from __future__ import annotations
from typing import Any, Callable
import jax
import jax.numpy as np
from ._regression_backing_funcs import reg_pmm_predict_func_legacy
from .basemodule import BaseModule
[docs]
class LegacyAffineObservablePMM(BaseModule):
"""
AffineObservablePMM is a module that implements the affine observable
Parametric Matrix Model (PMM), which is useful for generalize regression
"""
[docs]
def __init__(
self,
matrix_size: int = None,
num_eig: int = None,
output_size: int = None,
smoothing: float = None,
Ms: np.ndarray = None,
Ds: np.ndarray = None,
gs: np.ndarray = None,
init_magnitude: float = 1e-2,
) -> None:
"""
Initialize the LegacyAffineObservablePMM module. Represents a PMM which
evaluates the legacy affine observable PMM given by the process:
M(x) = M0 + x1 * M1 + ... + xp * Mp
-> (E, V) = eig(M(x)) [sorted in decreasing eigenvalue magnitude]
z_k = g_k
+ sum_{ij}^r (
[ |v_i^H D_{kij} v_j|^2 - 0.5 * ||D_{kij}||^2_2 ]
)
= g_k
- 0.5 * sum_ij^r ||D_{kij}||^2_2
+ sum_{ij}^r |v_i^H D_{kij} v_j|^2
By default this module is initialized to compute the smallest algebraic
eigenvalue (ground state).
Parameters
----------
matrix_size : int
Size of the PMM matrices (square). Shorthand `n`.
num_eig : int
Number of eigenvectors to use, the number of secondary matrices
will be r^2. Shorthand `r`.
output_size : int
Number of output features. Shorthand `k`.
smoothing : float, optional
Smoothing parameter, set to None/0.0 to disable.
Defaults to None/0.0.
Ms : np.ndarray, optional
Optional array of matrices M0, M1, ..., Mp that define the
parametric Hamiltonian. Each M must be Hermitian. If not
provided, the matrices will be randomly initialized when the
module is compiled.
Ds : np.ndarray, optional
Optional 5D array of matrices D_{kl} that define the
observables. Each D must be Hermitian. If not provided, the
observables will be randomly initialized when the module is
compiled.
gs : np.ndarray, optional
Optional vector of length r that defines the real biases for
the observable. If not provided, the biases will be
randomly initialized when the module is compiled.
init_magnitude : float, optional
Initial magnitude for the random matrices if Ms is not
provided, by default 1e-2.
"""
# input validation
if matrix_size is None or num_eig is None or output_size is None:
# module will be configured later, hopefully
self.matrix_size = None
self.num_eig = None
self.output_size = None
self.smoothing = None
self.input_size = None
self.Ms = None
self.Ds = None
self.gs = None
self.init_magnitude = None
return
if matrix_size <= 0 or not isinstance(matrix_size, int):
raise ValueError("matrix_size must be a positive integer")
if num_eig <= 0 or not isinstance(num_eig, int):
raise ValueError("num_eig must be a positive integer")
if num_eig > matrix_size:
raise ValueError(
"num_eig must be less than or equal to matrix_size, got"
f" num_eig={num_eig} and matrix_size={matrix_size}"
)
if output_size <= 0 or not isinstance(output_size, int):
raise ValueError("output_size must be a positive integer")
# if any of the parameters are provided, they must all be provided
if Ms is None and (Ds is not None or gs is not None):
raise ValueError(
"If Ds or gs are provided, Ms must also be provided"
)
if Ds is None and (gs is not None or Ms is not None):
raise ValueError(
"If gs or Ms are provided, Ds must also be provided"
)
if gs is None and (Ms is not None or Ds is not None):
raise ValueError(
"If Ms or Ds are provided, gs must also be provided"
)
if Ms is not None and Ds is not None and gs is not None:
if not isinstance(Ms, np.ndarray):
