modelscope + kubeflow

Dockerfile

# 使用稳定python瘦身版,减少镜像体积
FROM python:3.9.15-slim-buster
# 禁用pip缓存,减少镜像体积
ENV PIP_NO_CACHE_DIR=1

RUN sed -i 's/archive.ubuntu.com/mirrors.aliyun.com/g' /etc/apt/sources.list && \
	apt update && \
	apt-get install -y --no-install-recommends gcc ffmpeg libsm6 libxext6 g++ && \
	rm -rf /var/lib/apt/lists/*

# 安装依赖，没用到第三方包别往requirements.txt写
COPY *.whl ./
# 指定依赖镜像地址，加快下载
RUN python -m pip install --upgrade pip \
    && pip install -i https://pypi.tuna.tsinghua.edu.cn/simple *.whl \
    && rm -rf *.whl

COPY requirements.txt ./
RUN pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt


# 复制源码
COPY model ./model
COPY model.py ./

ENTRYPOINT ["python","-m","model"]

gistfile1.txt

modelscope==1.0.3
kserve
mmcls
mmcv
torchvision
transformers
fairseq
timm
unicodedata2
zhconv

model.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
import base64
import json
import os
import tempfile
from typing import Dict, Any

import kserve
import numpy
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks


class CustomJsonEncoder(json.JSONEncoder):
    def default(self, obj):
        if isinstance(obj, numpy.integer):
            return int(obj)
        elif isinstance(obj, numpy.floating):
            return float(obj)
        elif isinstance(obj, numpy.ndarray):
            return obj.tolist()
        return super(CustomJsonEncoder, self).default(obj)


def safe_json(output: Any) -> str:
    return json.dumps(
        output,
        ensure_ascii=False,
        allow_nan=False,
        indent=None,
        separators=(",", ":"),
        cls=CustomJsonEncoder,
    ).replace("</", "<\\/")


class AIModel(kserve.Model):
    def __init__(self, name: str):
        super().__init__(name)
        self.name = name
        self.load()

    def load(self):
        model = pipeline(Tasks.image_classification, model='model')
        self.model = model
        self.ready = True

    def predict(self, request: Dict) -> Dict:
        inputs = request["instances"]
        data = inputs[0]["image"]["b64"]

        img = base64.b64decode(data)
        with tempfile.NamedTemporaryFile(mode='wb', delete=False) as temp:
            temp.write(img)

        output = self.model(temp.name)
        os.remove(temp.name)

        return {"predictions": safe_json(output)}


if __name__ == "__main__":
    # 必须预留从环境变量中接收模型名字，MODEL_NAME这个key不能改，value影响到模型部署的URL: /v1/models/{MODEL_NAME}:predict
    model = AIModel(os.environ.get('MODEL_NAME', 'custom-model'))
    model.load()
    kserve.ModelServer(workers=1).start([model])