Xoryn-II / source /verify.py
PureOne's picture
Release Xoryn II finite-grid image recovery benchmark
8f0b027 verified
Raw History Blame Contribute Delete
20.5 kB
#!/usr/bin/env python3
"""Reproduce Xoryn II population checks, image experiment, and figures."""
from argparse import ArgumentParser
from itertools import product
import json
from pathlib import Path
import numpy as np
from sympy import Matrix, ZZ
from sympy.matrices.normalforms import smith_normal_form
from grid import Grid, OneTap, recover_sparse, restore_two_taps, signed_graph_nullity
def relative_orbit_error(observed, recovered, phase):
distance = np.linalg.norm(observed[..., None, :] * phase - recovered[..., None, :], axis=-1)
return distance.min(axis=-1) / np.linalg.norm(observed, axis=-1)
def population_checks(rng):
shapes = [(2,2), (2,3), (2,8), (3,3), (3,5), (4,4), (4,11), (5,5), (6,6), (7,12),
(8,8), (9,9), (11,11), (2,2,2), (2,2,5), (2,3,4), (3,3,3), (3,4,5), (4,5,6)]
one_error, linear_error, sparse_error = [], [], []
number_of_equations = 0
for shape in shapes:
grid = Grid(shape)
plan = OneTap(grid)
phase = grid.phases(grid.points)
for _ in range(4):
z = np.fft.fftn(rng.normal(size=shape), norm="ortho").ravel()
views = z*phase
moments = plan.estimates(views.real, views.imag @ plan.weights)
linear_error.append(float(np.max(abs(moments-grid.product(z, plan.rows)), initial=0)))
magnitude = np.sqrt(np.maximum(0, (views.real**2).mean(axis=0)
* np.where(grid.real, 1, 2)))
recovered = plan.recover(magnitude, moments, views.real[:,0].mean())
one_error.append(float(relative_orbit_error(z, recovered, phase)))
for density in (.35, .65, 1):
z = np.fft.fftn(rng.normal(size=shape), norm="ortho").ravel()
mask = rng.random(grid.size) < density
z *= mask[grid.representative]
error, constraints = restore_two_taps(grid, z)
sparse_error.append(error)
number_of_equations += constraints
assert max(linear_error + one_error + sparse_error) < 1e-9
return {"grids": shapes, "one_tap_cases": len(one_error),
"max_reconstruction_relative_error": max(one_error),
"max_selected_moment_error": max(linear_error),
"two_tap_cases": len(sparse_error), "tested_imaginary_coordinates": number_of_equations,
"max_two_tap_error": max(sparse_error)}
def relation_order(frequencies, p):
frequencies = np.asarray(frequencies)
size = len(frequencies)
candidates = np.array(tuple(product(range(-p, p+1), repeat=size)))
degree = np.abs(candidates).sum(axis=1)
useful = (degree > 0) & (degree <= p) & np.all((candidates @ frequencies)%p == 0, axis=1)
independent = any((int(a[0])*int(b[1])-int(a[1])*int(b[0]))%p != 0
for a in frequencies for b in frequencies)
expected = p**(2 if independent else 1)
for d in range(2,p+1):
rows = candidates[useful & (degree <= d)]
if len(rows) < size:
continue
diagonal = smith_normal_form(Matrix(rows.tolist()), domain=ZZ).diagonal()
index = abs(int(np.prod(diagonal)))
if index == expected:
return d
raise AssertionError("relations through p must generate")
def coded_higher_order(rng):
p = 5
grid = Grid((p,p))
phase = grid.phases(grid.points)
k, l = grid.anchors
z = np.zeros(grid.size, dtype=complex)
z[k], z[l] = .8*np.exp(.38j), 1.2*np.exp(.51j)
z[grid.minus[[k,l]]] = z[[k,l]].conj()
views = phase*z
C, S = views[:, [k,l]].real, views[:, [k,l]].imag
codes = np.array(tuple(product((-1,1), repeat=2)))
Q = (S @ codes.T) / np.sqrt(2)
