845 lines
34 KiB
Python
845 lines
34 KiB
Python
|
import bisect
|
||
|
|
from cmath import isnan
|
||
|
|
from email.mime import base
|
||
|
|
import matplotlib.pyplot as plt
|
||
|
|
import hashlib
|
||
|
|
import math
|
||
|
|
import numpy as np
|
||
|
|
import random
|
||
|
|
import statistics
|
||
|
|
|
||
|
|
from pkg_resources import get_distribution
|
||
|
|
from scipy import optimize, stats
|
||
|
|
from astropy import modeling
|
||
|
|
|
||
|
|
def encode(v):
|
||
|
|
byte_values = []
|
||
|
|
for i in range(0, math.ceil(len(v) / 8)):
|
||
|
|
x = 0
|
||
|
|
for j in range(0, 8):
|
||
|
|
index = i * 8 + j
|
||
|
|
if index >= len(v):
|
||
|
|
continue
|
||
|
|
x <<= 1
|
||
|
|
x |= int(v[index])
|
||
|
|
byte_values.append(x)
|
||
|
|
return bytearray(byte_values)
|
||
|
|
|
||
|
|
def sha(v):
|
||
|
|
x = encode(v)
|
||
|
|
m = hashlib.sha256()
|
||
|
|
m.update(x)
|
||
|
|
result = m.digest()
|
||
|
|
return result[0] & 0b1
|
||
|
|
|
||
|
|
def xor(v):
|
||
|
|
return np.sum(v[1:]) % 2
|
||
|
|
|
||
|
|
def hamming_distance(a, b, scratch):
|
||
|
|
np.logical_xor(a, b, scratch)
|
||
|
|
return sum(scratch)
|
||
|
|
|
||
|
|
def index_hash(indices):
|
||
|
|
return ','.join([str(index) for index in sorted(indices)])
|
||
|
|
|
||
|
|
def bin_div(a, b):
|
||
|
|
if a == 0 and b == 0:
|
||
|
|
return 2
|
||
|
|
if a == 1 and b == 0:
|
||
|
|
return -1
|
||
|
|
if a == 0 and b == 1:
|
||
|
|
return 0
|
||
|
|
return 1
|
||
|
|
|
||
|
|
class Candidate():
|
||
|
|
def __init__(self, indices):
|
||
|
|
self.indices = indices[:]
|
||
|
|
self.uplift = 0
|
||
|
|
|
||
|
|
def evaluate(self, x):
|
||
|
|
if len(x) in self.indices:
|
||
|
|
return 0
|
||
|
|
value = 1
|
||
|
|
for index in self.indices:
|
||
|
|
value *= x[index]
|
||
|
|
return value
|
||
|
|
|
||
|
|
def id(self):
|
||
|
|
return index_hash(self.indices)
|
||
|
|
|
||
|
|
def eval_str(self):
|
||
|
|
parts = []
|
||
|
|
for index in self.indices:
|
||
|
|
parts.append('x[' + str(index) + ']')
|
||
|
|
return '*'.join(parts)
|
||
|
|
|
||
|
|
class Probabilities():
|
||
|
|
def __init__(self):
|
||
|
|
self.N = 16
|
||
|
|
self.actual_N = self.N * 2
|
||
|
|
self.num_terms = 1
|
||
|
|
self.num_candidates = 100
|
||
|
|
# self.sample_size = self.N ** 2
|
||
|
|
self.sample_size = 64
|
||
|
|
self.p = np.zeros((self.actual_N + 1,))
|
||
|
|
self.p_temp = np.empty_like(self.p)
|
||
|
|
self.next_p = np.empty_like(self.p)
|
||
|
|
self.knowns = []
|
||
|
|
self.stops = set()
|
||
|
|
self.reset_p()
|
||
|
|
self.epoch = 0
|
||
|
|
|
||
|
|
self.inputs = np.zeros((self.sample_size, self.actual_N)).astype(np.int32)
|
||
|
|
self.raw_inputs = np.zeros((self.sample_size, self.N)).astype(np.int32)
|
||
|
|
self.masked_distances = np.zeros((self.sample_size, self.sample_size))
|
||
|
|
