嗨。我对MLFlow非常陌生，我想在我自己的ML模型上实现MLFlow项目。然而，我得到的是“无法在入口点中找到主要的”

Question

完整的错误消息如下：

ERROR mlflow.cli: === Could not find main among entry points [] or interpret main as a runnable script. Supported script file extensions: ['.py', '.sh'] ===

我也尝试了这里建议的解决方案，https://github.com/mlflow/mlflow/issues/1094，但结果是一样的。

下面我提供了运行MLflow项目所需的所有文件。

conda.yaml文件

name: lightgbm-example
channels:
  - conda-forge
dependencies:
  - python=3.6
  - pip
  - pip:
      - mlflow>=1.6.0
      - lightgbm
      - pandas
      - numpy

MLProject文件

name: lightgbm-example
conda_env: ~/Desktop/MLflow/conda.yaml
entry-points:
    main:
      parameters:
        learning_rate: {type: float, default: 0.1}
        colsample_bytree: {type: float, default: 1.0}
        subsample: {type: float, default: 1.0} 
      command: |
          python3 ~/Desktop/MLflow/Test.py \
            --learning-rate={learning_rate} \
            --colsample-bytree={colsample_bytree} \
            --subsample={subsample}

我的Test.py文件

import pandas as pd
import lightgbm as lgb
import numpy as np
import mlflow
import mlflow.lightgbm
import argparse
from sklearn.metrics import accuracy_score, confusion_matrix


def parse_args():
    parser = argparse.ArgumentParser(description="LightGBM example")
    parser.add_argument(
        "--learning-rate",
        type=float,
        default=0.1,
        help="learning rate to update step size at each boosting step (default: 0.3)",
    )
    parser.add_argument(
        "--colsample-bytree",
        type=float,
        default=1.0,
        help="subsample ratio of columns when constructing each tree (default: 1.0)",
    )
    parser.add_argument(
        "--subsample",
        type=float,
        default=1.0,
        help="subsample ratio of the training instances (default: 1.0)",
    )
    return parser.parse_args()

def find_specificity(c_matrix):
    specificity = c_matrix[1][1]/(c_matrix[1][1]+c_matrix[0][1])
    return specificity
    
    
def main():

    args = parse_args()

    df = pd.read_csv('~/Desktop/MLflow/Churn_demo.csv')
    train_df = df.sample(frac=0.8, random_state=25)
    test_df = df.drop(train_df.index)


        
    train_df.drop(['subscriberid'], axis = 1, inplace = True)
    test_df.drop(['subscriberid'], axis = 1, inplace = True)

    TrainX = train_df.iloc[:,:-1]
    TrainY = train_df.iloc[:,-1]

    TestX = test_df.iloc[:,:-1]
    TestY = test_df.iloc[:,-1]
    
    mlflow.lightgbm.autolog()
    
    dtrain = lgb.Dataset(TrainX, label=TrainY)
    dtest = lgb.Dataset(TestX, label=TestY)
    
    with mlflow.start_run():

        parameters = {
            'objective': 'binary',
            'device':'cpu',
            'num_threads': 6,
            'num_leaves': 127,
            'metric' : 'binary',
            'lambda_l2':5,
            'max_bin': 63,
            'bin_construct_sample_cnt' :2*1000*1000,
            'learning_rate': args.learning_rate,
            'colsample_bytree': args.colsample_bytree,
            'subsample': args.subsample,
            'verbose': 1
        }



        model = lgb.train(parameters,
                       dtrain,
                       valid_sets=dtest,
                       num_boost_round=10000,
                       early_stopping_rounds=10)
                       
               
        y_proba=model.predict(TestX)
        pred=np.where(y_proba>0.25,1,0) 
        conf_matrix = confusion_matrix(TestY,pred)
        
        specificity = find_specificity(conf_matrix)
        acc = accuracy_score(TestY,pred)
        
        mlflow.log_metric({"specificity" : specificity, "accuracy" : acc})


if __name__ == "__main__":
    main()

Answer 1

幸运的是，我的问题已经解决了。我列举了一些同样的错误的解决方案，如果你面临同样的问题，它可以在将来帮助你。

1. 文件名。文件名应该与MLFlow docs https://mlflow.org/中建议的相同。例如，不支持conda.yamp，但是conda.yaml，由于conda.yaml conda.yaml文件中存在不支持Tab的问题，请考虑使用MLProject文件名'P‘中的空格代替
2. 作为MLFlow 1.4之前的大写。但是后面的版本并不重要，正如这里所解释的，https://github.com/mlflow/mlflow/issues/1094
3. (In，my ) MLProject文件是空间敏感的。让https://github.com/mlflow/mlflow/tree/master/examples GitHub示例指导您。