panther.tuner.SkAutoTuner.Configs package

Submodules

panther.tuner.SkAutoTuner.Configs.LayerConfig module

class panther.tuner.SkAutoTuner.Configs.LayerConfig.LayerConfig(layer_names, params, separate=True, copy_weights=True)

Bases: object

Configuration for a single layer or group of layers to tune.

layer_names

Layer selector (string, list, or dict with criteria)

params

Dictionary of parameter names to ParamSpec or list of values

separate

Whether to tune each layer independently

copy_weights

Whether to copy weights when replacing layers

__init__(layer_names, params, separate=True, copy_weights=True)

Initialize a layer configuration.

Parameters:

• layer_names – Layer selector, can be: - A string: Regex pattern or substring - A list of strings: Multiple patterns or exact names - A dictionary with selection criteria (pattern, type, contains, indices, range)

• params – Dictionary of parameter names and their possible values

• separate – Whether to tune layers separately or jointly

• copy_weights – Whether to copy weights when replacing layers

clone()

Create a deep copy of this LayerConfig.

Returns:

A new LayerConfig instance with the same properties

panther.tuner.SkAutoTuner.Configs.LayerNameResolver module

class panther.tuner.SkAutoTuner.Configs.LayerNameResolver.LayerNameResolver(model, layer_map=None)

Bases: object

Provides intuitive ways to select layers for tuning in large models.

This resolver allows users to specify layers using patterns, types, or indices rather than requiring exact layer names.

__init__(model, layer_map=None)

Initialize the resolver with a model.

Parameters:

model – The neural network model to analyze

resolve(selectors)

Resolve layer name selectors to actual layer names in the model.

Parameters:

selectors –

One or more selectors to match layers. Can be:

• A single string pattern (e.g., “encoder.*attention”)

• A list of string patterns

• A dictionary with keys:

  • ’pattern’: String or list of regex patterns

  • ’type’: Layer type or list of types (e.g., ‘Linear’, ‘Conv2d’)

  • ’contains’: String that layer name must contain

  • ’indices’: Indices to select from matched layers (e.g., [0, 2, 4] for first, third, fifth)

  • ’range’: Range of indices as [start, end, step]

Returns:

List of resolved layer names that match the selectors

panther.tuner.SkAutoTuner.Configs.ParamSpec module

First-class parameter distribution specifications for modern HPO.

Enables expressing categorical, integer, and float distributions explicitly, which can be mapped to Optuna distributions or other HPO frameworks.

class panther.tuner.SkAutoTuner.Configs.ParamSpec.Categorical(choices)

Bases: object

A categorical parameter that takes values from a fixed set of choices.

Examples

Categorical([1, 2, 3]) # Integer choices Categorical([“relu”, “gelu”, “silu”]) # String choices Categorical([True, False]) # Boolean choices

choices

__init__(choices)

class panther.tuner.SkAutoTuner.Configs.ParamSpec.Int(low, high, step=1, log=False)

Bases: object

An integer parameter within a range [low, high].

Parameters:

• low – Lower bound (inclusive)

• high – Upper bound (inclusive)

• step – Step size for discrete values (default: 1)

• log – Whether to sample in log scale (useful for parameters like learning rate)

Examples

Int(1, 100) # Integer from 1 to 100 Int(8, 512, step=8) # Multiples of 8 from 8 to 512 Int(1, 1000, log=True) # Log-scale integer sampling

low

high

step = 1

log = False

__init__(low, high, step=1, log=False)

class panther.tuner.SkAutoTuner.Configs.ParamSpec.Float(low, high, step=None, log=False)

Bases: object

A floating-point parameter within a range [low, high].

Parameters:

• low – Lower bound (inclusive)

• high – Upper bound (inclusive)

• step – Step size for discrete values (None for continuous)

• log – Whether to sample in log scale

Examples

Float(0.0, 1.0) # Continuous float from 0 to 1 Float(1e-5, 1e-1, log=True) # Log-scale float (e.g., learning rate) Float(0.1, 1.0, step=0.1) # Discrete float values

low

high

step = None

log = False

__init__(low, high, step=None, log=False)

panther.tuner.SkAutoTuner.Configs.ParamSpec.is_param_spec(value)

Check if a value is a ParamSpec type.

panther.tuner.SkAutoTuner.Configs.ParamSpec.to_categorical(value)

Convert a list to a Categorical spec.

This is useful for backward compatibility with legacy list-based param definitions.

panther.tuner.SkAutoTuner.Configs.ParamSpec.get_param_choices(spec)

Get discrete choices from a param spec, if applicable.

Returns:

List of choices for Categorical, None for continuous Int/Float ranges.

panther.tuner.SkAutoTuner.Configs.ParamsResolver module

class panther.tuner.SkAutoTuner.Configs.ParamsResolver.ParamsResolver(model, layer_map=None, verbose=False)

Bases: object

A class to resolve parameters for the SkAutoTuner.

__init__(model, layer_map=None, verbose=False)

Initialize the resolver with a model.

Parameters:

• model – The neural network model to analyze

• verbose – Whether to print verbose logs

resolve(config)

Resolves and returns the parameters.

If config.params is a dictionary, it’s returned as is, wrapped in TuningConfigs. If config.params is “auto”, it automatically determines parameters based on layer types and sizes. Otherwise, raises a ValueError.

Returns:

The resolved parameters as TuningConfigs.

panther.tuner.SkAutoTuner.Configs.TuningConfigs module

class panther.tuner.SkAutoTuner.Configs.TuningConfigs.TuningConfigs(configs)

Bases: object

Collection of LayerConfig objects for tuning multiple layer groups.