raise ValueError("Ms must be a numpy array")
if not isinstance(Ds, np.ndarray):
raise ValueError("Ds must be a numpy array")
if not isinstance(gs, np.ndarray):
raise ValueError("gs must be a numpy array")
if Ms.shape != (Ms.shape[0], matrix_size, matrix_size):
raise ValueError(
"Ms must be a 3D array of shape (input_size+1,"
f" matrix_size, matrix_size) [({Ms.shape[0]},"
f" {matrix_size}, {matrix_size})], got {Ms.shape}"
)
if Ds.shape != (
output_size,
num_eig,
num_eig,
matrix_size,
matrix_size,
):
raise ValueError(
"Ds must be a 5D array of shape (output_size, num_eig,"
f" num_eig, matrix_size, matrix_size) [({output_size},"
f" {num_eig}, {num_eig}, {matrix_size}, {matrix_size})],"
f" got {Ds.shape}"
)
if gs.shape != (output_size,):
raise ValueError(
f"gs must be a 1D array of length {output_size}, got"
f" {gs.shape}"
)
# ensure Ms are Hermitian
if not np.allclose(Ms, Ms.conj().transpose((0, 2, 1))):
raise ValueError("Ms must be Hermitian matrices")
# ensure Ds are Hermitian
if not np.allclose(Ds, Ds.conj().transpose((0, 1, 2, 4, 3))):
raise ValueError("Ds must be Hermitian matrices")
# ensure Ds is symmetric in the 1, 2 dimensions
if not np.allclose(Ds, Ds.transpose((0, 2, 1, 3, 4))):
raise ValueError("Ds must be symmetric in the 1, 2 dimensions")
# ensure gs is real
if not np.isreal(gs).all():
raise ValueError("gs must be a real-valued array")
self.matrix_size = matrix_size
self.num_eig = num_eig
self.output_size = output_size
self.smoothing = smoothing if smoothing is not None else 0.0
self.input_size = (
(Ms.shape[0] - 1) if Ms is not None else None
) # number of input features
self.Ms = Ms # matrices M0, M1, ..., Mp
self.Ds = Ds
self.gs = gs
self.init_magnitude = init_magnitude
[docs]
def name(self) -> str:
return (
"LegacyAffineObservablePMM"
f" ({self.matrix_size}x{self.matrix_size}, num_eig={self.num_eig},"
f" output_size={self.output_size}, smoothing={self.smoothing})"
)
[docs]
def is_ready(self) -> bool:
return (
self.input_size is not None
and self.Ms is not None
and self.Ds is not None
and self.gs is not None
)
[docs]
def get_num_trainable_floats(self) -> int | None:
if not self.is_ready():
return None
# each matrix M is Hermitian, and so contains n * (n - 1) / 2 distinct
# complex numbers and n distinct real numbers on the diagonal
# the total number of trainable floats is then just n^2 per matrix
# so Ms contributes (p + 1) * n^2 floats
# each matrix D is also Hermitian, so it contributes
# k * l * (l + 1) / 2* n^2 floats
# gs contributes k floats
return (
(self.input_size + 1) * self.matrix_size**2
+ self.output_size
* (self.num_eig * (self.num_eig + 1) // 2)
* self.matrix_size**2
+ self.k
)
[docs]
def _get_callable(self) -> Callable[
[
tuple[np.ndarray, ...],
np.ndarray,
bool,
tuple[np.ndarray, ...],
Any,
],
tuple[np.ndarray, tuple[np.ndarray, ...]],
]:
return lambda params, input_NF, training, state, rng: (
reg_pmm_predict_func_legacy(
params[0][0], # A or M0
np.array(params[0][1:]), # Bs or M1, ..., Mp
np.array(params[1]), # Ds
params[2], # gs
input_NF, # X
self.smoothing, # smoothing
),
state, # state is not used in this module, return it unchanged
)
[docs]
def compile(self, rng: Any, input_shape: tuple[int, ...]) -> None:
# input shape must be 1D
if len(input_shape) != 1:
raise ValueError(
f"Input shape must be 1D, got {len(input_shape)}D shape: "
f"{input_shape}"
)
# if the module is already ready, just verify the input shape
if self.is_ready():
if self.input_size != input_shape[0]:
raise ValueError(
f"Input shape {input_shape} does not match the expected "
f"number of features {self.input_size}"
)
return
