Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Use vectorization for categorical distance #5147

Merged
merged 8 commits into from
Jan 9, 2024
Merged

Use vectorization for categorical distance #5147

Changes from 5 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
3257fa2
Use vectorization for categorical distance
nabenabe0928 Dec 9, 2023
5d6e590
Reduce time complexity from O(C**2) to O(C*U)
nabenabe0928 Dec 9, 2023
3b9db5c
Debug by test cases
nabenabe0928 Dec 9, 2023
7112b9a
Fix a bug for observations.size == 0
nabenabe0928 Dec 9, 2023
c1fe437
Fix an error for consider_prior=True in the cat dist calc
nabenabe0928 Dec 9, 2023
d7e1dda
Apply the suggestions by wang-san
nabenabe0928 Dec 22, 2023
25ba574
Make len(observations)==0 early-return in categorical dist
nabenabe0928 Dec 22, 2023
f600cd8
Refactor slightly
nabenabe0928 Dec 25, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
36 changes: 20 additions & 16 deletions optuna/samplers/_tpe/parzen_estimator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -185,29 +185,33 @@ def _calculate_categorical_distributions(
search_space: CategoricalDistribution,
parameters: _ParzenEstimatorParameters,
) -> _BatchedDistributions:
consider_prior = parameters.consider_prior or len(observations) == 0
choices = search_space.choices
n_choices = len(choices)
n_samples = observations.size + (parameters.consider_prior or observations.size == 0)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: the name n_samples might not be appropriate, because this includes prior. Maybe n_mixture may be better?


assert parameters.prior_weight is not None
weights = np.full(
shape=(len(observations) + consider_prior, len(search_space.choices)),
fill_value=parameters.prior_weight / (len(observations) + consider_prior),
shape=(n_samples, n_choices),
fill_value=parameters.prior_weight / n_samples,
)

if param_name in parameters.categorical_distance_func:
observed_indices = observations.astype(int)
if observed_indices.size == 0:
# Do nothing.
pass
elif param_name in parameters.categorical_distance_func:
# TODO(nabenabe0928): Think about how to handle combinatorial explosion.
# The time complexity is O(n_choices * used_indices.size), so n_choices cannot be huge.
used_indices, rev_indices = np.unique(observed_indices, return_inverse=True)
dist_func = parameters.categorical_distance_func[param_name]
for i, observation in enumerate(observations.astype(int)):
dists = [
dist_func(search_space.choices[observation], search_space.choices[j])
for j in range(len(search_space.choices))
]
exponent = -(
(np.array(dists) / max(dists)) ** 2
* np.log((len(observations) + consider_prior) / parameters.prior_weight)
* (np.log(len(search_space.choices)) / np.log(6))
)
weights[i] = np.exp(exponent)
dists = np.array([[dist_func(choices[i], c) for c in choices] for i in used_indices])
max_dists = np.max(dists, axis=1)
coef = np.log(n_samples / parameters.prior_weight) * np.log(n_choices) / np.log(6)
categorical_weights = np.exp(-((dists / max_dists[:, np.newaxis]) ** 2) * coef)
weights[: observed_indices.size] = categorical_weights[rev_indices]
else:
weights[np.arange(len(observations)), observations.astype(int)] += 1
weights[np.arange(observed_indices.size), observed_indices] += 1

weights /= weights.sum(axis=1, keepdims=True)
return _BatchedCategoricalDistributions(weights)

Expand Down