Skip to main content

← HomOps Reference

HomReshape

Change how your data is laid out, without touching the encrypted values.

RhShape
RunsServer
Changes shapeYes: new layout
Changes scaleNo: unchanged
Levels spent0
KeysNone

What it does​

HomReshape gives data a new shape (flattening, splitting, or merging axes) without changing its values. On the server, its backend applies the target shape to the ciphertext.
It can appear anywhere in a pipeline, including the encrypted hom section. It is not client-only; use client sections only when that placement suits the surrounding work.
Typical uses:
Flatten an image for a linear layer: (1, 28, 28) → (784,).
Restore an output shape after decryption: (10, 1) → (10,) in client_post.
Bridge convolution feature maps into a matmul-ready vector: (5, 8, 8) → (320,).

Signature​

HomReshape(dims)
HomReshape has no set_data; it carries no weights, only the target shape.

Parameters​

ParameterTypeDefaultDescription
dimsshapenoneThe target shape. One dimension may be -1, which is inferred from the total number of elements.
Rules that depend on the value
SettingRuleIf you break it
dimsThe total element count must be preserved: the product of dims must equal the product of the input shape (after any -1 is inferred).Compilation fails: the reshape is incompatible with the input size.
dimsAt most one dimension may be -1.Compilation fails.
dimsThe reshape may change which axis is the n axis, since the packing picks the axis that wastes the fewest slots. Worth knowing if a later op expects a particular n axis.Not an error, but see Concepts.

Requirements​

  • Element count preserved. Input and output must hold the same number of elements.
  • Any pipeline placement. It may run in hom on the server as well as in client sections.

Shape effect: yes​

The data is re-viewed under dims, and the n axis is re-inferred for the new layout.
InputdimsOutput
(1, 28, 28)(784,)(784,) flatten an image
(784,)(28, 28)(28, 28) restore a grid
(5, 8, 8)(320,)(320,) feature maps to a vector
(10, 1)(10,)(10,) drop a trailing axis
(512, 1)(-1,)(512,) with -1 inferred
One input passes through the reshape:
in(1, 28, 28)@scale S
client_preHomReshape((784,))
out(784,)@scale S

before encrypt: same values, new layout

Shape and the n axis are covered on Concepts.

Scale effect: no​

A reshape is a layout view only. Scale is unchanged.

Level budget​

Zero. Server-side reshape changes ciphertext layout without spending modulus-chain levels.

Keys​

None. Server-side reshape needs no evaluation key.

Example​

Flatten a 28×28 image before the first linear layer. This reshape may run in the encrypted hom section:
pipeline.pyPYTHON
from lattica_build.base_classes.hom_pipeline import HomomorphicPipeline
from lattica_build.operators.composite.sequential import SequentialHomOp
from lattica_build.operators import HomLinear, HomSquare, HomReshape

pipeline = HomomorphicPipeline(
    client_pre=[HomReshape((28 * 28,))],
    hom=SequentialHomOp(
        HomLinear((50, 28 * 28), bias=False, with_modswitch=False),
        HomSquare(with_modswitch=False),
        HomLinear((10, 50), bias=False, with_modswitch=False),
    ),
    input_shape=(1, 28, 28),
)
A reshape in client_post puts the decrypted result back into its expected shape:
pipeline.pyPYTHON · CLIENT_POST
client_post=[HomReshape((10,))]   # runs after decryption, on your machine
A reshape changes the shape but never the scale. Following one input through the pipeline above:
StageShapeScale
Pipeline input(1, 28, 28)S
After HomReshape(784,)S (unchanged)
After first HomLinear(50,)S
After HomSquare(50,)S²

See also​

  • HomUnsqueeze · HomSqueeze: insert or remove a size-1 axis on the server
  • HomSlice: extract an index or range along one axis
  • Repeat: fill the slots on the client, also client-only
  • Concepts: the n axis and how layout maps to slots

Next​

Back to

HomOps Reference

Browse all operators, or pick the next one from the sidebar.

→