Conformance tooling - Torch Lattice

torch_lattice_conformance.cli.main(argv=None)[source]¶

Return type:: None
Parameters:: argv (Sequence[str] | None)

class torch_lattice_conformance.generate.SparseInputSpec(channels, batch_size, spatial_shape, rows_per_batch, dtype=torch.float32, stride=(1, 1, 1))[source]¶

Bases: object

Parameters:

channels (int)
batch_size (int)
spatial_shape (tuple[int, int, int])
rows_per_batch (tuple[int, ...])
dtype (dtype)
stride (tuple[int, int, int])

channels: int¶

batch_size: int¶

spatial_shape: tuple[int, int, int]¶

rows_per_batch: tuple[int, ...]¶

dtype: dtype = torch.float32¶

stride: tuple[int, int, int] = (1, 1, 1)¶

class torch_lattice_conformance.generate.FuzzCase(name, family, seed, output_kind, input_mode, rtol, atol, metadata)[source]¶

Bases: object

Parameters:

name (str)
family (Literal['sparse_classifier', 'sparse_feature_chain', 'sparse_branch', 'target_branch', 'transpose_chain', 'generative_transpose', 'point_voxel', 'quantized_classifier'])
seed (int)
output_kind (Literal['dense', 'sparse'])
input_mode (Literal['sparse_kwargs', 'dense_kwargs'])
rtol (float)
atol (float)
metadata (dict[str, Any])

name: str¶

family: Literal['sparse_classifier', 'sparse_feature_chain', 'sparse_branch', 'target_branch', 'transpose_chain', 'generative_transpose', 'point_voxel', 'quantized_classifier']¶

seed: int¶

output_kind: Literal['dense', 'sparse']¶

input_mode: Literal['sparse_kwargs', 'dense_kwargs']¶

rtol: float¶

atol: float¶

metadata: dict[str, Any]¶

class torch_lattice_conformance.generate.SparseChainModel(stages, *, global_pool=None, head=None)[source]¶

Bases: Module

Parameters:

stages (list[nn.Module])
global_pool (nn.Module | None)
head (nn.Module | None)

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Parameters:: x (SparseTensor)

class torch_lattice_conformance.generate.SparseBranchModel(left, right, tail, *, merge, join)[source]¶

Bases: Module

Parameters:

left (nn.Module)
right (nn.Module)
tail (nn.Module)
merge (Literal['add', 'cat'])
join (Literal['inner', 'outer'])

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: SparseTensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.generate.TargetBranchModel(pre, target_conv)[source]¶

Bases: Module

Parameters:

pre (nn.Module)
target_conv (spnn.Conv3d)

forward(x, target)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

SparseTensor

Parameters:

x (SparseTensor)
target (SparseTensor)

class torch_lattice_conformance.generate.PointVoxelRoundTrip(*, voxel_size, origin, reduction, interpolation)[source]¶

Bases: Module

Parameters:

voxel_size (tuple[float, float, float])
origin (tuple[float, float, float])
reduction (Literal['sum', 'mean'])
interpolation (Literal['nearest', 'linear'])

forward(points, features, batch_indices, active_rows)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

Tensor

Parameters:

points (Tensor)
features (Tensor)
batch_indices (Tensor)
active_rows (Tensor)

torch_lattice_conformance.generate.main()[source]¶

Return type:: None

torch_lattice_conformance.e2e.main(argv=None)[source]¶

Return type:: None
Parameters:: argv (Sequence[str] | None)

torch_lattice_conformance.e2e.generate(root)[source]¶

Write the complete deterministic cross-runtime fixture corpus.

Return type:: None
Parameters:: root (Path)

class torch_lattice_conformance.e2e.SparseClassifier[source]¶

Bases: Module

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: Tensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.e2e.QuantizedClassifier[source]¶

Bases: Module

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: Tensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.e2e.TargetBranch[source]¶

Bases: Module

forward(x, target)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

SparseTensor

Parameters:

x (SparseTensor)
target (SparseTensor)

class torch_lattice_conformance.e2e.PointVoxel(*args, **kwargs)[source]¶

Bases: Module

Parameters:

args (Any)
kwargs (Any)

forward(points, features, batch_indices, active_rows)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

Tensor

Parameters:

points (Tensor)
features (Tensor)
batch_indices (Tensor)
active_rows (Tensor)

class torch_lattice_conformance.e2e.TransposeConvolution[source]¶

Bases: Module

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: SparseTensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.e2e.GenerativeTransposeConvolution[source]¶

Bases: Module

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: SparseTensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.e2e.NormalizedConvolution[source]¶

Bases: Module

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: SparseTensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.e2e.CanonicalKernelLayout[source]¶

Bases: Module

Non-cubic row-distinct kernel used to lock the artifact ABI.

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: SparseTensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.e2e.PoolTranspose[source]¶

Bases: Module

forward(source, target)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

SparseTensor

Parameters:

source (SparseTensor)
target (SparseTensor)

class torch_lattice_conformance.e2e.TargetTransposeConvolution[source]¶

Bases: Module

forward(source, target)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

SparseTensor

Parameters:

source (SparseTensor)
target (SparseTensor)

class torch_lattice_conformance.e2e.TrilinearUpsample[source]¶

Bases: Module

forward(source, target)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

SparseTensor

Parameters:

source (SparseTensor)
target (SparseTensor)

class torch_lattice_conformance.e2e.SparseReindex(*args, **kwargs)[source]¶

Bases: Module

Parameters:

args (Any)
kwargs (Any)

forward(source, target)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:

SparseTensor

Parameters:

source (SparseTensor)
target (SparseTensor)

class torch_lattice_conformance.e2e.GameleonReproductionBlock[source]¶

Bases: Module

Compact training/export gate for Gameleon’s sparse decoder routes.

forward(x)[source]¶

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Return type:: SparseTensor
Parameters:: x (SparseTensor)

class torch_lattice_conformance.migration.CompatCase(family, seed, output_kind)[source]¶

Bases: object

Parameters:

family (Literal['pointwise_chain', 'branch_add', 'branch_cat', 'global_pool', 'batchnorm_chain', 'spatial_subm_mapping', 'stride2_forward'])
seed (int)
output_kind (Literal['sparse', 'dense'])

family: Literal['pointwise_chain', 'branch_add', 'branch_cat', 'global_pool', 'batchnorm_chain', 'spatial_subm_mapping', 'stride2_forward']¶

seed: int¶

output_kind: Literal['sparse', 'dense']¶

torch_lattice_conformance.migration.main()[source]¶

Return type:: None

Conformance tooling¶