mlops-prototyping
Structure reproducible Jupyter notebooks with a fixed section layout, hoisted configuration, and leakage-free scikit-learn pipelines. Use when exploring a dataset, training a first model, or preparing a notebook for promotion.
Médéric HURIER (Fmind) · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add MLOps-Courses/mlops-coding-skills --skill mlops-prototyping技能说明
name: mlops-prototyping description: Structure reproducible Jupyter notebooks with a fixed section layout, hoisted configuration, and leakage-free scikit-learn pipelines. Use when exploring a dataset, training a first model, or preparing a notebook for promotion. license: MIT metadata: author: Médéric HURIER (Fmind) source: github.com/MLOps-Courses/mlops-coding-skills/tree/main/mlops-prototyping created: 2026-01-25 updated: 2026-08-10
MLOps Prototyping
Goal
To create standardized, reproducible, and production-ready prototypes in Jupyter notebooks. This skill enforces a structured layout (Imports -> Configs -> Load -> EDA -> Modeling -> Eval) and robust engineering practices (Pipelines, Split-Verification) to prevent technical debt and data leakage.
Prerequisites
- Language: Python 3.14
- Environment:
uvmanaged project (.venv), withipykernelin anotebookdependency group - Context: Executed within a
.ipynbfile or converting to one.
Instructions
1. Notebook Structure
Enforce the following linear sections in every notebook to ensure readability and maintainability.
- Title & Purpose: H1 Title and a brief description of the experiment goals.
- Imports: Group standard libraries, third-party, and usage-specific imports.
- Configs: Define Global Constants (paths, random seeds, hyperparameters) here. No magic numbers deeper in the code.
- Datasets: Load, validate, and split data.
- Analysis (EDA): Inspect target distributions and correlations.
- Modeling: Define and train
sklearn.pipeline.Pipelineobjects. - Evaluations: Compute metrics and visualize performance on held-out data.
2. Configuration Standards
Expose all "knobs" at the top of the notebook for easy experimentation.
-
Randomness: Define
RANDOM_STATE = 42and use it in splits and model initialization. -
Paths: Use
pathlibfor robust path handling.from pathlib import Path ROOT = Path("..") DATA_PATH = ROOT / "data" / "input.parquet" -
Hyperparameters: Group model params (e.g.,
N_ESTIMATORS,MAX_DEPTH). -
Toggles: Use booleans for expensive operations (e.g.,
USE_GPU = True,RUN_GRID_SEARCH = False).
3. Data Management
Ensure data integrity and prevent leakage.
- Loading: Prefer
pd.read_parquetfor speed/types, orpd.read_csv. - Splitting:
- Always split into
X_train,X_test,y_train,y_testbefore any data-dependent transformations (imputation, scaling). - Random Split: Use
sklearn.model_selection.train_test_splitwithstratifyfor balanced classification. - Time Series: Use
sklearn.model_selection.TimeSeriesSplitif data has a temporal dimension (do NOT shuffle). - Use
random_state=RANDOM_STATE.
- Always split into
4. Pipeline Construction
Prohibit raw data transformations on the full dataset.
-
Mandate: Use
sklearn.pipeline.PipelineorColumnTransformer. -
Why: Automation of
fiton train andtransformon test prevents data leakage. -
Example:
from sklearn.compose import ColumnTransformer from sklearn.impute import SimpleImputer from sklearn.pipeline import Pipeline from sklearn.preprocessing import StandardScaler CACHE = "./.cache" # Define a cache directory numeric_transformer = Pipeline( steps=[ ("imputer", SimpleImputer(strategy="median")), ("scaler", StandardScaler()), ] ) preprocessor = ColumnTransformer( transformers=[("num", numeric_transformer, numeric_features)] ) # Use 'memory' to cache transformer outputs, speeding up GridSearch model = Pipeline( steps=[ ("preprocessor", preprocessor), ("classifier", RandomForestClassifier()), ], memory=CACHE, )
5. Experiment Tracking from the Notebook
Prototypes are experiments; record them from the first run rather than retrofitting tracking later.
- Backend: Point MLflow at a SQL store, not the deprecated file store:
mlflow.set_tracking_uri("sqlite:///mlflow.db"). SQLite is a real SQLAlchemy backend, it supports the model registry, and it is the same store shape as a production Postgres — so moving up later is a URI change, not a rewrite. - Autologging:
mlflow.autolog()beforefitcaptures parameters, metrics, and the model for scikit-learn without extra code. - Scope: Keep one MLflow experiment per notebook question, and name runs after the hypothesis being tested.
6. Evaluation & Visualization
Go beyond accuracy/MSE.
- Metrics: Use
sklearn.metricsappropriate for the task (F1, ROC-AUC, RMSE, MAE). - Baselines: Compare against a "Dummy" model (mean/mode) to verify learning.
- Visualization:
- Regression: Residual plots, Actual vs Predicted.
- Classification: Confusion Matrix, ROC Curve, Precision-Recall.
- Feature Importance: Visualize
feature_importances_or SHAP values.
7. Transition to Production
Facilitate the move from notebook to python package (src/).
- Function Refactoring: Once a block of code is stable (e.g., a complex data cleaning step), refactor it into a function within the notebook. This makes moving it to a
.pyfile trivial later. - Cell Tagging: Use tags like
parameters(for Papermill) orexportto mark cells that should be part of the final documentation or automated pipeline. - Clean State: Ensure the notebook runs top-to-bottom (
Restart Kernel and Run All) without errors before committing. - Formatting: Ruff 0.16 formats Python inside Markdown too, so
mise run formatnormalizes the snippets you paste into notes and docs. Runmise run allbefore committing a notebook alongside package code. - Next step: mlops-industrialization covers the package layout the refactored functions move into.
Self-Correction Checklist
- No Magic Numbers: Are all parameters in