Skip to content

Patch Dropout Operator¤

Drop rectangular patches from images for regularization.

See Also¤

datarax.operators.modality.image.patch_dropout_operator ¤

PatchDropoutOperator - Operator for patch-based occlusion augmentation.

This operator extends ModalityOperator to provide patch-based dropout (occlusion).

Key Features:

  • Drops random rectangular patches from images
  • Configurable number of patches and patch size
  • Deterministic mode with fixed patch positions
  • Stochastic mode with random patch positions per sample
  • Full JAX compatibility with JIT compilation

Examples:

Basic usage:

config = PatchDropoutOperatorConfig(
    field_key="image",
    num_patches=4,
    patch_size=(8, 8),
    drop_value=0.0
)
op = PatchDropoutOperator(config, rngs=rngs)

logger module-attribute ¤

logger = getLogger(__name__)

PatchDropoutOperatorConfig dataclass ¤

PatchDropoutOperatorConfig(cacheable: bool = False, batch_stats_fn: Callable | Module | None = None, precomputed_stats: dict[str, Any] | None = None, stochastic: bool = False, stream_name: str | None = None, batch_strategy: str = 'vmap', *, field_key: str, target_key: str | None = None, auxiliary_fields: list[str] | None = None, clip_range: tuple[float, float] | None = None, preserve_auxiliary: bool = True, validate_domain_constraints: bool = True, num_patches: int = 4, patch_size: tuple[int, int] = (8, 8), drop_value: float = 0.0)

Bases: ModalityOperatorConfig

Configuration for PatchDropoutOperator.

Extends ModalityOperatorConfig with patch dropout-specific parameters.

Attributes:

Name Type Description
num_patches int

Number of rectangular patches to drop from each image. Default: 4
patch_size tuple[int, int]

Size of each patch as (height, width) tuple. Default: (8, 8)
drop_value float

Value to fill dropped patches with. Typically 0.0 for black or the mean image value. Default: 0.0
clip_range tuple[float, float] | None

Range for clipping output values. None means no clipping. Default: None (patch dropout preserves valid ranges)

num_patches class-attribute instance-attribute ¤

num_patches: int = field(default=4, kw_only=True)

patch_size class-attribute instance-attribute ¤

patch_size: tuple[int, int] = field(default=(8, 8), kw_only=True)

drop_value class-attribute instance-attribute ¤

drop_value: float = field(default=0.0, kw_only=True)

cacheable class-attribute instance-attribute ¤

cacheable: bool = False

batch_stats_fn class-attribute instance-attribute ¤

batch_stats_fn: Callable | Module | None = None

precomputed_stats class-attribute instance-attribute ¤

precomputed_stats: dict[str, Any] | None = None

stochastic class-attribute instance-attribute ¤

stochastic: bool = False

stream_name class-attribute instance-attribute ¤

stream_name: str | None = None

batch_strategy class-attribute instance-attribute ¤

batch_strategy: str = 'vmap'

field_key class-attribute instance-attribute ¤

field_key: str = field(kw_only=True)

target_key class-attribute instance-attribute ¤

target_key: str | None = field(default=None, kw_only=True)

auxiliary_fields class-attribute instance-attribute ¤

auxiliary_fields: list[str] | None = field(default=None, kw_only=True)

clip_range class-attribute instance-attribute ¤

clip_range: tuple[float, float] | None = field(default=None, kw_only=True)

preserve_auxiliary class-attribute instance-attribute ¤

preserve_auxiliary: bool = field(default=True, kw_only=True)

validate_domain_constraints class-attribute instance-attribute ¤

validate_domain_constraints: bool = field(default=True, kw_only=True)

PatchDropoutOperator ¤

PatchDropoutOperator(config: PatchDropoutOperatorConfig, *, rngs: Rngs)

Bases: ModalityOperator

Image patch dropout transformation operator.

Applies patch dropout by randomly dropping rectangular regions from images:

- Selects num_patches random positions
- Replaces each patch with drop_value
- Useful for occlusion robustness training

Supports three modes: 1. Deterministic: Fixed patch positions using fixed seed 2. Stochastic: Per-sample random patch positions from generate_random_params() 3. External params: Accept pre-generated random parameters

The operator works on single elements (H, W, C images) and is composed into batch processing via apply_batch() from the base class.