__init__(configs)

Initialize with a list of layer configurations.

Parameters:

configs – List of LayerConfig objects

add(config)

Add a LayerConfig, returning a new TuningConfigs.

remove(index)

Remove a LayerConfig by index, returning a new TuningConfigs.

replace(index, config)

Replace a LayerConfig at index, returning a new TuningConfigs.

clone()

Create a deep copy of this TuningConfigs.

merge(other)

Merge with another TuningConfigs, returning a new TuningConfigs.

filter(predicate)

Filter configs by predicate, returning a new TuningConfigs.

map(transform)

Apply a transformation to each config, returning a new TuningConfigs.

Module contents

class panther.tuner.SkAutoTuner.Configs.LayerConfig(layer_names, params, separate=True, copy_weights=True)

Bases: object

Configuration for a single layer or group of layers to tune.

layer_names

Layer selector (string, list, or dict with criteria)

params

Dictionary of parameter names to ParamSpec or list of values

separate

Whether to tune each layer independently

copy_weights

Whether to copy weights when replacing layers

__init__(layer_names, params, separate=True, copy_weights=True)

Initialize a layer configuration.

Parameters:

• layer_names – Layer selector, can be: - A string: Regex pattern or substring - A list of strings: Multiple patterns or exact names - A dictionary with selection criteria (pattern, type, contains, indices, range)

• params – Dictionary of parameter names and their possible values

• separate – Whether to tune layers separately or jointly

• copy_weights – Whether to copy weights when replacing layers

clone()

Create a deep copy of this LayerConfig.

Returns:

A new LayerConfig instance with the same properties

class panther.tuner.SkAutoTuner.Configs.TuningConfigs(configs)

Bases: object

Collection of LayerConfig objects for tuning multiple layer groups.

__init__(configs)

Initialize with a list of layer configurations.

Parameters:

configs – List of LayerConfig objects

add(config)

Add a LayerConfig, returning a new TuningConfigs.

remove(index)

Remove a LayerConfig by index, returning a new TuningConfigs.

replace(index, config)

Replace a LayerConfig at index, returning a new TuningConfigs.

clone()

Create a deep copy of this TuningConfigs.

merge(other)

Merge with another TuningConfigs, returning a new TuningConfigs.

filter(predicate)

Filter configs by predicate, returning a new TuningConfigs.

map(transform)

Apply a transformation to each config, returning a new TuningConfigs.

class panther.tuner.SkAutoTuner.Configs.LayerNameResolver(model, layer_map=None)

Bases: object

Provides intuitive ways to select layers for tuning in large models.

This resolver allows users to specify layers using patterns, types, or indices rather than requiring exact layer names.

__init__(model, layer_map=None)

Initialize the resolver with a model.

Parameters:

model – The neural network model to analyze

resolve(selectors)

Resolve layer name selectors to actual layer names in the model.

Parameters:

selectors –

One or more selectors to match layers. Can be:

• A single string pattern (e.g., “encoder.*attention”)

• A list of string patterns

• A dictionary with keys:

  • ’pattern’: String or list of regex patterns

  • ’type’: Layer type or list of types (e.g., ‘Linear’, ‘Conv2d’)

  • ’contains’: String that layer name must contain

  • ’indices’: Indices to select from matched layers (e.g., [0, 2, 4] for first, third, fifth)

  • ’range’: Range of indices as [start, end, step]

Returns:

List of resolved layer names that match the selectors

class panther.tuner.SkAutoTuner.Configs.ParamsResolver(model, layer_map=None, verbose=False)

Bases: object

A class to resolve parameters for the SkAutoTuner.

__init__(model, layer_map=None, verbose=False)

Initialize the resolver with a model.

Parameters:

• model – The neural network model to analyze

• verbose – Whether to print verbose logs

resolve(config)

Resolves and returns the parameters.

If config.params is a dictionary, it’s returned as is, wrapped in TuningConfigs. If config.params is “auto”, it automatically determines parameters based on layer types and sizes. Otherwise, raises a ValueError.

Returns:

The resolved parameters as TuningConfigs.

class panther.tuner.SkAutoTuner.Configs.Categorical(choices)

Bases: object

A categorical parameter that takes values from a fixed set of choices.

Examples

Categorical([1, 2, 3]) # Integer choices Categorical([“relu”, “gelu”, “silu”]) # String choices Categorical([True, False]) # Boolean choices

choices

__init__(choices)

class panther.tuner.SkAutoTuner.Configs.Int(low, high, step=1, log=False)

Bases: object

An integer parameter within a range [low, high].

Parameters:

• low – Lower bound (inclusive)

• high – Upper bound (inclusive)

• step – Step size for discrete values (default: 1)

• log – Whether to sample in log scale (useful for parameters like learning rate)

Examples

Int(1, 100) # Integer from 1 to 100 Int(8, 512, step=8) # Multiples of 8 from 8 to 512 Int(1, 1000, log=True) # Log-scale integer sampling

low

high

step = 1

log = False

__init__(low, high, step=1, log=False)

class panther.tuner.SkAutoTuner.Configs.Float(low, high, step=None, log=False)

Bases: object

A floating-point parameter within a range [low, high].

Parameters:

• low – Lower bound (inclusive)

• high – Upper bound (inclusive)

• step – Step size for discrete values (None for continuous)

• log – Whether to sample in log scale

Examples

Float(0.0, 1.0) # Continuous float from 0 to 1 Float(1e-5, 1e-1, log=True) # Log-scale float (e.g., learning rate) Float(0.1, 1.0, step=0.1) # Discrete float values

low

high

step = None

log = False

__init__(low, high, step=None, log=False)