.model = model
    context.logger.info("Init context... 100%") # ------------------------------

def handler(context, event):
    context.logger.info("Run sixx model")
    data = event.body
    buf = io.BytesIO(base64.b64decode(data["image"]))

    # threshold = float(data.get("threshold", SCORE_THRESHOLD))
    # context.user_data.model.conf = threshold

    image = Image.open(buf)
    imgArray = np.array(image)

    result = inference_detector(context.user_data.model, imgArray)

    if isinstance(result, tuple):
        bbox_result, segm_result = result
        if isinstance(segm_result, tuple):
            segm_result = segm_result[0]  # discard the `dim`
    else:
        bbox_result, segm_result = result, None

    img = image.copy()
    bboxes = np.vstack(bbox_result)
    labels = [
        np.full(bbox.shape[0], i, dtype=np.int32)
        for i, bbox in enumerate(bbox_result)
    ]
    labels = np.concatenate(labels)
    scores = bboxes[:, -1]

    inds = scores > SCORE_THRESHOLD
    scores = scores[inds]    # confidence
    labels = labels[inds]    # label
    bboxes = bboxes[inds, :] # points

    if show_mask and segm_result is not None:
        segms = mmcv.concat_list(segm_result)
        segms = [segms[i] for i in np.where(inds)[0]]
        if palette is None:
            palette = color_val_iter()
        colors = [next(palette) for _ in range(len(segms))]

    encoded_results = []
    if bboxes.shape[0] > 0:
        for i in range(bboxes.shape[0]):
            encoded_results.append({
                'confidence': scores[i],
                'label': classes[labels[i]]['name'],
                'points': bboxes[i].tolist(),
                'type': 'rectangle'
            })
    return context.Response(
        body=json.dumps(encoded_results),
        headers={},
        content_type='application/json',
        status_code=200
    )

2-2 아래 프로세스를 위해 handler에 entry point를 정의하고, main.py에 넣는다.

accept incoming HTTP requests
run inference
reply with detection results

main.py

# Copyright onesixx. All rights reserved.
import os 
# CONFIG = [os.path.join("./param", filenm) for filenm in os.listdir('./param') if filenm.endswith('.py')][0]
# CHKPNT = [os.path.join("./param", filenm) for filenm in os.listdir('./param') if filenm.endswith('.pth')][0] 

CONFIG = '/opt/nuclio/param/faster_rcnn_r50_fpn_1x_coco_001.py'
CHKPNT = '/opt/nuclio/param/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'
SCORE_THRESHOLD = 0.9
DEVICE  = "cpu" #"cuda:6"
PALETTE = "coco"

import mmcv
from mmdet.apis import init_detector
from mmdet.apis import inference_detector

import io
import base64
from PIL import Image
import numpy as np

import yaml    
import json
#from model_loader import ModelLoader

with open("/opt/nuclio/function.yaml", 'rb') as function_file:
    functionconfig = yaml.safe_load(function_file)
# with open("./function.yaml", 'rb') as function_file:     functionconfig = yaml.safe_load(function_file)
labels_spec = functionconfig['metadata']['annotations']['spec']
classes = eval(labels_spec)
# classes = [
#     { "id": 1, "name": "person" },
#     { "id": 2, "name": "bicycle" },
# ...
#     { "id":90, "name": "toothbrush" }
# ]
def init_context(context):
    context.logger.info("Init context... 0%") # --------------------------------

    model_handler = init_detector(CONFIG, CHKPNT, device=DEVICE)
    #model_handler = ModelLoader(classes) 
    context.user_data.model = model_handler

    context.logger.info("Init context... 100%") # ------------------------------

def handler(context, event):
    context.logger.info("Run sixx model")
    data = event.body
    buf = io.BytesIO(base64.b64decode(data["image"]))
    # threshold = float(data.get("threshold", SCORE_THRESHOLD))
    # context.user_data.model.conf = threshold
    