self.distances = np.zeros((self.sample_size, self.sample_size))
|
||
|
|
self.xor_square = np.zeros((self.sample_size, self.sample_size))
|
||
|
|
self.nn = np.zeros((self.sample_size, self.sample_size)).astype(np.int32)
|
||
|
|
self.nn_distances = np.zeros((self.sample_size, 2)).astype(np.int32)
|
||
|
|
self.base_outputs = np.zeros((self.sample_size)).astype(np.int32)
|
||
|
|
self.outputs = np.zeros((self.sample_size)).astype(np.int32)
|
||
|
|
self.expected_outputs = np.zeros((self.sample_size)).astype(np.int32)
|
||
|
|
self.output_xor = np.zeros((self.sample_size)).astype(np.int32)
|
||
|
|
self.mask = np.zeros((self.sample_size))
|
||
|
|
self.numerators = np.zeros((self.sample_size))
|
||
|
|
self.denominators = np.zeros((self.sample_size))
|
||
|
|
self.coherences = np.zeros((self.sample_size))
|
||
|
|
self.max_coherences = np.zeros((self.actual_N + 1))
|
||
|
|
self.max_candidates = [None for _ in range(0, self.actual_N)]
|
||
|
|
self.uplifts = np.zeros((self.actual_N))
|
||
|
|
self.uplift_means = np.zeros((self.actual_N))
|
||
|
|
self.uplift_medians = np.zeros((self.actual_N))
|
||
|
|
self.uplift_convergences = np.zeros((self.actual_N))
|
||
|
|
# self.subspace_uplift_samples = [[] for _ in range(0, self.actual_N)]
|
||
|
|
self.superspace_uplift_samples = []
|
||
|
|
self.subspace_uplifts = np.zeros((self.actual_N))
|
||
|
|
self.uplift_ranges = [[0, 0] for _ in range(0, self.actual_N)]
|
||
|
|
self.uplift_stddevs = np.zeros((self.actual_N))
|
||
|
|
|
||
|
|
self.last_index = -1
|
||
|
|
self.last_pvalue = -1
|
||
|
|
self.left_half = True
|
||
|
|
|
||
|
|
self.samples = 10
|
||
|
|
self.num_bins = 1000
|
||
|
|
# self.samples = 200
|
||
|
|
self.base_coherence_samples = np.zeros((self.samples))
|
||
|
|
self.coherence_samples = np.zeros((self.actual_N, self.samples))
|
||
|
|
self.subspace_uplift_samples = np.zeros((self.actual_N, self.samples))
|
||
|
|
self.subspace_uplift_weights = np.zeros((self.actual_N, self.samples))
|
||
|
|
|
||
|
|
self.layers = []
|
||
|
|
self.layer_confidence = {}
|
||
|
|
self.base = None
|
||
|
|
|
||
|
|
self.scratch = np.zeros((self.N,))
|
||
|
|
|
||
|
|
self.last_value = -1
|
||
|
|
self.rounds = 0
|
||
|
|
self.average_delta_over_null = 0
|
||
|
|
self.visited = set()
|
||
|
|
|
||
|
|
self.candidate_pool = []
|
||
|
|
self.candidate_ids = set()
|
||
|
|
self.has_added_layer = False
|
||
|
|
|
||
|
|
def randomize_inputs(self):
|
||
|
|
for i in range(0, self.sample_size):
|
||
|
|
for j in range(0, self.N):
|
||
|
|
val = random.randint(0, 1)
|
||
|
|
self.raw_inputs[i][j] = val
|
||
|
|
self.inputs[i][j * 2] = val
|
||
|
|
self.inputs[i][j * 2 + 1] = val ^ 1
|
||
|
|
|
||
|
|
def populate_distances(self):
|
||
|
|
self.nn.fill(-1)
|
||
|
|
self.nn_distances.fill(-1)
|
||
|
|
for i in range(0, len(self.raw_inputs)):
|
||
|
|
x_a = self.raw_inputs[i]