power = np.abs(z[k])**2
fourth_cosine = C[:,0]**4
# (C+iS)^5 = C^5 - 10 C^3 S^2 + 5 C S^4
# + i S (5 C^4 - 10 C^2 S^2 + S^4).
s2 = power - C[:,0]**2
real = np.mean(C[:,0]**5 - 10*C[:,0]**3*s2 + 5*C[:,0]*s2**2)
imag_poly = 5*fourth_cosine - 10*C[:,0]**2*s2 + s2**2
weighted = np.sqrt(2)*np.mean(imag_poly[:,None] * Q * codes[:,0])
error = abs((real + 1j*weighted) - z[k]**5)
assert error < 1e-12
degrees = {}
for prime in (5,7):
g = Grid((prime,prime))
counts = {}
example = relation_order(g.points[g.anchors], prime)
assert example == prime
for _ in range(30):
support = rng.choice(g.odd, size=rng.integers(2,5), replace=False)
value = relation_order(g.points[support], prime)
counts[str(value)] = counts.get(str(value),0) + 1
degrees[str(prime)] = {"random_supports_by_order": counts, "two_anchor_order": example}
return {"fifth_moment_error": float(error), "orders": degrees}
def evaluate_distance(grid, w, anchor, block=16):
# Distances to a character from Fourier correlation, and to sampled cubic rows.
spectrum = (np.fft.fftn(w.reshape(grid.shape))/grid.size).ravel()
d_orbit = np.sqrt(max(0,2-2*np.max(spectrum.real)))
total = 0
full_frequency = np.arange(grid.size)
for start in range(0,len(anchor),block):
a = anchor[start:start+block, None]
b = full_frequency[None, :]
ab = grid.add(a,b)
total += np.sum(abs(w[a]*w[b]*w[ab].conj()-1)**2)
d_anchor = np.sqrt(total/(len(anchor)*grid.size))
d_double = np.sqrt(np.mean(abs(w*w*w[grid.add(full_frequency,full_frequency)].conj()-1)**2))
return float(d_orbit), float(d_anchor), float(d_double)
def stability_checks(rng):
results = []
for shape in [(17,17),(32,32),(64,64),(128,128)]:
grid = Grid(shape)
log_size = int(np.ceil(8*np.log(2*grid.size)))
anchor = rng.integers(0,grid.size,size=log_size)
freq = np.bincount(anchor, minlength=grid.size).reshape(shape) / len(anchor)
bias = np.fft.fftn(freq).real.ravel()
delta = float(np.sqrt(2-2*np.max(bias[1:])))
assert delta >= 1
scale = 2*np.sqrt(2)*(3+np.sqrt(np.count_nonzero(grid.real)))/delta
ratios = []
for epsilon in (.005,.05,.25,1,3):
for _ in range(4):
a = rng.normal(size=grid.size)
w = np.exp(1j*epsilon*(a-a[grid.minus]))
d, A, B = evaluate_distance(grid,w,anchor)
ratios.append(d/(B+scale*A))
w = np.full(grid.size,-1.+0.j)
w[anchor] = w[grid.minus[anchor]] = w[0] = 1
d,A,B = evaluate_distance(grid,w,anchor)
ratios.append(d/(B+scale*A))
assert max(ratios) <= 1+1e-9
results.append({"shape": shape, "anchors": len(anchor), "anchor_separation": delta,
"max_proven_bound_ratio": max(ratios), "sign_flip_example": {
"translation_distance": d, "anchor_only_error": A, "doubling_error": B}})
return results
def exact_digit_support(grid, images):
"""A Fourier value at an eighth root is zero exactly when its cyclotomic remainder is zero."""
powers = (grid.points @ grid.points.T) % 8
residues = np.arange(4)
basis = (powers[:,:,None] == residues).astype(np.int32)
basis -= (powers[:,:,None] == residues+4).astype(np.int32)
coefficients = np.einsum("np,kpr->nkr",images.reshape(-1,64).astype(np.int32),basis)
support = (coefficients != 0).any(axis=-1)
support[:,0] = False
return support
def image_evidence(rng, dataset):
images = dataset["digits"]
grid = Grid((8,8))
plan = OneTap(grid)
phase = grid.phases(grid.points)
z = np.fft.fft2(images, axes=(1,2), norm="ortho").reshape(-1,grid.size)
z[:,0] = 0
z *= np.sqrt(grid.size)/np.linalg.norm(z,axis=1,keepdims=True)
support = exact_digit_support(grid,images)
assert np.array_equal(support,abs(z)>1e-10)
valid = np.flatnonzero(support[:,1:].all(axis=1))
reconstruction_errors = []
for start in range(0,len(valid),32):
coefficients = z[valid[start:start+32]]
view = coefficients[:,None,:]*phase
m = plan.estimates(view.real,view.imag@plan.weights)
magnitude = np.sqrt(np.maximum(0,(view.real**2).mean(axis=1)*np.where(grid.real,1,2)))
rec = plan.recover(magnitude,m)
reconstruction_errors.extend(relative_orbit_error(coefficients,rec,phase).tolist())
assert max(reconstruction_errors)<2e-8
sparse = np.flatnonzero(~support[:,1:].all(axis=1))
patterns, membership = np.unique(support[sparse],axis=0,return_inverse=True)
sparse_errors = []
for i,mask in enumerate(patterns):
coefficients = z[sparse[membership==i]].copy()
coefficients[:,~mask] = 0
rec = recover_sparse(grid,coefficients,mask,phase)
sparse_errors.extend(relative_orbit_error(coefficients,rec,phase).tolist())
