Use log axis if parameter's distribution is log scale

[toshihikoyanase]

Feature Request

Related to optuna/optuna#666.

Motivation

Currently, the slice plot is always shown in linear scale. The parameters in LogUniformDistribution or IntLogUniformDistribution tend to be plotted around left-edge, and it is sometimes difficult for users to analyze results. On the other hand, Optuna's visualization function automatically uses log axes for such parameters.

For example, the following slice plots are rendered using optuna/examples/pytorch/pytorch_simple.py. Optuna's plot visualizes the peak between 0.001 to 0.01 more clearly than Oputna-dashboard's figure.

(To write the results to the RDBStorage, following fix is required.)

--- a/examples/pytorch/pytorch_simple.py
+++ b/examples/pytorch/pytorch_simple.py
@@ -123,7 +123,7 @@ def objective(trial):
 
 
 if __name__ == "__main__":
-    study = optuna.create_study(direction="maximize")
+    study = optuna.create_study(direction="maximize", storage="sqlite:///foo.db")
     study.optimize(objective, n_trials=100, timeout=600)
 
     pruned_trials = study.get_trials(deepcopy=False, states=[TrialState.PRUNED])

Optuna Dashboard

Optuna

Description

Please convert axes to the log scale when parameters' distributions are log scale.