p5_03 - Train Production Model (deployable)¶

Get the Video Course

Step 3 of 3 (final). Load the winning config from p5_02's YAML artifact. Train the final production model. No HP search, no threshold gymnastics. Fast, deterministic, CI/CD-friendly. This is the deployable artifact - the model you ship to inference services.

1. What we are going to do?¶

p5_02 discovered the production-ready hyperparameters via GridSearchCV and saved them to artifacts/p5_02_best_config.yaml. This script:

1. Reads that YAML config
2. Trains the final production model with those exact hyperparameters
3. Evaluates on the test set
4. Saves the model + metadata to models/p5_03_production/

That is it. No search. No tuning. No threshold magic. Just train with the known-good config and save.

The production handoff pattern¶

This is the MLOps re-training loop. p5_02 writes artifacts/p5_02_best_config.yaml (left box). p5_03 reads that YAML (right box). The model produced (models/p5_03_production/model_latest.pkl) is what Section 08's inference API loads. Re-run p5_02 on fresh data → YAML auto-updates → re-run p5_03 → new production model. Zero manual copy-paste of hyperparameters between training and deployment.

Why a separate script? In real MLOps you re-train your model regularly (daily / weekly) on fresh data. You do NOT re-run hyperparameter search every time - that is expensive and the result barely changes. Instead, you cache the hyperparameters once (p5_02), and the re-training job (THIS script) just loads them and trains. Re-running p5_02 on fresh data updates the artifact, and the next run of this script automatically picks up the new config.

2. What exactly is it?¶

The script is purposefully minimal - it does training only. Notable code blocks:

• with open(INPUT_ARTIFACT) as f: config = yaml.safe_load(f) - reads the YAML artifact
• best_params = config['best_params'] - extracts the model params. class_weight={0:1, 1:100} arrives as a dict with int keys, ready to hand to sklearn as-is
• pipeline = create_pipeline(..., model_params=best_params) - same create_pipeline() helper as p3_01 / p4_03 / p5_01-02
• pipeline.fit(X_train, y_train) - one train call, no search wrapping
• pipeline.predict(X_test) - default threshold 0.5 (because the aggressiveness is already baked into class_weight)
• save_pipeline(...) - saves to models/p5_03_production/

If p5_02 has not been run, this script errors with a clear message asking you to run it first.

3. How to run it¶

Step 1 - make sure you are in the mlops-env1 conda environment¶

conda info --envs
# If active env is not mlops-env1:
conda deactivate
conda activate mlops-env1

Step 2 - go to the project folder¶

cd 07_Hyperparameter_Tuning/ccfd-project

Step 3 - make sure data + FE1 CSV + p5_02 config exist¶

python p1_01_generate_initial_dataset.py            # only if data does not exist
python p4_01_feature_engineering_fe1.py              # only if FE1 CSV does not exist
python p5_02_class_weight_gridsearch.py              # creates artifacts/p5_02_best_config.yaml

Step 4 - run this script¶

python p5_03_train_production_model.py

Runtime: ~2-3 seconds. Just a single model train + test evaluation + save.

4. What is the result output?¶

The script prints, in order:

• Header banner: p5_03 - TRAIN PRODUCTION MODEL (uses p5_02 winning config)
• STEP 1: loads artifacts/p5_02_best_config.yaml and prints the config

Production config (from p5_02):
  C: 0.1
  l1_ratio: 1.0
  solver: saga
  max_iter: 5000
  class_weight: {0: 1, 1: 100}

(Note: random_state is set internally by create_pipeline() for reproducibility - not part of the search grid, so it does not appear in the loaded config.)

The hyperparameter we inherited from p5_02 - why class_weight={0:1, 1:100} matters¶

The class_weight={0:1, 1:100} value in the config above is the bottom row of this image (the green one). p5_02 picked it from a 5-value grid because it matches the 60:1 dollar cost ratio of fraud detection. p5_03 just trains with that pre-discovered winner - no search, no second-guessing. This is what "deterministic re-training from a cached config" looks like in practice.

• STEP 2: load FE1 CSV + split
• STEP 3: train production model (no search)
• STEP 4: test-set evaluation block (full print_metrics() output)
• STEP 5: save model + metadata to models/p5_03_production/
• Closing FULL IMPROVEMENT JOURNEY summary:

FULL IMPROVEMENT JOURNEY (final)
======================================================================

  Stage 1 - Baseline (p3_01, no FE):           $39,625
  Stage 2 - + FE1 (p4_03, amount_ratio):       $39,700   (+$75)
  Stage 3 - + RandomSearch (p5_01):             typically HURT (narrow grid)
  Stage 4 - + GridSearch w/ cw (p5_02):         $50,550   (finds the right params)
  Stage 5 - Production model (p5_03, this):    $50,550

  Total change vs Stage 1 baseline:            +$10,925

  The production model is saved at models/p5_03_production/
  Re-train it any time by re-running p5_02 (updates artifact) then this script.

5. Complete metrics table¶

For the production model (config inherited from p5_02), evaluated at default threshold 0.5:

Metrics match p5_02 exactly because we use the same config on the same train/test split with the same random seed.

Output artifacts: - models/p5_03_production/model_latest.pkl - deployable model - models/p5_03_production/metadata_latest.json - includes best_params, config_source, production_role - results/logistic_regression_p5_03_production_*.json

Where this fits in Section 07¶

Step 3 of 3. The final deployable. See the section README for the full 3-script arc and the comparison table.

After this section, you have a tuned production model at models/p5_03_production/. Section 08 uses this model to serve predictions via a FastAPI endpoint and compare against the baseline inference API from Section 04.

Back to 07 - Hyperparameter Tuning