assert max(sparse_errors)<2e-8
# A predeclared balanced test compares each known template with its point inversion.
nontrivial = relative_orbit_error(z[valid],z[valid].conj(),phase)>1e-8
test = rng.choice(valid[nontrivial],size=512,replace=False)
labels = np.tile([1,-1],len(test))
prototypes = np.repeat(z[test],2,axis=0)
actual = np.where(labels[:,None]>0,prototypes,prototypes.conj())
rows = plan.rows[plan.uses_tap]
truth = grid.product(prototypes, rows).imag
b,c = grid.representative[rows[:,1]],grid.representative[rows[:,2]]
a = rows[:,0]
tap_factor = plan.tap_factors[plan.uses_tap]
accuracy = []
for count in (32,128):
for sigma in (.5,1,2,4):
prediction, full = [],[]
for j in range(0,len(labels),32):
k = min(j+32,len(labels))
view = actual[j:k,None,:]*phase[rng.integers(grid.size,size=(k-j,count))]
re_noise = rng.normal(0,sigma,size=(k-j,count,grid.size))
im_noise = rng.normal(0,sigma,size=re_noise.shape)
cosine = (view.real+re_noise)[...,grid.representative]
positive_sine = (view.imag+im_noise)[...,grid.representative]
sine = positive_sine*grid.sign*np.where(grid.real,0,1)
tap = sine@plan.weights
mixed = (tap[...,None]*(cosine[...,b]*cosine[...,c]-sigma**2*(b==c))).mean(axis=1)
score = (mixed*tap_factor*truth[j:k]).sum(axis=1)
prediction.extend(np.where(score>=0,1,-1))
uncompressed = cosine+1j*sine
F = grid.product(uncompressed, rows).mean(axis=1).imag
full.extend(np.where((F*truth[j:k]).sum(axis=1)>=0,1,-1))
accuracy.append({"views_per_trial": count, "coordinate_noise_sd": sigma,
"one_tap_accuracy": float(np.mean(prediction==labels)),
"full_view_accuracy": float(np.mean(full==labels)),
"cosine_only_bayes_accuracy": .5})
exemplar = test[:3]
example = {"template_indices": exemplar.tolist(),
"digit_labels": dataset["digit_labels"][exemplar].tolist()}
results={"dataset_images": len(images), "full_support_images": len(valid),
"reconstruction_max_relative_error": max(reconstruction_errors),
"sparse_support_images": len(sparse), "sparse_support_patterns": len(patterns),
"sparse_max_relative_error": max(sparse_errors),
"reconstruction_median_relative_error": float(np.median(reconstruction_errors)),
"one_tap_readings_per_view": int((grid.size+np.count_nonzero(grid.real))//2+1),
"full_readings_per_view": grid.size, "templates_in_recognition_test": len(test),
"trials_per_noisy_setting": len(labels), "recognition": accuracy, "example": example}
return results, images[exemplar]
def rgb_evidence(dataset):
grid = Grid((32,32))
plan = OneTap(grid)
phase = grid.phases(grid.points)
weight = np.ones(3)/np.sqrt(3)
factor_real = np.where(grid.real,1,2)
factor_imag = np.where(grid.real,0,grid.sign*2*np.sqrt(len(grid.odd)))
corners = dataset["rgb_corners"].tolist()
cases = []
best_examples = {}
for crops, filename in zip(dataset["rgb_crops"], dataset["rgb_names"]):
for source, (y,x) in zip(crops, corners):
patch = source.astype(float)/255
z = np.fft.fft2(patch, axes=(0,1),norm="ortho").reshape(grid.size,3)
anchor = z@weight
if np.min(abs(anchor[1:])) <= 1e-10:
cases.append({"source": Path(filename).name, "top_left": [y,x], "full_anchor": False})
continue
hviews = phase*anchor
c = hviews.real
tap_axis = hviews.imag@plan.weights
tap_dense = hviews[:,grid.odd].imag.sum(axis=1)/np.sqrt(len(grid.odd))
moments = plan.estimates(c,tap_axis)
magnitude = np.sqrt((c*c).mean(axis=0)*factor_real)
