Review Hyperopt trials

  • 5 minutes

When you use Hyperopt to try multiple hyperparameter combinations, you can review the details of each trial. You can review these details in two ways:

  • View the MLflow run for each trial.
  • Use the Trials class to capture run details.

View the MLflow run for each trial

In Azure Databricks, calls to the Hyperopt fmin function automatically generate MLflow experiment runs that you can view in the Azure Databricks portal; providing you with an easy way to view the full set of hyperparameters and metrics for each trial, as shown here:

Screenshot of the MLflow run for a Hyperopt trial.

Use the Trials class

Hyperopt includes a Trials class that logs details for each trial that is run during an fmin function call. The following code example shows how to use the Trials class.

from hyperopt import Trials

# Create a Trials object to track each run
trial_runs = Trials()

argmin = fmin(
  fn=objective,
  space=search_space,
  algo=algo,
  max_evals=100,
  trials=trial_runs)

print("Best param values: ", argmin)

# Get details from each trial run
print ("trials:")
for trial in trial_runs.trials:
    print ("\n", trial)

The full detail output for a collection of trials resembles this example:

{'state': 2, 'tid': 0, 'spec': None, 'result': {'loss': -0.8571428571428571, 'status': 'ok'}, 'misc': {'tid': 0, 'cmd': ('domain_attachment', 'FMinIter_Domain'), 'workdir': None, 'idxs': {'Iterations': [0], 'Regularization': [0]}, 'vals': {'Iterations': [1], 'Regularization': [0.4965634473237057]}}, 'exp_key': None, 'owner': None, 'version': 0, 'book_time': datetime.datetime(2023, 3, 17, 22, 45, 24, 993000), 'refresh_time': datetime.datetime(2023, 3, 17, 22, 46, 30, 374000)}

 {'state': 2, 'tid': 1, 'spec': None, 'result': {'loss': -0.8857142857142857, 'status': 'ok'}, 'misc': {'tid': 1, 'cmd': ('domain_attachment', 'FMinIter_Domain'), 'workdir': None, 'idxs': {'Iterations': [1], 'Regularization': [1]}, 'vals': {'Iterations': [9], 'Regularization': [0.8446551490616772]}}, 'exp_key': None, 'owner': None, 'version': 0, 'book_time': datetime.datetime(2023, 3, 17, 22, 46, 30, 379000), 'refresh_time': datetime.datetime(2023, 3, 17, 22, 47, 34, 155000)}

 {'state': 2, 'tid': 2, 'spec': None, 'result': {'loss': -0.9523809523809523, 'status': 'ok'}, 'misc': {'tid': 2, 'cmd': ('domain_attachment', 'FMinIter_Domain'), 'workdir': None, 'idxs': {'Iterations': [2], 'Regularization': [2]}, 'vals': {'Iterations': [9], 'Regularization': [0.3931915704555482]}}, 'exp_key': None, 'owner': None, 'version': 0, 'book_time': datetime.datetime(2023, 3, 17, 22, 47, 34, 160000), 'refresh_time': datetime.datetime(2023, 3, 17, 22, 48, 45, 986000)}

 ...

Tip

For more information about the details recorded by the Trials class, see the Hyperopt documentation.