MLflow_doc_zh_cn/logging-first-model.ipynb

记录第一个模型¶

In [1]:

from mlflow import MlflowClient
from pprint import pprint
from sklearn.ensemble import RandomForestRegressor

创建一个 MLflow 客户端对象¶

使用 mlflow server --host 127.0.0.1 --port 8080 启动一个服务。

In [2]:

# NOTE: 请确保与你启动的服务地址一致

client = MlflowClient(tracking_uri="http://127.0.0.1:8080")

使用 MLflow 客户端 API 搜索创建的实验¶

Let's take a look at the Default Experiment that is created for us.

This safe 'fallback' experiment will store Runs that we create if we don't specify a new experiment.

In [3]:

# Search experiments without providing query terms behaves effectively as a 'list' action

all_experiments = client.search_experiments()

print(all_experiments)

[<Experiment: artifact_location='mlflow-artifacts:/345492218691480896', creation_time=1705671588054, experiment_id='345492218691480896', last_update_time=1705671588054, lifecycle_stage='active', name='MLflow Quickstart', tags={}>, <Experiment: artifact_location='mlflow-artifacts:/0', creation_time=1705670739503, experiment_id='0', last_update_time=1705670739503, lifecycle_stage='active', name='Default', tags={}>]

In [4]:

# Extract the experiment name and lifecycle_stage

default_experiment = [
    {"name": experiment.name, "lifecycle_stage": experiment.lifecycle_stage}
    for experiment in all_experiments
    if experiment.name == "Default"
][0]

pprint(default_experiment)

{'lifecycle_stage': 'active', 'name': 'Default'}

创建一个新实验¶

In this section, we'll:

• create a new MLflow Experiment
• apply metadata in the form of Experiment Tags

In [5]:

experiment_description = (
    "This is the grocery forecasting project. "
    "This experiment contains the produce models for apples."
)

experiment_tags = {
    "project_name": "grocery-forecasting",
    "store_dept": "produce",
    "team": "stores-ml",
    "project_quarter": "Q3-2023",
    "mlflow.note.content": experiment_description,
}

produce_apples_experiment = client.create_experiment(name="Apple_Models", tags=experiment_tags)

In [6]:

# Use search_experiments() to search on the project_name tag key

apples_experiment = client.search_experiments(
    filter_string="tags.`project_name` = 'grocery-forecasting'"
)

pprint(apples_experiment[0])

<Experiment: artifact_location='mlflow-artifacts:/715035017833909618', creation_time=1705818634613, experiment_id='715035017833909618', last_update_time=1705818634613, lifecycle_stage='active', name='Apple_Models', tags={'mlflow.note.content': 'This is the grocery forecasting project. This '
                        'experiment contains the produce models for apples.',
 'project_name': 'grocery-forecasting',
 'project_quarter': 'Q3-2023',
 'store_dept': 'produce',
 'team': 'stores-ml'}>

In [8]:

# Access individual tag data

print(apples_experiment[0].tags["team"])

stores-ml

运行第一个模型训练¶

In this section, we'll:

• create a synthetic data set that is relevant to a simple demand forecasting task
• start an MLflow run
• log metrics, parameters, and tags to the run
• save the model to the run
• register the model during model logging

生成苹果需求的综合数据¶

Keep in mind that this is purely for demonstration purposes.

The demand value is purely artificial and is deliberately covariant with the features. This is not a particularly realistic real-world scenario (if it were, we wouldn't need Data Scientists!).

In [9]:

import pandas as pd
import numpy as np
from datetime import datetime, timedelta


def generate_apple_sales_data_with_promo_adjustment(base_demand: int = 1000, n_rows: int = 5000):
    """
    Generates a synthetic dataset for predicting apple sales demand with seasonality and inflation.

    This function creates a pandas DataFrame with features relevant to apple sales.
    The features include date, average_temperature, rainfall, weekend flag, holiday flag,
    promotional flag, price_per_kg, and the previous day's demand. The target variable,
    'demand', is generated based on a combination of these features with some added noise.

    Args:
        base_demand (int, optional): Base demand for apples. Defaults to 1000.
        n_rows (int, optional): Number of rows (days) of data to generate. Defaults to 5000.

    Returns:
        pd.DataFrame: DataFrame with features and target variable for apple sales prediction.