h_recovered = plan.recover(magnitude,moments,c[:,0].mean())
cos = (phase[:,:,None]*z).real
cross_re = np.einsum("pk,pkc->kc",c,cos)/grid.size*factor_real[:,None]
cross_im = np.einsum("p,pkc->kc",tap_dense,cos)/grid.size*factor_imag[:,None]
recovered = np.empty_like(z)
recovered[1:] = (cross_re[1:]+1j*cross_im[1:]).conj()/h_recovered[1:,None].conj()
recovered[0] = cos[:,0].mean(axis=0)
align = np.argmin(np.linalg.norm(phase*anchor-h_recovered,axis=1))
aligned = recovered*phase[align,:,None].conj()
error = float(np.linalg.norm(aligned-z)/np.linalg.norm(z))
non_dc_error = float(np.linalg.norm(aligned[1:]-z[1:])/np.linalg.norm(z[1:]))
cases.append({"source": Path(filename).name, "top_left": [y,x],
"full_anchor": True, "relative_error": error,
"relative_non_dc_error": non_dc_error})
contrast = float(np.std(patch,axis=(0,1)).mean())
best = best_examples.get(filename)
if best is None or contrast > best[0]:
pixels = np.fft.ifft2(aligned.reshape(32,32,3),axes=(0,1),norm="ortho").real
best_examples[filename] = (contrast,(Path(filename).name,patch,pixels))
maximum = max(r["relative_error"] for r in cases if r["full_anchor"])
non_dc = max(r["relative_non_dc_error"] for r in cases if r["full_anchor"])
assert max(maximum, non_dc) < 1e-8
return {"shape": [32,32,3], "crops_tested": len(cases),
"crops_with_full_anchor": sum(r["full_anchor"] for r in cases),
"max_relative_error": maximum, "max_relative_non_dc_error": non_dc,
"full_readings_per_view": 3*grid.size,
"two_tap_readings_per_view": int(3*(grid.size+np.count_nonzero(grid.real))//2+2),
"cases": cases}, [entry[1] for entry in best_examples.values()]
def make_figure(path, images, results):
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
plt.rcParams.update({"font.family":"DejaVu Sans","font.size":10,
"axes.spines.top":False,"axes.spines.right":False})
f = plt.figure(figsize=(11,3.65),facecolor="white")
gs = f.add_gridspec(2,5, width_ratios=[1,1,1,.28,4], wspace=.3, hspace=.15)
inversion = (-np.arange(8))%8
for j in range(3):
for row in range(2):
ax=f.add_subplot(gs[row,j])
im=images[j] if row==0 else images[j][np.ix_(inversion,inversion)]
ax.imshow(im,cmap="cividis",vmin=0,vmax=16,interpolation="nearest")
ax.axis("off")
if j==0:
ax.text(-.16,.5,"Original" if row==0 else "Inverted",rotation=90,
ha="center",va="center",transform=ax.transAxes,color="#253a4a")
ax=f.add_subplot(gs[:,4])
data=[r for r in results["recognition"] if r["views_per_trial"]==128]
x=[r["coordinate_noise_sd"] for r in data]
ax.plot(x,[r["one_tap_accuracy"]*100 for r in data],lw=2.1,marker="o",
color="#0d756a",label="One fixed sine reading")
ax.plot(x,[r["full_view_accuracy"]*100 for r in data],lw=2.0,marker="s",
color="#46658a",label="Full-view cubic score")
ax.axhline(50,lw=1.4,color="#a34731",linestyle="--",label="Folded-data Bayes limit")
ax.set(xticks=x,ylim=(45,102), xlabel="Noise SD per real coordinate",ylabel="Pair accuracy (%)")
ax.legend(loc="lower left",fontsize=8)
ax.grid(alpha=.25)
ax.set_title("Point-inversion recognition · 128 views",fontweight="bold",color="#15304a")
f.subplots_adjust(left=.07,right=.98,top=.88,bottom=.17)
f.savefig(path,bbox_inches="tight")
plt.close(f)
def make_rgb_figure(path, examples):
import matplotlib.pyplot as plt
f, axes = plt.subplots(1,2*len(examples),figsize=(8.8,2.35),facecolor="white")