    # buf = './imgs/demo.jpg'
    image = Image.open(buf)
    imgArray = np.array(image)
    # result = inference_detector(model_handler, imgArray)
    result = inference_detector(context.user_data.model, imgArray)

    if isinstance(result, tuple):
        bbox_result, segm_result = result
        if isinstance(segm_result, tuple):
            segm_result = segm_result[0]  # discard the `dim`
    else:
        bbox_result, segm_result = result, None

    img = image.copy()
    bboxes = np.vstack(bbox_result)
    labels = [
        np.full(bbox.shape[0], i, dtype=np.int32)
        for i, bbox in enumerate(bbox_result)
    ]
    labels = np.concatenate(labels)
    scores = bboxes[:, -1]

    inds = scores > SCORE_THRESHOLD
    scores = scores[inds]    # confidence
    labels = labels[inds]    # label
    bboxes = bboxes[inds, :] # points

    # if show_mask and segm_result is not None:
    #     segms = mmcv.concat_list(segm_result)
    #     segms = [segms[i] for i in np.where(inds)[0]]
    #     if palette is None:
    #         palette = color_val_iter()
    #     colors = [next(palette) for _ in range(len(segms))]

    encoded_results = []
    if bboxes.shape[0] > 0:
        for i in range(bboxes.shape[0]):
            encoded_results.append({
                'confidence': float(scores[i]),
                'label': classes[labels[i]]['name'],
                'points': bboxes[i][:4].tolist(),
                'type': 'rectangle'
            })
    return context.Response(
        body=json.dumps(encoded_results),
        headers={},
        content_type='application/json',
        status_code=200
    )

# Copyright onesixx. All rights reserved.
import os 
# CONFIG = [os.path.join("./param", filenm) for filenm in os.listdir('./param') if filenm.endswith('.py')][0]
# CHKPNT = [os.path.join("./param", filenm) for filenm in os.listdir('./param') if filenm.endswith('.pth')][0] 

CONFIG = '/opt/nuclio/param/faster_rcnn_r50_fpn_1x_coco_001.py'
CHKPNT = '/opt/nuclio/param/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'
SCORE_THRESHOLD = 0.9
DEVICE  = "cpu" #"cuda:6"
PALETTE = "coco"

import mmcv
from mmdet.apis import init_detector
from mmdet.apis import inference_detector

import io
import base64
from PIL import Image
import numpy as np

import yaml    
import json
#from model_loader import ModelLoader

with open("/opt/nuclio/function.yaml", 'rb') as function_file:
    functionconfig = yaml.safe_load(function_file)
# with open("./function.yaml", 'rb') as function_file:     functionconfig = yaml.safe_load(function_file)
labels_spec = functionconfig['metadata']['annotations']['spec']
classes = labels_spec
# classes = [
#     { "id": 1, "name": "person" },
#     { "id": 2, "name": "bicycle" },
# ...
#     { "id":90, "name": "toothbrush" }
# ]
def init_context(context):
    context.logger.info("Init context... 0%") # --------------------------------

    model_handler = init_detector(CONFIG, CHKPNT, device=DEVICE)
    #model_handler = ModelLoader(classes) 
    context.user_data.model = model_handler

    context.logger.info("Init context... 100%") # ------------------------------

def handler(context, event):
    context.logger.info("Run sixx model")
    data = event.body
    buf = io.BytesIO(base64.b64decode(data["image"]))
    # threshold = float(data.get("threshold", SCORE_THRESHOLD))
    # context.user_data.model.conf = threshold
    
    # buf = './imgs/demo.jpg'
    image = Image.open(buf)
    imgArray = np.array(image)
    # result = inference_detector(model, imgArray)
    result = inference_detector(context.user_data.model, imgArray)

    if isinstance(result, tuple):
        bbox_result, segm_result = result
        if isinstance(segm_result, tuple):
            segm_result = segm_result[0]  # discard the `dim`
    else:
        bbox_result, segm_result = result, None

    img = image.copy()
    bboxes = np.vstack(bbox_result)
    labels = [
        np.full(bbox.shape[0], i, dtype=np.int32)
        for i, bbox in enumerate(bbox_result)
    ]
    labels = np.concatenate(labels)
    scores = bboxes[:, -1]

    inds = scores > SCORE_THRESHOLD
    scores = scores[inds]    # confidence
    labels = labels[inds]    # label
    bboxes = bboxes[inds, :] # points