    Example:
        >>> df = generate_apple_sales_data_with_seasonality(base_demand=1200, n_rows=6000)
        >>> df.head()
    """

    # Set seed for reproducibility
    np.random.seed(9999)

    # Create date range
    dates = [datetime.now() - timedelta(days=i) for i in range(n_rows)]
    dates.reverse()

    # Generate features
    df = pd.DataFrame(
        {
            "date": dates,
            "average_temperature": np.random.uniform(10, 35, n_rows),
            "rainfall": np.random.exponential(5, n_rows),
            "weekend": [(date.weekday() >= 5) * 1 for date in dates],
            "holiday": np.random.choice([0, 1], n_rows, p=[0.97, 0.03]),
            "price_per_kg": np.random.uniform(0.5, 3, n_rows),
            "month": [date.month for date in dates],
        }
    )

    # 随着时间的推移引入通货膨胀（年）
    df["inflation_multiplier"] = 1 + (df["date"].dt.year - df["date"].dt.year.min()) * 0.03

    # 考虑到苹果收获的季节性
    df["harvest_effect"] = np.sin(2 * np.pi * (df["month"] - 3) / 12) + np.sin(
        2 * np.pi * (df["month"] - 9) / 12
    )

    # 根据收获效果修改price_per_kg
    df["price_per_kg"] = df["price_per_kg"] - df["harvest_effect"] * 0.5

    # 调整促销期，使其与滞后高峰收获期 1 个月一致
    peak_months = [4, 10]  # months following the peak availability
    df["promo"] = np.where(
        df["month"].isin(peak_months),
        1,
        np.random.choice([0, 1], n_rows, p=[0.85, 0.15]),
    )

    # 根据特征生成目标变量
    base_price_effect = -df["price_per_kg"] * 50
    seasonality_effect = df["harvest_effect"] * 50
    promo_effect = df["promo"] * 200

    df["demand"] = (
        base_demand
        + base_price_effect
        + seasonality_effect
        + promo_effect
        + df["weekend"] * 300
        + np.random.normal(0, 50, n_rows)
    ) * df[
        "inflation_multiplier"
    ]  # 引入随机噪声

    # Add previous day's demand
    df["previous_days_demand"] = df["demand"].shift(1)
    df["previous_days_demand"].fillna(method="bfill", inplace=True)  # fill the first row

    # Drop temporary columns
    df.drop(columns=["inflation_multiplier", "harvest_effect", "month"], inplace=True)

    return df

In [60]:

# Generate the dataset!

data = generate_apple_sales_data_with_promo_adjustment(base_demand=1_000, n_rows=1_000)

data[-20:]

/tmp/ipykernel_1170/724292291.py:84: FutureWarning: Series.fillna with 'method' is deprecated and will raise in a future version. Use obj.ffill() or obj.bfill() instead.
  df["previous_days_demand"].fillna(method="bfill", inplace=True)  # fill the first row

Out[60]:

	date	average_temperature	rainfall	weekend	holiday	price_per_kg	promo	demand	previous_days_demand
980	2024-01-02 15:05:05.013229	34.130183	1.454065	0	0	1.449177	0	999.306290	1029.418398
981	2024-01-03 15:05:05.013227	32.353643	9.462859	0	0	2.856503	0	842.129427	999.306290
982	2024-01-04 15:05:05.013225	18.816833	0.391470	0	0	1.326429	0	990.616709	842.129427
983	2024-01-05 15:05:05.013223	34.533012	2.120477	0	0	0.970131	0	1068.802075	990.616709
984	2024-01-06 15:05:05.013222	23.057202	2.365705	1	0	1.049931	0	1346.486305	1068.802075
985	2024-01-07 15:05:05.013220	34.810165	3.089005	1	0	2.035149	0	1329.564672	1346.486305
986	2024-01-08 15:05:05.013218	29.208905	3.673292	0	0	2.518098	0	1086.143402	1329.564672
987	2024-01-09 15:05:05.013216	16.428676	4.077782	0	0	1.268979	0	1093.207186	1086.143402
988	2024-01-10 15:05:05.013214	32.067512	2.734454	0	0	0.762317	0	1069.939894	1093.207186
989	2024-01-11 15:05:05.013213	31.938203	13.883486	0	0	1.153301	0	994.409540	1069.939894
990	2024-01-12 15:05:05.013211	18.024055	7.544061	0	0	0.610703	0	1078.323183	994.409540
991	2024-01-13 15:05:05.013209	20.681067	18.820490	1	0	1.533488	0	1328.499120	1078.323183
992	2024-01-14 15:05:05.013207	16.010132	7.705941	1	0	1.632498	1	1548.922141	1328.499120
993	2024-01-15 15:05:05.013198	18.766455	6.274840	0	0	2.806554	0	956.412724	1548.922141
994	2024-01-16 15:05:05.013196	27.948793	23.705246	0	0	0.829464	0	1090.592622	956.412724
995	2024-01-17 15:05:05.013194	28.661072	10.329865	0	0	2.290591	0	936.465043	1090.592622
996	2024-01-18 15:05:05.013192	10.821693	3.575645	0	0	0.897473	0	1016.336362	936.465043
997	2024-01-19 15:05:05.013190	21.108560	6.221089	0	0	1.093864	0	1063.698477	1016.336362
998	2024-01-20 15:05:05.013187	29.451301	5.021463	1	0	2.493085	0	1306.255235	1063.698477
999	2024-01-21 15:05:05.013172	19.261458	0.438381	1	0	2.610422	0	1207.188828	1306.255235

训练并记录模型数据¶

We're now ready to import our model class and train a RandomForestRegressor

In [61]:

import mlflow
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score


# Use the fluent API to set the tracking uri and the active experiment
mlflow.set_tracking_uri("http://127.0.0.1:8080")

# Sets the current active experiment to the "Apple_Models" experiment and returns the Experiment metadata
apple_experiment = mlflow.set_experiment("Apple_Models")

# Define a run name for this iteration of training.
# If this is not set, a unique name will be auto-generated for your run.
run_name = "apples_rf_test"

# Define an artifact path that the model will be saved to.
artifact_path = "rf_apples"

In [62]:

# Split the data into features and target and drop irrelevant date field and target field
X = data.drop(columns=["date", "demand"])
y = data["demand"]

# 将数据分为训练集和测试集
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.2, random_state=42)

params = {
    "n_estimators": 100,
    "max_depth": 6,
    "min_samples_split": 10,
    "min_samples_leaf": 4,
    "bootstrap": True,
    "oob_score": False,
    "random_state": 888,
}

# Train the RandomForestRegressor
rf = RandomForestRegressor(**params)

# Fit the model on the training data
rf.fit(X_train, y_train)

# Predict on the validation set
y_pred = rf.predict(X_val)

# Calculate error metrics
mae = mean_absolute_error(y_val, y_pred)
mse = mean_squared_error(y_val, y_pred)
rmse = np.sqrt(mse)
r2 = r2_score(y_val, y_pred)

# Assemble the metrics we're going to write into a collection
metrics = {"mae": mae, "mse": mse, "rmse": rmse, "r2": r2}

# Initiate the MLflow run context
with mlflow.start_run(run_name=run_name) as run:
    # Log the parameters used for the model fit
    mlflow.log_params(params)

    # Log the error metrics that were calculated during validation
    mlflow.log_metrics(metrics)

    # Log an instance of the trained model for later use
    mlflow.sklearn.log_model(sk_model=rf, input_example=X_val, artifact_path=artifact_path)

/home/deck/miniconda3/envs/mlflow/lib/python3.10/site-packages/mlflow/models/signature.py:358: UserWarning: Hint: Inferred schema contains integer column(s). Integer columns in Python cannot represent missing values. If your input data contains missing values at inference time, it will be encoded as floats and will cause a schema enforcement error. The best way to avoid this problem is to infer the model schema based on a realistic data sample (training dataset) that includes missing values. Alternatively, you can declare integer columns as doubles (float64) whenever these columns may have missing values. See `Handling Integers With Missing Values <https://www.mlflow.org/docs/latest/models.html#handling-integers-with-missing-values>`_ for more details.
  input_schema = _infer_schema(input_example)

Success!¶

You've just logged your first MLflow model!

Navigate to the MLflow UI to see the run that was just created (named "apples_rf_test", logged to the Experiment "Apple_Models").