for i,(name,original,recovered) in enumerate(examples):
for j,image in enumerate((original,recovered)):
ax = axes[2*i+j]
ax.imshow(np.clip(image,0,1),interpolation="nearest")
ax.axis("off")
title = name.split(".")[0].capitalize()+(" reference" if j==0 else " recovered")
ax.set_title(title,color="#15304a",fontsize=10,fontweight="bold")
f.subplots_adjust(wspace=.15,left=.02,right=.98,bottom=.02,top=.85)
f.savefig(path,bbox_inches="tight")
plt.close(f)
def write_report_table(path, results):
p = results["population"]
d = results["digits"]
c = results["rgb"]
number = lambda x: ("%0.2e" % x).replace("e-", r"\times10^{-") + "}"
rows = [r"\begin{tabularx}{\linewidth}{@{}Xr@{}}",
r"\toprule Result & Measurement\\ \midrule",
f"Fully supported recorded digits, out of {d['dataset_images']:,} & {d['full_support_images']:,}"+r"\\",
r"Largest noiseless relative translation-orbit error on these digits & " +
"$"+number(d["reconstruction_max_relative_error"])+"$"+r"\\",
f"Digit readings per view, one-tap versus full & {d['one_tap_readings_per_view']} versus {d['full_readings_per_view']}"+r"\\",
f"Remaining digits with spectral holes, recovered with two taps & {d['sparse_support_images']}"+r"\\",
r"Largest noiseless relative error on those digits & $"+
number(d["sparse_max_relative_error"])+"$"+r"\\",
f"Natural-image RGB crops reconstructed, out of {c['crops_tested']} & {c['crops_with_full_anchor']}"+r"\\",
r"Largest noiseless relative error on RGB crops, excluding DC & $"+
number(c["max_relative_non_dc_error"])+"$"+r"\\",
f"RGB readings per view, two-tap versus full & {c['two_tap_readings_per_view']:,} versus {c['full_readings_per_view']:,}"+r"\\",
f"Two-tap transfer across {p['two_tap_cases']} synthetic cases & {p['tested_imaginary_coordinates']:,} cubic coordinate checks"+r"\\",
r"\bottomrule\end{tabularx}", "",
r"\medskip\noindent\textbf{Noisy point-inversion decisions.} Each setting uses 1,024 balanced trials across 512 known digit templates.",
"",
r"{\small\begin{tabularx}{\linewidth}{@{}rcc>{\centering\arraybackslash}Xr@{}}",
r"\toprule Views & Noise SD & One fixed tap & Full-view cubic & Folded Bayes\\ \midrule"]
for r in d["recognition"]:
rows.append(f"{r['views_per_trial']} & {r['coordinate_noise_sd']:g} & "
f"{100*r['one_tap_accuracy']:.1f}"+r"\% & "+
f"{100*r['full_view_accuracy']:.1f}"+r"\% & 50.0\%\\")
rows += [r"\bottomrule\end{tabularx}}", ""]
path.write_text("\n".join(rows))
def main():
parser = ArgumentParser()
parser.add_argument("--output",type=Path,default=Path("."))
args = parser.parse_args()
args.output.mkdir(parents=True,exist_ok=True)
rng = np.random.default_rng(471826)
report = {"seed":471826,"population":population_checks(rng),
"tensor_nullities":[signed_graph_nullity(p) for p in (3,5,7,11,13,17)],
"sparse_coded":coded_higher_order(rng),"conditioning":stability_checks(rng)}
with np.load(Path(__file__).with_name("benchmark_inputs.npz"), allow_pickle=False) as dataset:
report["digits"],images = image_evidence(rng,dataset)
report["rgb"],examples = rgb_evidence(dataset)
(args.output/"results.json").write_text(json.dumps(report,indent=2)+"\n")
make_figure(args.output/"recognition.pdf",images,report["digits"])
make_rgb_figure(args.output/"rgb.pdf",examples)
write_report_table(args.output/"results_table.tex",report)
print(json.dumps(report,indent=2))
if __name__=="__main__":
main()