# Copyright 2021 CR-Suite Development Team
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import jax.numpy as jnp
from jax import vmap, jit, lax
from .defs import RecoverySolution, HTPState
from cr.nimble.dsp import (hard_threshold,
hard_threshold_sorted,
build_signal_from_indices_and_values)
import cr.sparse.dict as crdict
import cr.sparse.lop as lop
[docs]def matrix_solve(Phi, y, K, normalized=False, step_size=None, max_iters=None, res_norm_rtol=1e-4):
"""Solves the sparse recovery problem :math:`y = \\Phi x + e` using Hard Thresholding Pursuit for matrices
"""
## Initialize some constants for the algorithm
M, N = Phi.shape
# squared norm of the signal
y_norm_sqr = y.T @ y
max_r_norm_sqr = y_norm_sqr * (res_norm_rtol ** 2)
if not normalized and step_size is None:
step_size = 0.98 / crdict.upper_frame_bound(Phi)
if max_iters is None:
max_iters = M
min_iters = min(3*K, 20)
def compute_step_size(h, I):
h_I = h[I]
Phi_I = Phi[:, I]
# Step size calculation
Ah = Phi_I @ h_I
mu = h_I.T @ h_I / (Ah.T @ Ah)
return mu
def get_step_size(h, I):
return compute_step_size(h, I) if normalized else step_size
def init():
# Data for the previous approximation [r = y, x = 0]
I_prev = jnp.arange(0, K)
x_I_prev = jnp.zeros(K)
r_norm_sqr_prev = y_norm_sqr
# Assume previous estimate to be zero and conduct first iteration
# compute the correlations of atoms with signal y
h = Phi.T @ y
mu = get_step_size(h, I_prev)
# update
x = mu * h
# threshold
I, x_I = hard_threshold(x, K)
# Form the subdictionary of corresponding atoms
Phi_I = Phi[:, I]
# Compute new residual
r = y - Phi_I @ x_I
# Compute residual norm squared
r_norm_sqr = r.T @ r
return HTPState(x_I=x_I, I=I, r=r, r_norm_sqr=r_norm_sqr,
iterations=1,
I_prev=I_prev, x_I_prev=x_I_prev, r_norm_sqr_prev=r_norm_sqr_prev)
def iteration(state):
I_prev = state.I
x_I_prev = state.x_I
r_norm_sqr_prev = state.r_norm_sqr
# compute the correlations of dictionary atoms with the residual
h = Phi.T @ state.r
# current approximation
x = build_signal_from_indices_and_values(N, state.I, state.x_I)
# Step size calculation
mu = get_step_size(h, I_prev)
# update
x = x + mu * h
# threshold
I, x_I = hard_threshold_sorted(x, K)
# Form the subdictionary of corresponding atoms
Phi_I = Phi[:, I]
# Solve least squares over the selected K indices
x_I, _, _, _ = jnp.linalg.lstsq(Phi_I, y)
# Compute new residual
y_hat = Phi_I @ x_I
r = y - y_hat
# Compute residual norm squared
r_norm_sqr = r.T @ r
return HTPState(x_I=x_I, I=I, r=r, r_norm_sqr=r_norm_sqr,
iterations=state.iterations+1,
I_prev=I_prev, x_I_prev=x_I_prev, r_norm_sqr_prev=r_norm_sqr_prev
)
def cond(state):
# limit on residual norm
a = state.r_norm_sqr > max_r_norm_sqr
# limit on number of iterations
b = state.iterations < max_iters
c = jnp.logical_and(a, b)
# checking if support is still changing
d = jnp.any(jnp.not_equal(state.I, state.I_prev))
# consider support change only after some iterations
d = jnp.logical_or(state.iterations < min_iters, d)
c = jnp.logical_and(c,d)
# overall condition
return c
state = lax.while_loop(cond, iteration, init())
return RecoverySolution(x_I=state.x_I, I=state.I, r=state.r, r_norm_sqr=state.r_norm_sqr,
iterations=state.iterations, length=Phi.shape[1])