|
||
|
|
for j in range(0, len(self.raw_inputs)):
|
||
|
|
if i == j:
|
||
|
|
continue
|
||
|
|
x_b = self.raw_inputs[j]
|
||
|
|
distance = hamming_distance(x_a, x_b, self.scratch)
|
||
|
|
if (self.nn_distances[i][0] < 0 or distance < self.nn_distances[i][0]) and distance > 0:
|
||
|
|
self.nn_distances[i][0] = distance
|
||
|
|
self.nn_distances[i][1] = 1
|
||
|
|
self.nn[i][0] = j
|
||
|
|
elif distance == self.nn_distances[i][0]:
|
||
|
|
count = self.nn_distances[i][1]
|
||
|
|
self.nn_distances[i][1] = count + 1
|
||
|
|
self.nn[i][count] = j
|
||
|
|
# self.distances[i][j] = 1.0 / (2 ** (distance - 1)) if distance > 0 else 0
|
||
|
|
self.distances[i][j] = 1.0 / (distance ** 12) if distance > 0 else 0
|
||
|
|
|
||
|
|
def compute_expected_outputs(self):
|
||
|
|
for i in range(0, len(self.raw_inputs)):
|
||
|
|
self.expected_outputs[i] = xor(self.raw_inputs[i])
|
||
|
|
|
||
|
|
def compute_base_outputs(self):
|
||
|
|
if self.base is None:
|
||
|
|
self.base_outputs.fill(0)
|
||
|
|
return
|
||
|
|
for i in range(0, len(self.inputs)):
|
||
|
|
self.base_outputs[i] = self.base(self.inputs[i])
|
||
|
|
|
||
|
|
def mat_coherence(self):
|
||
|
|
np.abs(self.output_xor, self.mask)
|
||
|
|
np.subtract(self.output_xor, self.mask, self.mask)
|
||
|
|
np.divide(self.mask, 2.0, self.mask)
|
||
|
|
np.add(1.0, self.mask, self.mask)
|
||
|
|
self.xor_square.fill(0)
|
||
|
|
np.copyto(self.masked_distances, self.distances)
|
||
|
|
masked_distances_t = self.masked_distances.transpose()
|
||
|
|
for i in range(0, len(self.xor_square)):
|
||
|
|
self.xor_square[i] = self.output_xor
|
||
|
|
np.multiply(self.masked_distances[i], self.mask, self.masked_distances[i])
|
||
|
|
np.multiply(masked_distances_t[i], self.mask, masked_distances_t[i])
|
||
|
|
np.sum(self.masked_distances, axis=0, out=self.denominators)
|
||
|
|
self.xor_square = self.xor_square.transpose()
|
||
|
|
np.logical_xor(self.xor_square, self.output_xor, self.xor_square)
|
||
|
|
np.multiply(self.xor_square, self.masked_distances, self.xor_square)
|
||
|
|
np.sum(self.xor_square, axis=0, out=self.numerators)
|
||
|
|
np.divide(self.numerators, self.denominators, self.coherences)
|
||
|
|
mean = np.nanmean(self.coherences)
|
||
|
|
if isnan(mean):
|
||
|
|
mean = 1.0
|
||
|
|
return 1.0 - mean
|
||
|
|
|
||
|
|
def nn_coherence(self):
|
||
|
|
for i in range(0, len(self.output_xor)):
|
||
|
|
total = 0
|
||
|
|
y_a = self.output_xor[i]
|
||
|
|
[distance, count] = self.nn_distances[i]
|
||
|
|
for index in range(0, count):
|
||
|
|
j = self.nn[i][index]
|
||
|
|
y_b = self.output_xor[j]
|
||
|
|
total += 1 if y_a == 1 and y_b == 1 or y_a == 0 and y_b == 0 else 0
|
||
|
|
self.coherences[i] = total / count
|
||
|
|
return np.mean(self.coherences)
|
||
|
|
|
||
|
|
def coherence(self, outputs=None):
|
||
