def fixed(ctx, batch_size, filename, rows, cols, plot, verbose, **kwargs):
"""View/save images of dataset given a fixed latent factor."""
dataset = ctx.obj['dataset']
add_gin(ctx, 'config', ['evaluate/dataset/{}.gin'.format(dataset)])
parse(ctx)
with gin.unlock_config():
gin.bind_parameter('disentangled.visualize.show.output.show_plot',
plot)
gin.bind_parameter('disentangled.visualize.show.output.filename',
filename)
if rows is not None:
gin.bind_parameter('disentangled.visualize.fixed_factor_data.rows',
rows)
if cols is not None:
gin.bind_parameter('disentangled.visualize.fixed_factor_data.cols',
cols)
num_values_per_factor = disentangled.dataset.get(
dataset).num_values_per_factor
dataset = disentangled.dataset.get(dataset).supervised()
fixed, _ = disentangled.metric.utils.fixed_factor_dataset(
dataset, batch_size, num_values_per_factor)
disentangled.visualize.fixed_factor_data(fixed,
rows=gin.REQUIRED,
cols=gin.REQUIRED,
verbose=verbose)
def gui(ctx, **kwargs):
"""Qualitiative interactive evaluation of disentanglement."""
add_gin(ctx, "config", ["evaluate/visual/gui.gin"])
parse(ctx, set_seed=True)
dataset = ctx.obj["dataset"].pipeline()
disentangled.visualize.gui(ctx.obj["model"], dataset)
def latent1d(ctx, rows, cols, plot, filename, **kwargs):
"""Latent space traversal in 1D."""
add_gin(ctx, "config", ["evaluate/visual/latent1d.gin"])
parse(ctx, set_seed=True)
with gin.unlock_config():
gin.bind_parameter("disentangled.visualize.show.output.show_plot",
plot)
if filename is not None:
gin.bind_parameter("disentangled.visualize.show.output.filename",
filename)
if rows is not None:
gin.bind_parameter("disentangled.visualize.traversal1d.dimensions",
rows)
if cols is not None:
gin.bind_parameter("disentangled.visualize.traversal1d.steps",
cols)
dataset = ctx.obj["dataset"].pipeline()
disentangled.visualize.traversal1d(
ctx.obj["model"],
dataset,
dimensions=gin.REQUIRED,
offset=gin.REQUIRED,
skip_batches=gin.REQUIRED,
steps=gin.REQUIRED,
)
def loglikelihood(ctx, **kwargs):
"""The logarithmic likelihood of the input given the representation."""
add_gin(ctx, "config", ["metric/loglikelihood.gin"])
parse(ctx, set_seed=True)
metric = disentangled.metric.loglikelihood(
ctx.obj["model"],
dataset=gin.REQUIRED,
)
breakpoint()
disentangled.metric.log_metric(metric,
name=ctx.obj["model_str"],
metric_name=gin.REQUIRED)
def collapsed(ctx, **kwargs):
"""Number of latent dimensions collapsed to prior."""
add_gin(ctx, "config", ["metric/collapsed.gin"])
parse(ctx, set_seed=True)
metric = disentangled.metric.collapsed(
ctx.obj["model"],
dataset=gin.REQUIRED,
tolerance=gin.REQUIRED,
)
disentangled.metric.log_metric(metric,
name=ctx.obj["model_str"],
metric_name=gin.REQUIRED)
def mig(ctx, **kwargs):
"""Mutual Information GAP.
Quantitative disentanglement metric."""
add_gin(ctx, "config", ["metric/mig.gin"])
parse(ctx, set_seed=True)
metric = disentangled.metric.mutual_information_gap(
ctx.obj["model"],
dataset=gin.REQUIRED,
encoding_dist=gin.REQUIRED,
)
disentangled.metric.log_metric(metric,
metric_name=gin.REQUIRED,
name=ctx.obj["model_str"])
def dmig(ctx, **kwargs):
"""Discrete Mutual Information Gap.
Quantitative disentanglement metric."""
add_gin(ctx, "config", ["metric/dmig.gin"])
parse(ctx, set_seed=True)
metric = disentangled.metric.discrete_mutual_information_gap(
ctx.obj["model"],
dataset=gin.REQUIRED,
points=gin.REQUIRED,
batch_size=gin.REQUIRED)
disentangled.metric.log_metric(metric,
metric_name=gin.REQUIRED,
name=ctx.obj["model_str"])
def gini_index(ctx, **kwargs):
"""Quantitative sparsity metric of representation."""
add_gin(ctx, "config", ["metric/gini.gin"])
parse(ctx, set_seed=True)
metric = disentangled.metric.gini_index(
ctx.obj["model"],
dataset=gin.REQUIRED,
points=gin.REQUIRED,
batch_size=gin.REQUIRED,
tolerance=gin.REQUIRED,
)
disentangled.metric.log_metric(metric,
name=ctx.obj["model_str"],
metric_name=gin.REQUIRED)
def factorvae_score(ctx, **kwargs):
"""The FactorVAE-score.
