注册模型¶

注册一个模型非常简单:

service.register(
    model_name="mymodel",
    model=model,
    entrypoint="predict",
)

如果我有多个模型，或者有多个版本呢?

你可以注册多个模型，每个模型可以有不同的版本:

service.register(
    model_name="my-model",
    model=my_model,
    entrypoint="predict",
)
service.register(
    model_name="my-model",
    model=my_model_v1,
    entrypoint="predict",
    version_name="v1,
)

service.register(
    model_name="your-model",
    model=your_model,
    entrypoint="predict",
)
service.register(
    model_name="your-model",
    model=your_model_v1,
    entrypoint="predict",
    version_name="v1,
)
service.register(
    model_name="your-model",
    model=your_model_v2,
    entrypoint="predict",
    version_name="v2,
)

参数¶

参数	类似	默认值（如有）	细节
model_name	str		模型名称
model	object		模型Python对象，或者模型文件路径
version_name	str	None	版本名称
entrypoint	str	None	用来预测的函数名称
metadata	dict	None	模型基础信息
handler	object	None	Hanlder 类
load_now	bool	True	是否立刻载入模型

例子¶

Model Name¶

from pinferencia import Server


def predict(data):
    return sum(data)

service = Server()
service.register(
    model_name="mymodel",
    model=predict,
)

Model¶

Model ObjectFunction

from pinferencia import Server


class MyModel:
    def predict(self, data):
        return sum(data)


model = MyModel()

service = Server()
service.register(
    model_name="mymodel",
    model=model,
    entrypoint="predict
)

from pinferencia import Server


def predict(data):
    return sum(data)

service = Server()
service.register(
    model_name="mymodel",
    model=predict,
)

Version名称¶

没有提供版本名的模型会用 default 版本名.

from pinferencia import Server


def add(data):
    return data[0] + data[1]

def substract(data):
    return data[0] + data[1]

service = Server()
service.register(
    model_name="mymodel",
    model=add,
    version_name="add", # (1)
)
service.register(
    model_name="mymodel",
    model=substract,
    version_name="substract", # (2)
)

预测地址在 http://127.0.0.1/v1/models/mymodel/versions/add/predict
预测地址在 http://127.0.0.1/v1/models/mymodel/versions/substract/predict

Entrypoint¶

from pinferencia import Server


class MyModel:
    def add(self, data):
        return data[0] + data[1]

    def substract(self, data):
        return data[0] - data[1]


model = MyModel()

service = Server()
service.register(
    model_name="mymodel",
    model=model,
    version_name="add", # (1)
    entrypoint="add", # (3)
)
service.register(
    model_name="mymodel",
    model=model,
    version_name="substract", # (2)
    entrypoint="substract", # (4)
)

预测地址在 http://127.0.0.1/v1/models/mymodel/versions/add/predict
预测地址在 http://127.0.0.1/v1/models/mymodel/versions/substract/predict
add 函数会被用来预测.
substract 函数会被用来预测.

Metadata¶

默认API¶

Pinferencia 默认metadata架构支持 platform 和 device

这些信息仅供展示。

from pinferencia import Server


def predict(data):
    return sum(data)

service = Server()
service.register(
    model_name="mymodel",
    model=predict,
    metadata={
        "platform": "Linux",
        "device": "CPU+GPU",
    }
)

Kserve API¶

Pinferencia 同时支持 Kserve API.

对于 Kserve V2, 模型metadata支持: - platform - inputs - outputs

inputs 和 outputs 会决定模型收到的数据和返回的数据类型.

from pinferencia import Server


def predict(data):
    return sum(data)

service = Server(api="kserve") # (1)
service.register(
    model_name="mymodel",
    model=predict,
    metadata={
        "platform": "mac os",
        "inputs": [
            {
                "name": "integers", # (2)
                "datatype": "int64",
                "shape": [1],
                "data": [1, 2, 3],
            }
        ],
        "outputs": [
            {"name": "sum", "datatype": "int64", "shape": -1, "data": 6}, # (3)
            {"name": "product", "datatype": "int64", "shape": -1, "data": 6},
        ],
    }
)

如果要使用 Kserve API 需要在实例化服务时设置 api="kserve"。
如果请求包含多组数据，只有 intergers 数据会被传递给模型。
输出数据会被转换为 int64。datatype 字段仅支持numpy 数据类型. 如果类型转换失败，响应里会多出 error 字段。

Handler¶

关于Handler的细节，请查看Handlers.

from pinferencia import Server
from pinferencia.handlers import PickleHandler


class MyPrintHandler(PickleHandler):

    def predict(self, data):
        print(data)
        return self.model.predict(data)


def predict(data):
    return sum(data)

service = Server()
service.register(
    model_name="mymodel",
    model=predict,
    handler=MyPrintHandler
)

Load Now¶

import joblib

from pinferencia import Server


class JoblibHandler(BaseHandler):
    def load_model(self):
        return joblib.load(self.model_path)


service = Server(model_dir="/opt/models")
service.register(
    model_name="mymodel",
    model="/path/to/model.joblib",
    entrypoint="predict",
    handler=JoblibHandler,
    load_now=True,
)