def configs() -> Configs:
c = Configs()
c.add(
name="target_gflops",
type=float,
default=2.0,
strategy="constant",
description="Target (Giga) Floating Point Operations per Second.",
)
return c
def configs():
c = Configs()
c.add(
name="hidden_dim",
type=int,
default=128,
strategy="choice",
choices=[128, 256, 512, 1024],
description="Number of hiden units.",
)
return c
def configs() -> Configs:
c = Configs()
c.add(
name="extract_features_after_layer",
default="",
type=str,
description=(
"Layer name after which to extract features. "
"Nested layers may be selected using dot-notation, "
"e.g. `block.subblock.layer1`"
),
)
return c
def configs() -> Configs:
c = Configs()
c.add(
name="learning_rate",
type=float,
default=0.001,
choices=(5e-7, 5e-1),
strategy="loguniform",
description="Learning rate.",
)
c.add(
name="optimizer_beta1",
type=float,
default=0.9,
choices=(0, 0.999),
strategy="uniform",
description="Beta 1.",
)
c.add(
name="optimizer_beta2",
type=float,
default=0.999,
choices=(0, 0.99999),
strategy="uniform",
description="Beta 2.",
)
c.add(
name="weight_decay",
type=float,
default=1e-2,
choices=(1e-6, 1e-1),
strategy="loguniform",
description="Weight decay.",
)
return c
def configs() -> Configs:
c = Configs()
c.add(
name="metric_selection",
default=default_config,
type=str,
strategy="constant",
description="Selection key for MetricSelector.",
choices=list(mapping.keys()),
)
for Metric in metric_set:
if hasattr(Metric, "configs"):
c += Metric.configs()
return c
def configs() -> Configs:
c = Configs()
c.add(
name="optimization_metric",
default="loss",
type=str,
choices=MetricMixin.metric_names(),
description=
"Name of the performance metric that should be optimized",
)
c.add(
name="test_ensemble",
type=int,
default=0,
strategy="constant",
description=
"Flag indicating whether the test dataset should yield a clip ensemble.",
)
c.add(
name="gpus",
type=str,
default=None,
strategy="constant",
description=
"Which gpus should be used. Can be either the number of gpus (e.g. '2') or a list of gpus (e.g. ('0,1').",
)
c.add(
name="loss",
type=str,
default="cross_entropy",
choices=loss_names,
strategy="constant",
description="Loss function used during optimisation.",
)
return c
def configs(self) -> Configs:
c = Configs()
c.add(
name="trials",
default=30,
type=int,
description="Number of trials in the hyperparameter search",
)
c.add(
name="gpus_per_trial",
default=0,
type=int,
description="Number of GPUs per trail in the hyperparameter search",
)
c.add(
name="optimization_metric",
default="loss",
type=str,
choices=self.Module.metric_names(),
description="Name of the performance metric that should be optimized",
)
c.add(
name="from_hparam_space_file",
default=None,
type=str,
description="Path to file with specification for the search space during hyper-parameter optimisation",
)
return c
def configs() -> Configs:
c = Configs()
c.add(
name="learning_rate",
type=float,
default=0.1,
choices=(5e-2, 5e-1),
strategy="loguniform",
description="Learning rate.",
)
c.add(
name="weight_decay",
type=float,
default=1e-5,
choices=(1e-6, 1e-3),
strategy="loguniform",
description="Weight decay.",
)
c.add(
name="momentum",
type=float,
default=0.9,
choices=(0, 0.999),
strategy="uniform",
description="Momentum.",
)
return c
def configs() -> Configs:
c = Configs()
c.add(
name="unfreeze_from_epoch",
type=int,
default=-1,
description=
"Number of epochs to wait before starting gradual unfreeze. If -1, unfreeze is omitted.",
)
c.add(
name="unfreeze_layers_must_include",
type=str,
default="",
description=
"String that must be contained in layer names which should be unfrozen. If empty, this feature is disabled.",
)
c.add(
name="unfreeze_epoch_step",
type=int,
default=1,
description="Number of epochs to train before next unfreeze.",
)
c.add(
name="unfreeze_layers_initial",
type=int,
default=1,
strategy="choice",
description=
"Number layers to unfreeze initially. If `-1`, it will be equal to total_layers",
)
c.add(
name="unfreeze_layer_step",
type=int,
default=1,
description=
"Number additional layers to unfreeze at each unfreeze step. If `-1`, all layers are unfrozon after a step",
)
c.add(
name="unfreeze_layers_max",
type=int,
default=-1,
description=
"Maximum number of layers to unfreeze. If `-1`, it will be equal to total_layers",
)
return c