Examples:

Deterministic patch dropout:

config = PatchDropoutOperatorConfig(
    field_key="image",
    num_patches=4,
    patch_size=(16, 16),
    drop_value=0.0,
    stochastic=False
)
operator = PatchDropoutOperator(config, rngs=nnx.Rngs(0))
result, state, metadata = operator.apply(data, state, metadata)

Stochastic patch dropout with random positions:

config = PatchDropoutOperatorConfig(
    field_key="image",
    num_patches=8,
    patch_size=(8, 8),
    drop_value=0.5,
    stochastic=True
)
operator = PatchDropoutOperator(config, rngs=nnx.Rngs(0))
# Use apply_batch() for automatic random param generation
result, state, metadata = operator.apply_batch(batch_data, state, metadata)

Parameters:

Name Type Description Default
config PatchDropoutOperatorConfig

Configuration for patch dropout operation

 required
rngs Rngs

RNG streams for stochastic operations

 required

config instance-attribute ¤

config: PatchDropoutOperatorConfig = config

rngs instance-attribute ¤

rngs = rngs

name instance-attribute ¤

name = static(name)

stochastic instance-attribute ¤

stochastic = static(stochastic)

stream_name instance-attribute ¤

stream_name = static(stream_name)

generate_random_params ¤

generate_random_params(rng: Array, data_shapes: dict[str, tuple[int, ...]]) -> dict[str, Array]

Generate random patch positions for stochastic mode.

In stochastic mode, this pre-generates random patch positions for the entire batch. This approach avoids RNG state mutations inside vmapped apply().

Parameters:

Name Type Description Default
rng Array

JAX random key

 required
data_shapes dict[str, tuple[int, ...]]

Dictionary mapping field keys to their shapes. Used to determine batch size and image dimensions.

 required

Returns:

Type Description
dict[str, Array]

Dictionary with:

  • "patch_positions": Array of patch top-left positions Shape: (batch_size, num_patches, 2) where last dim is (y, x)

Raises:

Type Description
KeyError

If field_key not in data_shapes

apply ¤

apply(data: dict[str, Array], state: dict[str, Any], metadata: dict[str, Any], random_params: dict[str, Array] | None = None, stats: dict[str, Any] | None = None) -> tuple[dict[str, Array], dict[str, Any], dict[str, Any]]

Apply patch dropout transformation to a single element.

This operates on single elements (e.g., one image of shape [H, W, C]). For batch processing, use apply_batch() which handles random param generation.

Parameters:

Name Type Description Default
data dict[str, Array]

Input data dictionary. Must contain field specified by config.field_key

 required
state dict[str, Any]

Operator state (unused for patch dropout, passed through)

 required
metadata dict[str, Any]

Metadata dictionary (passed through unchanged)

 required
random_params dict[str, Array] | None

Optional random parameters from generate_random_params(). If config.stochastic=True and this is provided, uses random_params["patch_positions"] for patch locations.

 None
stats dict[str, Any] | None

Optional statistics dictionary (unused)

 None

Returns:

Type Description
tuple[dict[str, Array], dict[str, Any], dict[str, Any]]

Tuple of (transformed_data, state, metadata) - transformed_data: Data dict with patches dropped from target field - state: Unchanged state dict - metadata: Unchanged metadata dict
Note

CRITICAL: Always check config.stochastic flag, not whether random_params is None. apply_batch() always passes random_params even in deterministic mode.

get_operation_stats ¤

get_operation_stats() -> dict[str, int]

Get operation statistics.

Note: This method converts JAX arrays to Python ints for introspection. It is intended for use outside of JIT-compiled functions.

Returns:

Type Description
dict[str, int]

Dictionary with 'applied_count' and 'skipped_count'

reset_operation_stats ¤

reset_operation_stats() -> None

Reset operation statistics to zero.

Note: Creates new JAX arrays to reset the counters.

compute_statistics ¤

compute_statistics(data: Any) -> dict[str, Any] | None

Compute statistics from data using batch_stats_fn.

If batch_stats_fn is not configured, returns None. Computed statistics are cached in _computed_stats.

Parameters:

Name Type Description Default
data Any

Input data to compute statistics from

 required

Returns:

Type Description
dict[str, Any] | None

Dictionary of statistics, or None if no batch_stats_fn configured

get_statistics ¤

get_statistics() -> dict[str, Any] | None

Get current statistics.

Returns precomputed_stats if configured (unless reset was called), otherwise returns cached computed statistics, or None if no statistics available.

Returns:

Type Description
dict[str, Any] | None

Dictionary of statistics, or None if no statistics available

set_statistics ¤

set_statistics(stats: dict[str, Any]) -> None

Manually set statistics.

This overwrites any previously computed statistics and clears reset flag.

Parameters:

Name Type Description Default
stats dict[str, Any]

Dictionary of statistics to set

 required

reset_statistics ¤

reset_statistics() -> None

Reset all statistics to None.

This clears both computed statistics and marks that precomputed_stats should be ignored (via internal flag). After reset, get_statistics() will return None until new statistics are set or computed.

reset_cache ¤

reset_cache() -> None

Clear the cache.

