Hyperparameter Tuning with DVC

CI/CD for Machine Learning

Ravi Bhadauria

Machine Learning Engineer

Hyperparameter tuning workflow

Hyperparameter tuning
- Input: Parameter sweep ranges
- Output: Best parameters
Training
- Input: Best parameters
- Output: Metrics and plots (already covered)
Loose coupling for independent training
- Hyperparameter tuning sufficient but not necessary
Both jobs are dataset dependent

# Contents of hp configuration
{
    "n_estimators": [2, 4, 5],
    "max_depth": [10, 20, 50],
    "random_state": [1993]
}

# Contents of best parameters
{
    "n_estimators": 5,
    "max_depth": 20,
    "random_state": 1993
}

Training code changes

Changes in Python Hyperparameter tuning script

# Load hyperparameters from the JSON file
with open("rfc_best_params.json", "r") as params_file:
  rfc_params = json.load(params_file)


# Define and train model
model = RandomForestClassifier(**rfc_params)
model.fit(X_train, y_train)

Hyperparameter Tuning with GridSearch

# Define the model and hyperparameter search space
model = RandomForestClassifier()
param_grid = json.load(open("hp_config.json", "r"))


# Perform GridSearch with five fold CV
grid_search = GridSearchCV(model, param_grid, cv=5)
grid_search.fit(X_train, y_train)


# Get the best hyperparameters
best_params = grid_search.best_params_
with open("rfc_best_params.json", "w") as outfile:
        json.dump(best_params, outfile)

DVC YAML changes

Hyperparameter Tuning

stages:
  preprocess: ...
  train: ...
  hp_tune:
    cmd: python hp_tuning.py
    deps:
    - processed_dataset/weather.csv
    - hp_config.json
    - hp_tuning.py
    outs: # Not tracking best parameters
      - hp_tuning_results.md:
          cache: false

Training

stages:
  preprocess: ...
  hp_tune: ...
  train:
    cmd: python train.py
    deps:
    - processed_dataset/weather.csv
    - rfc_best_params.json # Best parameters
    - train.py
    metrics:
      - metrics.json:
          cache: false

Triggering individual stages

Stages can be triggered independently dvc repro <stage_name>
Force run hyperparameter tuning stage dvc repro -f hp_tune
- Ensures best parameter file will update
Training can be run with dvc repro train
Both stages trigger preprocessing step as dependency

Hyperparameter Run Output

mean_test_score	std_test_score	max_depth	n_estimators	random_state
0.999733	0.000413118	20	5	1993
0.999307	0.000574418	50	5	1993
0.99888	0.000617378	10	5	1993
0.997813	0.00117333	10	4	1993

Changes in Python hyperparameter tuning script

# Save the results of hyperparameter tuning
cv_results = pd.DataFrame(grid_search.cv_results_)
markdown_table = cv_results.to_markdown(index=False)
with open("hp_tuning_results.md", "w") as markdown_file:
  markdown_file.write(markdown_table)

Summary

Hyperparameter tuning route
- Branch name hp_tune/<some-string>
- Make changes to search configuration
- Manually open a PR
  - Force runs DVC pipeline dvc repro -f hp_tune
  - Uses cml pr create to create a new training PR with best parameters
- Force push a commit to training PR to kick off model training job
Manual route
- Branch name train/<some-string>
- Edit best parameters file and commit changes
- Manually open a PR to kick off model training job

Let's practice!

CI/CD for Machine Learning