Skip to main content

← HomOps Reference

HomUnsqueeze

Insert a size-1 axis into an encrypted tensor's shape.

UqShape
RunsServer
Changes shapeYes - adds a size-1 axis
Changes scaleNo - unchanged
Levels spentNone
RotatesNo
KeysNone

What it does​

HomUnsqueeze inserts a size-1 dimension at a chosen axis of the ciphertext's logical shape. It is layout-only: no arithmetic on the encrypted values, matching PyTorch's unsqueeze(dim) on encrypted tensors.
Reach for it when a later op expects an extra axis for broadcasting, or when you want to pair with HomSqueeze around a reduction that keeps or drops dims.

Signature​

HomUnsqueeze(dim)
No set_data.

Parameters​

ParameterTypeDefaultDescription
dimintrequiredAxis index where size 1 is inserted. Negative indexing is supported.
Rules that depend on the value
SettingRuleIf you break it
dimMust be a valid insert position for the current rank.Compilation fails.
n axisIf you insert at or before the n axis, that axis index may shift in the new shape.Not an error - but see Concepts.

Requirements​

  • Valid insert position. dim must be legal for the current rank.
  • No keys, no levels. Layout only.

Shape effect - yes​

Rule: insert 1 at dim.
InputdimOutput
(128, 512)0(1, 128, 512)
(1, 128, 512)2(1, 128, 1, 512)

Scale effect - no​

Scale is unchanged.

Level budget​

None. Unsqueeze does not drop a prime from the modulus chain.

Keys​

None.

Example​

Insert a size-1 axis before a later broadcast or reduction:
pipeline.pyPYTHON
from lattica_build.operators import HomUnsqueeze
from lattica_build.base_classes.hom_pipeline import HomomorphicPipeline
from lattica_build.operators.composite.sequential import SequentialHomOp

pipeline = HomomorphicPipeline(
    hom=SequentialHomOp(
        HomUnsqueeze(2),   # (1, 128, 512) → (1, 128, 1, 512)
    ),
    input_shape=(1, 128, 512),
)

See also​