# otherwise, initialize the matrices
self.input_size = input_shape[0] # number of input features
rng_Mreal, rng_Mimag, rng_Dreal, rng_Dimag, rng_g = jax.random.split(
rng, 5
)
self.Ms = self.init_magnitude * (
jax.random.normal(
rng_Mreal,
(self.input_size + 1, self.matrix_size, self.matrix_size),
dtype=np.complex64,
)
+ 1j
* jax.random.normal(
rng_Mimag,
(self.input_size + 1, self.matrix_size, self.matrix_size),
dtype=np.complex64,
)
)
# ensure the matrices are Hermitian
self.Ms = (self.Ms + self.Ms.conj().transpose((0, 2, 1))) / 2.0
# initialize Ds
self.Ds = self.init_magnitude * (
jax.random.normal(
rng_Dreal,
(
self.output_size,
self.num_eig,
self.num_eig,
self.matrix_size,
self.matrix_size,
),
dtype=np.complex64,
)
+ 1j
* jax.random.normal(
rng_Dimag,
(
self.output_size,
self.num_eig,
self.num_eig,
self.matrix_size,
self.matrix_size,
),
dtype=np.complex64,
)
)
# ensure the Ds are Hermitian
self.Ds = (self.Ds + self.Ds.conj().transpose((0, 1, 2, 4, 3))) / 2.0
# ensure Ds is symmetric in the 1, 2 dimensions
self.Ds = (self.Ds + self.Ds.transpose((0, 2, 1, 3, 4))) / 2.0
# initialize gs
self.gs = self.init_magnitude * jax.random.normal(
rng_g, (self.output_size,), dtype=np.float32
)
[docs]
def get_output_shape(
self, input_shape: tuple[int, ...]
) -> tuple[int, ...]:
return (self.output_size,)
[docs]
def get_hyperparameters(self) -> dict[str, Any]:
return {
"matrix_size": self.matrix_size,
"num_eig": self.num_eig,
"output_size": self.output_size,
"input_size": self.input_size,
"smoothing": self.smoothing,
"init_magnitude": self.init_magnitude,
}
[docs]
def set_hyperparameters(self, hyperparams: dict[str, Any]) -> None:
if self.Ms is not None or self.Ds is not None or self.gs is not None:
raise ValueError(
"Cannot set hyperparameters after the module has parameters"
)
super(LegacyAffineObservablePMM, self).set_hyperparameters(hyperparams)
[docs]
def get_params(self) -> tuple[np.ndarray, ...]:
return (self.Ms, self.Ds, self.gs)
[docs]
def set_params(self, params: tuple[np.ndarray, ...]) -> None:
if not isinstance(params, tuple) or not all(
isinstance(p, np.ndarray) for p in params
):
raise ValueError("params must be a tuple of numpy arrays")
if len(params) != 3:
raise ValueError(f"Expected 3 parameter arrays, got {len(params)}")
Ms, Ds, gs = params
if Ms.shape != (
self.input_size + 1,
self.matrix_size,
self.matrix_size,
):
raise ValueError(
"Ms must be a 3D array of shape (input_size+1, matrix_size,"
f" matrix_size) [({self.input_size + 1}, {self.matrix_size},"
f" {self.matrix_size})], got {Ms.shape}"
)
if Ds.shape != (
self.output_size,
self.num_eig,
self.num_eig,
self.matrix_size,
self.matrix_size,
):
raise ValueError(
"Ds must be a 4D array of shape (output_size, num_eig,"
f" num_eig, matrix_size, matrix_size) [({self.output_size},"
f" {self.num_eig}, {self.num_eig}, {self.matrix_size},"
f" {self.matrix_size})], got {Ds.shape}"
)
if gs.shape != (self.output_size,):
raise ValueError(
f"gs must be a 1D array of length {self.output_size}, got"
f" {gs.shape}"
)
# ensure Ms are Hermitian
if not np.allclose(Ms, Ms.conj().transpose((0, 2, 1))):
raise ValueError("Ms must be Hermitian matrices")
# ensure Ds are Hermitian
if not np.allclose(Ds, Ds.conj().transpose((0, 1, 2, 4, 3))):
raise ValueError("Ds must be Hermitian matrices")
# ensure Ds is symmetric in the 1, 2 dimensions
if not np.allclose(Ds, Ds.transpose((0, 2, 1, 3, 4))):
raise ValueError("Ds must be symmetric in the 1, 2 dimensions")
# ensure gs is real
if not np.isreal(gs).all():
raise ValueError("gs must be a real-valued array")
self.Ms = Ms
self.Ds = Ds
self.gs = gs