matrix_solve_jit = jit(matrix_solve, static_argnums=(2),
static_argnames=("normalized", "step_size", "max_iters", "res_norm_rtol"))
[docs]def operator_solve(Phi, y, K, normalized=False, step_size=None, max_iters=None, res_norm_rtol=1e-4):
"""Solves the sparse recovery problem :math:`y = \\Phi x + e` using Hard Thresholding Pursuit for linear operators
"""
## Initialize some constants for the algorithm
M, N = Phi.shape
trans = Phi.trans
# squared norm of the signal
y_norm_sqr = y.T @ y
max_r_norm_sqr = y_norm_sqr * (res_norm_rtol ** 2)
if not normalized and step_size is None:
step_size = 0.98 / lop.upper_frame_bound(Phi)
if max_iters is None:
max_iters = M
min_iters = min(3*K, 20)
def compute_step_size(h, I):
h_I = h[I]
Phi_I = Phi.columns(I)
# Step size calculation
Ah = Phi_I @ h_I
mu = h_I.T @ h_I / (Ah.T @ Ah)
return mu
def get_step_size(h, I):
return compute_step_size(h, I) if normalized else step_size
def init():
# Data for the previous approximation [r = y, x = 0]
I_prev = jnp.arange(0, K)
x_I_prev = jnp.zeros(K)
r_norm_sqr_prev = y_norm_sqr
# Assume previous estimate to be zero and conduct first iteration
# compute the correlations of atoms with signal y
h = trans(y)
mu = get_step_size(h, I_prev)
# update
x = mu * h
# threshold
I, x_I = hard_threshold(x, K)
# Form the subdictionary of corresponding atoms
Phi_I = Phi.columns(I)
# Compute new residual
r = y - Phi_I @ x_I
# Compute residual norm squared
r_norm_sqr = r.T @ r
return HTPState(x_I=x_I, I=I, r=r, r_norm_sqr=r_norm_sqr,
iterations=1,
I_prev=I_prev, x_I_prev=x_I_prev, r_norm_sqr_prev=r_norm_sqr_prev)
def iteration(state):
I_prev = state.I
x_I_prev = state.x_I
r_norm_sqr_prev = state.r_norm_sqr
# compute the correlations of dictionary atoms with the residual
h = trans(state.r)
# current approximation
x = build_signal_from_indices_and_values(N, state.I, state.x_I)
# Step size calculation
mu = get_step_size(h, I_prev)
# update
x = x + mu * h
# threshold
I, x_I = hard_threshold_sorted(x, K)
# Form the subdictionary of corresponding atoms
Phi_I = Phi.columns(I)
# Solve least squares over the selected K indices
x_I, r_I_norms, rank_I, s_I = jnp.linalg.lstsq(Phi_I, y)
# Compute new residual
y_hat = Phi_I @ x_I
r = y - y_hat
# Compute residual norm squared
r_norm_sqr = r.T @ r
return HTPState(x_I=x_I, I=I, r=r, r_norm_sqr=r_norm_sqr,
iterations=state.iterations+1,
I_prev=I_prev, x_I_prev=x_I_prev, r_norm_sqr_prev=r_norm_sqr_prev
)
def cond(state):
# limit on residual norm
a = state.r_norm_sqr > max_r_norm_sqr
# limit on number of iterations
b = state.iterations < max_iters
c = jnp.logical_and(a, b)
# checking if support is still changing
d = jnp.any(jnp.not_equal(state.I, state.I_prev))
# consider support change only after some iterations
d = jnp.logical_or(state.iterations < min_iters, d)
c = jnp.logical_and(c,d)
# overall condition
return c
state = lax.while_loop(cond, iteration, init())
return RecoverySolution(x_I=state.x_I, I=state.I, r=state.r, r_norm_sqr=state.r_norm_sqr,
iterations=state.iterations, length=Phi.shape[1])
operator_solve_jit = jit(operator_solve, static_argnums=(0, 2),
static_argnames=("normalized", "step_size", "max_iters", "res_norm_rtol"))
solve = operator_solve_jit
