SlicedWassersteinDistance

Inheritance Diagram

Inheritance diagram of ashpy.metrics.sliced_wasserstein_metric.SlicedWassersteinDistance

class ashpy.metrics.sliced_wasserstein_metric.SlicedWassersteinDistance(model_selection_operator=<built-in function lt>, logdir='/home/docs/checkouts/readthedocs.org/user_builds/ashpy/checkouts/latest/docs/source/log', resolution=128, resolution_min=16, patches_per_image=64, patch_size=7, random_sampling_count=1, random_projection_dim=147, use_svd=False)[source]

Bases: ashpy.metrics.metric.Metric

Sliced Wasserstein Distance. Used as metric in Progressive Growing of GANs 1.

1

Progressive Growing of GANs https://arxiv.org/abs/1710.10196

Methods

__init__([model_selection_operator, logdir, …])

Initialize the Metric.

log(step)

Log the SWD mean and each sub-metric

model_selection(checkpoint, global_step)

Perform model selection for each sub-metric

reset_states()

Reset the state of the metric and the state of each child metric.

update_state(context)

Update the internal state of the metric, using the information from the context object.

Attributes

best_folder

Retrieve the folder used to save the best model when doing model selection.

best_model_sel_file

Retrieve the path to JSON file containing the measured performance of the best model.

logdir

Retrieve the log directory.

metric

Retrieve the tf.keras.metrics.Metric object.

model_selection_operator

Retrieve the operator used for model selection.

name

Retrieve the metric name.

__init__(model_selection_operator=<built-in function lt>, logdir='/home/docs/checkouts/readthedocs.org/user_builds/ashpy/checkouts/latest/docs/source/log', resolution=128, resolution_min=16, patches_per_image=64, patch_size=7, random_sampling_count=1, random_projection_dim=147, use_svd=False)[source]

Initialize the Metric.

Parameters
  • model_selection_operator (typing.Callable) –

    The operation that will be used when model_selection is triggered to compare the metrics, used by the update_state. Any typing.Callable behaving like an operator is accepted.

    Note

    Model selection is done ONLY if an operator is specified here.

  • logdir (str) – Path to the log dir, defaults to a log folder in the current directory.

  • resolution (int) – Image Resolution, defaults to 128

  • resolution_min (int) – Min Resolution achieved by the metric

  • patches_per_image (int) – Number of patches to extract per image per Laplacian level.

  • patch_size (int) – Width of a square patch.

  • random_sampling_count (int) – Number of random projections to average.

  • random_projection_dim (int) – Dimension of the random projection space.

  • use_svd (bool) – experimental method to compute a more accurate distance.

Return type

None

log(step)[source]

Log the SWD mean and each sub-metric

model_selection(checkpoint, global_step)[source]

Perform model selection for each sub-metric

Return type

None

reset_states()[source]

Reset the state of the metric and the state of each child metric.

Return type

None

update_state(context)[source]

Update the internal state of the metric, using the information from the context object.

Parameters

context (ashpy.contexts.GANContext) – An AshPy Context Object that carries all the information the Metric needs.

Return type

None