Quantitative disentanglement metric."""
add_gin(ctx, "config", ["metric/factorvae_score.gin"])
parse(ctx, set_seed=True)
metric = disentangled.metric.factorvae_score(
ctx.obj["model"],
dataset=gin.REQUIRED,
training_points=gin.REQUIRED,
test_points=gin.REQUIRED,
tolerance=gin.REQUIRED,
)
disentangled.metric.log_metric(metric,
name=ctx.obj["model_str"],
metric_name=gin.REQUIRED)
def evaluate(ctx, model, path, **kwargs):
"""Interface for evaluating models.
Quantitative and qualitative evaluation methods are provided.
"""
ctx.obj["model"] = disentangled.model.utils.load(path, model)
ctx.obj["model_str"] = model
method, dataset = model.split("/")
ctx.obj["dataset"] = disentangled.dataset.get(dataset)
ctx.obj["method_str"] = method
ctx.obj["dataset_str"] = dataset
add_gin(ctx, "gin_param", ["HP_SWEEP_VALUES=None"])
add_gin(ctx, "gin_param", ["log_metric.path='{}'".format(path)])
add_gin(ctx, "config", ["evaluate/evaluate.gin"])
add_gin(ctx, "config", ["evaluate/dataset/" + dataset + ".gin"])
add_gin(ctx, "config", ["evaluate/model/" + method + ".gin"])
def latent2d(ctx, rows, cols, plot, filename, **kwargs):
"""Latent space traversal in 2D."""
add_gin(ctx, "config", ["evaluate/visual/latent2d.gin"])
parse(ctx, set_seed=True)
with gin.unlock_config():
gin.bind_parameter("disentangled.visualize.show.output.show_plot",
plot)
gin.bind_parameter("disentangled.visualize.show.output.filename",
filename)
if rows is not None:
gin.bind_parameter("disentangled.visualize.traversal2d.rows", rows)
if cols is not None:
gin.bind_parameter("disentangled.visualize.traversal2d.cols", cols)
dataset = ctx.obj["dataset"].pipeline()
disentangled.visualize.traversal2d(ctx.obj["model"], dataset)
def train(ctx,
model,
config,
gin_param,
gin_file,
path,
checkpoint,
overwrite=False,
print_operative=False):
"""Interface for training models by using syntax MODEL/DATASET.
Training is executed with default config values found in config/train directory.
See config files or print operative configuration for configurable parameters.
\b
Example of modifying default config:
disentangled train BetaVAE/DSprites -p iterations=1000
\b
Available Models:
BetaVAE
BetaSVAE
BetaTCVAE
FactorVAE
Available Datasets:
Shapes3d
DSprites
CelebA
"""
add_gin(ctx, 'config', ['train/{}.gin'.format(model)], insert=True)
if overwrite is True:
add_gin(ctx,
'gin_param'['disentangled.model.utils.save.overwrite=True'])
parse(ctx)
disentangled.training.run_training(model=gin.REQUIRED,
dataset=gin.REQUIRED,
iterations=gin.REQUIRED,
path=path,
checkpoint=checkpoint)
if print_operative:
print(gin.operative_config_str())
def examples(ctx, filename, rows, cols, plot, **kwargs):
"""View or save example images of dataset."""
dataset = ctx.obj['dataset']
add_gin(ctx, 'config', ['evaluate/dataset/{}.gin'.format(dataset)])
parse(ctx)
with gin.unlock_config():
gin.bind_parameter('disentangled.visualize.show.output.show_plot',
plot)
gin.bind_parameter('disentangled.visualize.show.output.filename',
filename)
if rows is not None:
gin.bind_parameter('disentangled.visualize.data.rows', rows)
if cols is not None:
gin.bind_parameter('disentangled.visualize.data.cols', cols)
dataset = disentangled.dataset.get(dataset).pipeline()
disentangled.visualize.data(dataset, rows=gin.REQUIRED, cols=gin.REQUIRED)
def cli(ctx, **kwargs):
"""Train and evaluate disentangled representation learning models."""
add_gin(ctx, "config", ["config.gin"])