Are oracles trained on normalized x or unnormalized x?

[yuanqidu]

Dear authors,

Thanks for open sourcing this library. I'm trying to understand if the oracles were trained on normalized or unnormalized x. Thank you for your help. If yes, what dataset they were used for normalizing them? Since when we use the oracle, we can only normalize them by the provided dataset (smaller), would this cause problems as input to the oracle (if normalized on a larger dataset)?

Best,
Yuanqi

[brandontrabucco]

Hello yuanqidu,

Thanks for your interest in design-bench!

The benchmark keeps two copies of an MBO dataset---a private internal version that matches the format expected by the oracle model, and a public version exposed to the user for benchmarking their own algorithms.

The code for this separation is located in four functions:

dataset_to_oracle_x:

design-bench/design_bench/oracles/oracle_builder.py

Line 261 in e529395

def dataset_to_oracle_x(self, x_batch, dataset=None):

dataset_to_oracle_y:

design-bench/design_bench/oracles/oracle_builder.py

Line 315 in e529395

def dataset_to_oracle_y(self, y_batch, dataset=None):

oracle_to_dataset_x:

design-bench/design_bench/oracles/oracle_builder.py

Line 359 in e529395

def oracle_to_dataset_x(self, x_batch, dataset=None):

oracle_to_dataset_y:

design-bench/design_bench/oracles/oracle_builder.py

Line 414 in e529395

def oracle_to_dataset_y(self, y_batch, dataset=None):

This boilerplate code handles the conversion from oracle format to public format.

For a particular task from design-bench, you can find out what format the oracle expects by checking task.oracle.expect_normalized_x and task.oracle.expect_normalized_y.

Design-Bench internally manages these to ensure the format is correct when task.predict(xs) is called, where xs is a batch of designs you want to evaluate.

[brandontrabucco]

This section of the README may help too:

import design_bench
task = design_bench.make('TFBind8-Exact-v0')

# convert x to logits of a categorical probability distribution
task.map_to_logits()
discrete_x = task.to_integers(task.x)

# normalize the inputs to have zero mean and unit variance
task.map_normalize_x()
original_x = task.denormalize_x(task.x)

# normalize the outputs to have zero mean and unit variance
task.map_normalize_y()
original_y = task.denormalize_y(task.y)

# remove the normalization applied to the outputs
task.map_denormalize_y()
normalized_y = task.normalize_y(task.y)

# remove the normalization applied to the inputs
task.map_denormalize_x()
normalized_x = task.normalize_x(task.x)

# convert x back to integers
task.map_to_integers()
continuous_x = task.to_logits(task.x)