# Kubeflow實作筆記：新增 Random Forest 演算法
## 目錄
[TOC]
## 參考資料

:::warning
建議先看：
https://towardsdatascience.com/kubeflow-pipelines-how-to-build-your-first-kubeflow-pipeline-from-scratch-2424227f7e5
:::
https://github.com/FernandoLpz/Kubeflow_Pipelines

https://hub.docker.com/r/fernandolpz/only-tests/tags

https://hub.docker.com/repository/docker/lightnighttw/kubeflow/general

## 架構
每個方塊代表一個組件，每個組件都是一個容器。
# Random Forest Image
## 建立新資料夾randomForest

## Dockerfile

FROM python:3.8-slim
WORKDIR /pipelines
COPY requirements.txt /pipelines
RUN pip install -r requirements.txt
COPY randomforest.py /pipelines

## ramdom_forest.yaml

name: Random Forest classifier
description: Train a random forest classifier

inputs:
- {name: Data, type: LocalPath, description: 'Path where data is stored.'}
outputs:
- {name: Accuracy, type: Float, description: 'Accuracy metric'}

implementation:
  container:
    image: lightnighttw/kubeflow:random_forest_v4
    command: [
      python, randomforest.py,

      --data,
      {inputPath: Data},

      --accuracy,
      {outputPath: Accuracy},

    ]

## randomforest.py

import json

import argparse
from pathlib import Path

from sklearn.metrics import accuracy_score
from sklearn.ensemble import RandomForestClassifier

def _randomforest(args):

    # Open and reads file "data"
    with open(args.data) as data_file:
        data = json.load(data_file)
    
    # The excted data type is 'dict', however since the file
    # was loaded as a json object, it is first loaded as a string
    # thus we need to load again from such string in order to get 
    # the dict-type object.
    data = json.loads(data)

    x_train = data['x_train']
    y_train = data['y_train']
    x_test = data['x_test']
    y_test = data['y_test']
    
    # Initialize and train the model
    model = RandomForestClassifier(n_estimators=100, criterion = 'gini')
    model.fit(x_train, y_train)

    # Get predictions
    y_pred = model.predict(x_test)
    
    # Get accuracy
    accuracy = accuracy_score(y_test, y_pred)
    
    # Save output into file
    with open(args.accuracy, 'w') as accuracy_file:
        accuracy_file.write(str(accuracy))



if __name__ == '__main__':

    # Defining and parsing the command-line arguments
    parser = argparse.ArgumentParser(description='My program description')
    parser.add_argument('--data', type=str)
    parser.add_argument('--accuracy', type=str)

    args = parser.parse_args()

    # Creating the directory where the output file will be created (the directory may or may not exist).
    Path(args.accuracy).parent.mkdir(parents=True, exist_ok=True)
    
    _randomforest(args)

## requirements.txt
:::danger
sklearn在新版的pip已經不能用，需要pip install scikit-learn
:::

scikit-learn

## 上傳到Docker Hub
[安裝docker](https://www.docker.com)
在ramdom_forest.yaml定義了image的取用位址lightnighttw/kubeflow:random_forest_v4(需自訂)
[Docker Hub](https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&cad=rja&uact=8&ved=2ahUKEwjwy-SU35f-AhWRU94KHU9aDq4QFnoECBEQAQ&url=https%3A%2F%2Fhub.docker.com%2F&usg=AOvVaw08oOGeIqomrCmPs9p07hDk)
沒帳號就申請一個

#將以上四個檔案下載到自己的電腦的新資料夾裡
#登入
docker login -u "docker用户名"
#docker build
docker build --platform=linux/amd64 -t <docker-registry-username>/<docker-image-name>:<tag_name> . -f Dockerfile
#推送到Docker Hub
docker push <docker-registry-username>/<docker-image-name>:<tag_name>
Ex: docker push lightnighttw/kubeflow:random_forest_v4

如果不行就用我的[image](https://hub.docker.com/repository/docker/lightnighttw/kubeflow/general)
# 建立pipeline
## pipeline.py
把前面所有組建的資訊(yaml)引入編譯成管線的yaml檔

import kfp
from kfp import dsl
from kfp.components import func_to_container_op

@func_to_container_op
def show_results(decision_tree : float, logistic_regression : float, random_forest : float) -> None:
    # Given the outputs from decision_tree and logistic regression components
    # the results are shown.

    print(f"Decision tree (accuracy): {decision_tree}")
    print(f"Logistic regression (accuracy): {logistic_regression}")
    print(f"Random forest (accuracy): {random_forest}")

def add_resource_constraints(op: dsl.ContainerOp):
    return op.set_cpu_request("1").set_cpu_limit("2")


@dsl.pipeline(name='Three Pipeline', description='Applies Decision Tree, random forest and Logistic Regression for classification problem.')
def first_pipeline():

    # Loads the yaml manifest for each component
    download = kfp.components.load_component_from_file('download_data/download_data.yaml')
    decision_tree = kfp.components.load_component_from_file('decision_tree/decision_tree.yaml')
    logistic_regression = kfp.components.load_component_from_file('logistic_regression/logistic_regression.yaml')
    random_forest = kfp.components.load_component_from_file('randomForest/random_forest.yaml')

    # Run download_data task
    download_task = add_resource_constraints(download())

    # Run tasks "decison_tree" and "logistic_regression" given
    # the output generated by "download_task".
    decision_tree_task = add_resource_constraints(decision_tree(download_task.output))
    logistic_regression_task = add_resource_constraints(logistic_regression(download_task.output))
    random_forest_task = add_resource_constraints(random_forest(download_task.output))
    # Given the outputs from "decision_tree" and "logistic_regression"
    # the component "show_results" is called to print the results.
    add_resource_constraints(show_results(decision_tree_task.output, logistic_regression_task.output, random_forest_task.output))



if __name__ == '__main__':
    kfp.compiler.Compiler().compile(first_pipeline, 'three_pipelines.yaml')
    # kfp.Client().create_run_from_pipeline_func(basic_pipeline, arguments={})

## 編譯

## 上傳pipeline
若成功執行