    # if show_mask and segm_result is not None:
    #     segms = mmcv.concat_list(segm_result)
    #     segms = [segms[i] for i in np.where(inds)[0]]
    #     if palette is None:
    #         palette = color_val_iter()
    #     colors = [next(palette) for _ in range(len(segms))]

    encoded_results = []
    if bboxes.shape[0] > 0:
        for i in range(bboxes.shape[0]):
            encoded_results.append({
                'confidence': float(scores[i]),
                'label': classes[labels[i]]['name'],
                'points': bboxes[i][:4].tolist(),
                'type': 'rectangle'
            })
    return context.Response(
        body=json.dumps(encoded_results),
        headers={},
        content_type='application/json',
        status_code=200
    )

from model_handler import ModelHandler

3. deploy

새로운 serverless 함수를 사용하기 위해서는
(위 Builtin model에서 했던것처럼) nuctl 명령어로 deploy를 해야한다.

function.yaml
main.py
model_handler.py

방법1)

$ nuctl deploy --project-name cvat \\
  --path     cvat/serverless/sixx/mmdet/faster_rcnn_x/nuclio/ \\
  --volume   `pwd`/serverless/common:/opt/nuclio/common \\
  --platform local

방법2)

$ serverless/deploy_cpu.sh \\
serverless/pytorch/facebookresearch/detectron2/retinanet/

$ docker ps -a 

CONTAINER ID   IMAGE                                        COMMAND                  CREATED          STATUS                    PORTS                                                      NAMES
264a2739eeb6   cvat/sixx.mm.fast:latest                     "conda run -n mmdet3…"   47 seconds ago   Up 46 seconds (healthy)   0.0.0.0:49377->8080/tcp                                    nuclio-nuclio-mmdet-faster-rcnn-x
e71c29400634   gcr.io/iguazio/alpine:3.15                   "/bin/sh -c '/bin/sl…"   5 hours ago      Up 5 hours                                                                           nuclio-local-storage-reader

Issue

https://github.com/opencv/cvat/issues/3457 : Steps for custom model deployment

http://How to upload DL which built by myself? : How to upload DL which built by myself?

https://github.com/opencv/cvat/issues/5551 : Mmdetection MaskRCNN serverless support for semi-automatic annotation

https://github.com/opencv/cvat/issues/4909: Load my own Yolov5 model on cvat by nuclio

mmdetection 모델변환

https://mmdetection.readthedocs.io/en/latest/useful_tools.html#model-conversion

Exporting MMDetection models to ONNX format
https://medium.com/axinc-ai/exporting-mmdetection-models-to-onnx-format-3ec839c38ff

openvino : https://da2so.tistory.com/63

Serving pre-trained ML/DL models
https://docs.mlrun.org/en/stable/tutorial/03-model-serving.html#serving-pre-trained-ml-dl-models

docker 정리

더 이상 컨테이너에 연결되지 않고, 태그가 없어진 이미지를 Dangling image

https://sarc.io/index.php/aws/1921-docker

docker rmi $(docker images -f "dangling=true" -q)

docker exec -it nuclio-nuclio-mmdet-faster-rcnn-x bash

Nuclio 모델배포 : custom model

모델배포 : custom model – mmdetection

0. inference demo 확인

1. MakeConfig

2. dockerfile

function.yaml

Nuclio에 적용

main.py

mmdetection 모델변환

docker 정리

onesixx

vision

Nuclio 모델배포 : mmdetect

vision

Nuclio 모델배포 : custom model – Mask

vision

Nuclio Deploying Functions

Nuclio 모델배포 : custom model

모델배포 : custom model – mmdetection

0. inference demo 확인

1. MakeConfig

2. dockerfile

function.yaml

Nuclio에 적용

main.py

mmdetection 모델변환

docker 정리

onesixx

Related Posts

vision

Nuclio 모델배포 : mmdetect

vision

Nuclio 모델배포 : custom model – Mask

vision

Nuclio Deploying Functions