Only has effect if cacheable=True in config.

copy ¤

copy(*, config: DataraxModuleConfig | None = None, rngs: Rngs | None = None, name: str | None = None) -> DataraxModule

Create a copy of this module with optional config/parameter changes.

This allows creating a new module instance with modified configuration while preserving other attributes. Useful for hyperparameter tuning.

Parameters:

Name Type Description Default
config DataraxModuleConfig | None

New config (if None, uses current config)

 None
rngs Rngs | None

New RNG state (if None, uses current rngs)

 None
name str | None

New name (if None, uses current name)

 None

Returns:

Type Description
DataraxModule

New module instance with updated parameters

Examples:

Change configuration¤

new_config = DataraxModuleConfig(cacheable=True) new_module = module.copy(config=new_config)

Change name only¤

renamed = module.copy(name="new_name")

Note

Subclasses can override this method to provide more fine-grained control over copying, such as allowing individual config field updates without requiring dataclass replace().

get_state ¤

get_state() -> dict[str, Any]

Get module state for checkpointing.

This method implements the Checkpointable protocol using NNX state management. It extracts all state variables from the module and converts them to a serializable format.

Returns:

Type Description
dict[str, Any]

A dictionary containing the internal state of the component.

set_state ¤

set_state(state: dict[str, Any]) -> None

Restore module state from a checkpoint.

This method implements the Checkpointable protocol using NNX state management. It restores the module state from a serialized format. Restoration is strict: checkpoint structure must match module state.

Parameters:

Name Type Description Default
state dict[str, Any]

A dictionary containing the internal state to restore.

 required

Raises:

Type Description
TypeError

If state is not a dictionary.
ValueError

If checkpoint structure does not match module state.

clone ¤

clone() -> DataraxModule

Create a new instance with the same state as this module.

Uses NNX's clone function for proper deep cloning of all state.

Returns:

Type Description
DataraxModule

A new module instance with the same state.

requires_rng_streams ¤

requires_rng_streams() -> list[str] | None

Get the list of RNG streams required by this module.

Returns:

Type Description
list[str] | None

A list of required RNG stream names, or None if no RNG streams
list[str] | None

are required.

ensure_rng_streams ¤

ensure_rng_streams(stream_names: list[str]) -> None

Ensure that the required RNG streams are available.

Parameters:

Name Type Description Default
stream_names list[str]

A list of available RNG stream names.

 required

Raises:

Type Description
ValueError

If a required RNG stream is not available.

get_output_structure ¤

get_output_structure(sample_data: PyTree, sample_state: PyTree) -> tuple[PyTree, PyTree]

Declare output PyTree structure for vmap axis specification.

Default uses jax.eval_shape to discover structure automatically. Override for efficiency or when eval_shape doesn't work (e.g., data-dependent shapes).

Parameters:

Name Type Description Default
sample_data PyTree

Single element data (not batched)

 required
sample_state PyTree

Single element state (not batched)

 required

Returns:

Type Description
PyTree

Tuple of (output_data_structure, output_state_structure) with None leaves.
PyTree

The structure (keys/nesting) matters, leaf values are ignored.
Example override for operator that adds keys

def get_output_structure(self, sample_data, sample_state): out_data = { **jax.tree.map(lambda _: None, sample_data), "score": None, "alignment": None, } return out_data, sample_state

apply_batch ¤

apply_batch(batch: Batch, stats: dict[str, Any] | None = None) -> Batch

Process entire batch with vmap and optional RNG generation.

This method implements the batch processing logic for both stochastic and deterministic modes. It uses static branching on self.stochastic for JIT compilation efficiency.

The implementation delegates to _vmap_apply() for the shared computational core, then wraps the result in a Batch object.

Parameters:

Name Type Description Default
batch Batch

Input batch (Batch[Element] structure)

 required
stats dict[str, Any] | None

Optional statistics (if None, uses get_statistics())

 None

Returns:

Type Description
Batch

Transformed batch with same structure
Note

This method is concrete (not abstract). Subclasses typically don't override it, but can if they need custom batch processing logic.

output_spec ¤

output_spec(input_spec: PyTree) -> PyTree

Return the operator's output spec given an input spec.

Most operators (normalization, additive noise, simple element-wise transforms) do not change shape; the default returns input_spec unchanged. Shape-changing operators (Resize, Crop, Reshape) MUST override this method.

Parameters:

Name Type Description Default
input_spec PyTree

PyTree of jax.ShapeDtypeStruct describing the input element (matching the upstream DataSourceModule.element_spec() or another operator's output_spec).

 required

Returns:

Type Description
PyTree

PyTree of jax.ShapeDtypeStruct describing the operator's output.
PyTree

By default, equal to input_spec.