|
|
if outputs is None:
|
||
|
|
outputs = self.outputs
|
||
|
|
np.logical_xor(outputs, self.expected_outputs, self.output_xor)
|
||
|
|
return self.nn_coherence()
|
||
|
|
# return self.mat_coherence()
|
||
|
|
coherences = []
|
||
|
|
for i in range(0, len(self.output_xor)):
|
||
|
|
y_a = self.output_xor[i]
|
||
|
|
numerator = 0
|
||
|
|
denominator = 0
|
||
|
|
for j in range(0, len(self.output_xor)):
|
||
|
|
if i == j:
|
||
|
|
continue
|
||
|
|
y_b = self.output_xor[j]
|
||
|
|
weight = self.distances[i][j]
|
||
|
|
denominator += weight
|
||
|
|
if y_a == 0 and y_b == 0 or y_a == 1 and y_b == 1:
|
||
|
|
numerator += weight
|
||
|
|
coherence = numerator / denominator if denominator > 0 else 0
|
||
|
|
coherences.append(coherence)
|
||
|
|
|
||
|
|
raw_coherence = sum(coherences) / len(coherences)
|
||
|
|
check_coherence = self.mat_coherence()
|
||
|
|
|
||
|
|
return raw_coherence
|
||
|
|
|
||
|
|
def div_coherence(self):
|
||
|
|
coherences = []
|
||
|
|
for i in range(0, len(self.output_xor)):
|
||
|
|
y_a = self.output_xor[i]
|
||
|
|
if y_a < 0:
|
||
|
|
continue
|
||
|
|
numerator = 0
|
||
|
|
denominator = 0
|
||
|
|
for j in range(0, len(self.output_xor)):
|
||
|
|
if i == j:
|
||
|
|
continue
|
||
|
|
y_b = self.output_xor[j]
|
||
|
|
if y_b < 0:
|
||
|
|
continue
|
||
|
|
weight = self.distances[i][j]
|
||
|
|
denominator += weight
|
||
|
|
if y_a == 0 and y_b == 0 or y_a == 1 and y_b == 1:
|
||
|
|
numerator += weight
|
||
|
|
# if y_a < 0 or y_b < 0:
|
||
|
|
# numerator += weight
|
||
|
|
coherence = numerator / denominator if denominator > 0 else 0
|
||
|
|
coherences.append(coherence)
|
||
|
|
if len(coherences) == 0:
|
||
|
|
return 1.0
|
||
|
|
return sum(coherences) / len(coherences)
|
||
|
|
|
||
|
|
def normalize_p(self):
|
||
|
|
check = self.knowns[:]
|
||
|
|
for i in range(0, len(self.p)):
|
||
|
|
if self.p[i] < 0:
|
||
|
|
self.p[i] = 0
|
||
|
|
for i in range(0, len(self.p)):
|
||
|
|
if i in self.knowns:
|
||
|
|
flip = i ^ 0b1
|
||
|
|
self.p[i] = 0.0
|
||
|
|
self.p[flip] = 0.0
|
||
|
|
else:
|
||
|
|
check.append(i)
|
||
|
|
stop_id = index_hash(check)
|
||
|
|
check.pop()
|
||
|
|
if stop_id in self.stops:
|
||
|
|
self.p[i] = 0.0
|
||
|
|
total = np.sum(self.p)
|
||
|
|
if total > 0:
|
||
|
|
for i in range(0, len(self.p)):
|
||
|
|
self.p[i] = self.p[i] / total
|
||
|
|
|
||
|
|
def reset_p(self):
|
||
|
|
self.p.fill(1.0)
|
||
|
|
self.normalize_p()
|
||
|
|
|
||
|
|
def threshold(self):
|
||
|
|
# return (1.0 / (self.num_terms - len(self.knowns))) - (self.epoch / 100)
|
||
|
|
return 1.0 - (self.epoch / 1000)
|
||
|
|
|
||
|
|
def get_converged_index(self):
|
||
|
|
for i in range(0, len(self.p)):
|
||
|
|
if self.p[i] > self.threshold():
|
||
|
|
return i
|
||
|
|
return None
|
||
|
|
|
||
|
|
def add_layer(self):
|
||
|
|
