Required prerequisites
Describe the feature
Add built-in APIs for common quantum embeddings.
--
(1) Amplitude encoding:
Maps and normalizes classical features (list, ndarray or cudaq.State) to a quantum state by encoding values as amplitudes: $x \mapsto \sum_i x_i |i\rangle$, including pad as an argument for vectors not sized to $2^n$ with the specified pad value, e.g., zero-padding to the nearest power of $2^n$ by setting pad = 0. The featured API looks like:
state_vector = cudaq.amplitude_encode([0.5, 0.5, 0.5], pad = 0)
Returns cudaq.State with values [0.5773, 0.5773, 0.5773, 0] (L2-norm = 1)
--
(2) Angular encoding:
Encodes classical features (list or ndarray) as qubit rotations (e.g., with specified Ry gates): $x_i \mapsto R_y(x_i)|0\rangle$:
@cudaq.kernel
def kernel(angles: list[float]):
q = cudaq.qvector(3)
cudaq.angular_encode(q, angles, rotation='Y') # or cudaq.kernels.angular_encode
The corresponding kernel:
print(cudaq.draw(kernel, [0.1, 0.2, 0.3]))
╭─────────╮
q0 : ┤ ry(0.1) ├
├─────────┤
q1 : ┤ ry(0.2) ├
├─────────┤
q2 : ┤ ry(0.3) ├
╰─────────╯
--
Both embedding methods are commonly used in QML, enabling developers to design models more easily and intuitively.
Comment for Unitary Hack:
We are happy to accept PRs that implement either one encoding and preferably both.
Required prerequisites
Describe the feature
Add built-in APIs for common quantum embeddings.
--
(1) Amplitude encoding:$x \mapsto \sum_i x_i |i\rangle$ , including $2^n$ with the specified pad value, e.g., zero-padding to the nearest power of $2^n$ by setting
Maps and normalizes classical features (
list,ndarrayorcudaq.State) to a quantum state by encoding values as amplitudes:padas an argument for vectors not sized topad = 0. The featured API looks like:Returns
cudaq.Statewith values[0.5773, 0.5773, 0.5773, 0](L2-norm = 1)--
(2) Angular encoding:$x_i \mapsto R_y(x_i)|0\rangle$ :
Encodes classical features (
listorndarray) as qubit rotations (e.g., with specified Ry gates):The corresponding kernel:
--
Both embedding methods are commonly used in QML, enabling developers to design models more easily and intuitively.
Comment for Unitary Hack:
We are happy to accept PRs that implement either one encoding and preferably both.