ml_switcheroo.frameworks.keras¶

Keras (v3) Framework Adapter.

This module implements the adapter for the Keras 3 framework, supporting multi-backend translation (JAX/Torch/TensorFlow).

It handles: 1. Math: Mapping keras.ops.* (backend-agnostic math). 2. Layers: Mapping keras.layers.*. 3. Discovery: Runtime introspection of the Keras API surface. 4. Ghost Mode: Silent fallback when Keras is not installed. 5. Weight Migration: Loading/saving .h5 or .keras files via h5py.

Attributes¶

keras

Classes¶

KerasAdapter

Adapter for Keras v3 (Multi-backend).

Module Contents¶

ml_switcheroo.frameworks.keras.keras = None[source]¶

class ml_switcheroo.frameworks.keras.KerasAdapter[source]¶

Adapter for Keras v3 (Multi-backend).

Provides definitions for Keras Core Ops, Layers, and Models.

display_name: str = 'Keras'¶

inherits_from: str | None = None¶

ui_priority: int = 25¶

property search_modules: List[str]¶

Returns list of search modules.

If in Ghost Mode, returns empty list to prevent Scaffolder from attempting to import missing modules.

Returns:: Module names.
Return type:: List[str]

property unsafe_submodules: Set[str]¶

Submodules to exclude from recursive scans.

Returns:: Empty set for Keras properly scoped imports.
Return type:: Set[str]

property import_alias: Tuple[str, str]¶

Default import alias.

Returns:: (“keras”, “keras”).
Return type:: Tuple[str, str]

property import_namespaces: Dict[str, ml_switcheroo.frameworks.base.ImportConfig]¶

Namespace mapping for import fixer.

Returns:: Configuration for import injection.
Return type:: Dict[str, ImportConfig]

property discovery_heuristics: Dict[str, List[str]]¶

Regex patterns for scaffold categorization.

Returns:: Patterns mapping tiers to paths.
Return type:: Dict[str, List[str]]

property test_config: Dict[str, str]¶

Templates for test code generation.

Returns:: Test Harness values.
Return type:: Dict[str, str]

property harness_imports: List[str]¶

Imports for verification harness.

Returns:: Empty list.
Return type:: List[str]

get_harness_init_code() → str[source]¶

Init code for verification harness.

Returns:: Empty string.
Return type:: str

get_to_numpy_code() → str[source]¶

Returns code to convert Keras tensors to NumPy.

Returns:: Conversion logic checking for numpy property.
Return type:: str

property supported_tiers: List[ml_switcheroo.enums.SemanticTier]¶

Supported semantic tiers.

Returns:: [ARRAY_API, NEURAL, EXTRAS].
Return type:: List[SemanticTier]

property declared_magic_args: List[str]¶

List of framework-specific magic arguments.

Returns:: Empty list as Keras handles state implicitly.
Return type:: List[str]

property structural_traits: ml_switcheroo.frameworks.base.StructuralTraits¶

Structural transformation rules.

Returns:: Configuration object.
Return type:: StructuralTraits

property plugin_traits: Any¶

Plugin behavior flags.

Returns:: Object defining capabilities.
Return type:: PluginTraits

property definitions: Dict[str, ml_switcheroo.frameworks.base.StandardMap]¶

Static mappings for Keras. Loaded dynamically from frameworks/definitions/keras.json.

Returns:: Definitions.
Return type:: Dict[str, StandardMap]

property rng_seed_methods: List[str]¶

Global seed methods.

Returns:: Names like “set_random_seed”.
Return type:: List[str]

collect_api(category: ml_switcheroo.frameworks.base.StandardCategory) → List[ml_switcheroo.core.ghost.GhostRef][source]¶

Collects API definitions via introspection or snapshot.

Parameters:: category (StandardCategory) – Category to scan.
Returns:: Found items.
Return type:: List[GhostRef]

convert(data: Any) → Any[source]¶

Converts input data to Keras Tensor.

Parameters:: data (Any) – Input data.
Returns:: Keras Tensor or original data.
Return type:: Any

get_serialization_imports() → List[str][source]¶

Imports for saving/loading.

Returns:: Imports.
Return type:: List[str]

get_serialization_syntax(op: str, file_arg: str, object_arg: str | None = None) → str[source]¶

Syntax for saving/loading models.

Parameters:

op (str) – ‘save’ or ‘load’.
file_arg (str) – Path string.
object_arg (Optional[str]) – Object name.

Returns:

Generated python code.

Return type:

str

get_weight_conversion_imports() → List[str][source]¶

Returns imports required for the generated weight migration script.

Returns:: List of import statements.
Return type:: List[str]

get_weight_load_code(path_var: str) → str[source]¶: Returns python code to load a checkpoint. Stub implemented as Keras models contain structure + weights, making raw dict handling tricky.

get_tensor_to_numpy_expr(tensor_var: str) → str[source]¶: Returns python expression string that converts tensor_var from Keras tensor to numpy array.

get_weight_save_code(state_var: str, path_var: str) → str[source]¶

Returns Python code to save the dictionary state_var (mapping flat keys to numpy arrays) to path_var using h5py.

Parameters:

state_var (str) – Variable name of the state dictionary.
path_var (str) – Variable name of the output path.

Returns:

Generated Python code block.

Return type:

str

get_device_syntax(device_type: str, device_index: str | None = None) → str[source]¶

Syntax for device scoping.

Parameters:

device_type (str) – ‘cuda’, ‘cpu’.
device_index (Optional[str]) – Index.

Returns:

Generated code.

Return type:

str

get_device_check_syntax() → str[source]¶

Syntax for checking GPU availability.

Returns:: Logic expression.
Return type:: str

get_rng_split_syntax(rng_var: str, key_var: str) → str[source]¶

Keras handles RNG state internally.

Returns:: “pass”.
Return type:: str

apply_wiring(snapshot: Dict[str, Any]) → None[source]¶

Applies configuration wiring.

Parameters:: snapshot – Snapshot dict.

get_doc_url(api_name: str) → str | None[source]¶

Provides a search URL for Keras documentation as direct API mapping is non-trivial.

Parameters:: api_name (str) – API path.
Returns:: URL.
Return type:: Optional[str]

get_tiered_examples() → Dict[str, str][source]¶

Returns tiered examples.

Returns:: Map of tiers to examples.
Return type:: Dict[str, str]