self.has_added_layer = True
|
||
|
|
self.add_stop()
|
||
|
|
layer = Candidate(self.knowns)
|
||
|
|
self.layers.append(layer)
|
||
|
|
self.base = self.cache_layers()
|
||
|
|
self.knowns.pop()
|
||
|
|
self.reset_p()
|
||
|
|
|
||
|
|
def random_sample(self):
|
||
|
|
self.randomize_inputs()
|
||
|
|
self.populate_distances()
|
||
|
|
self.compute_expected_outputs()
|
||
|
|
self.compute_base_outputs()
|
||
|
|
return self.coherence(self.base_outputs)
|
||
|
|
|
||
|
|
def random_candidate(self):
|
||
|
|
indices = self.knowns[:]
|
||
|
|
np.copyto(self.p_temp, self.p)
|
||
|
|
self.p_temp[self.actual_N] = 0
|
||
|
|
total = np.sum(self.p_temp)
|
||
|
|
if total == 0:
|
||
|
|
return None
|
||
|
|
np.divide(self.p_temp, total, self.p_temp)
|
||
|
|
for _ in range(0, self.num_terms - len(self.knowns)):
|
||
|
|
index = np.random.choice(len(self.p_temp), 1, p=self.p_temp)[0]
|
||
|
|
indices.append(index)
|
||
|
|
flip = index ^ 0b1
|
||
|
|
self.p_temp[index] = 0
|
||
|
|
self.p_temp[flip] = 0
|
||
|
|
for i in range(0, len(self.p_temp)):
|
||
|
|
if i not in indices:
|
||
|
|
indices.append(i)
|
||
|
|
stop_id = index_hash(indices)
|
||
|
|
indices.pop()
|
||
|
|
if stop_id in self.stops:
|
||
|
|
self.p_temp[i] = 0.0
|
||
|
|
total = np.sum(self.p_temp)
|
||
|
|
if total == 0:
|
||
|
|
return None
|
||
|
|
np.divide(self.p_temp, total, self.p_temp)
|
||
|
|
return Candidate(indices)
|
||
|
|
|
||
|
|
def seed_candidate_pool(self):
|
||
|
|
for _ in range(0, self.num_candidates):
|
||
|
|
candidate = self.random_candidate()
|
||
|
|
if candidate is None:
|
||
|
|
continue
|
||
|
|
candidate_id = candidate.id()
|
||
|
|
if candidate_id in self.candidate_ids:
|
||
|
|
continue
|
||
|
|
self.candidate_pool.append(candidate)
|
||
|
|
self.candidate_ids.add(candidate_id)
|
||
|
|
|
||
|
|
def add_stop(self):
|
||
|
|
stop_id = index_hash(self.knowns)
|
||
|
|
self.stops.add(stop_id)
|
||
|
|
|
||
|
|
def get_distribution(self, candidate, half = 1):
|
||
|
|
count = 0
|
||
|
|
for i in range(0, len(self.inputs)):
|
||
|
|
value = candidate.evaluate(self.inputs[i])
|
||
|
|
if value == half:
|
||
|
|
self.output_xor[i] = self.base_outputs[i] ^ self.expected_outputs[i]
|
||
|
|
count += 1
|
||
|
|
else:
|
||
|
|
self.output_xor[i] = -1
|
||
|
|
# return (count, self.mat_coherence())
|
||
|
|
return (count, self.nn_coherence())
|
||
|
|
|
||
|
|
def err(self, fitted_model, bins, hist):
|
||
|
|
err = 0
|
||
|
|
for i in range(0, self.num_bins):
|
||
|
|
x = bins[i + 1]
|
||
|
|
y = hist[i]
|
||
|
|
delta = fitted_model(x) - y
|
||
|
|
err += delta * delta
|
||
|
|
return err / self.num_bins
|
||
|
|
|
||
|
|
def update(self):
|
||
|
|
sample = self.epoch
|
||
|
|
self.epoch += 1
|
||
|
|
|
||
|
|
base_coherence = self.random_sample()
|
||
|
|
self.base_coherence_samples[sample] = base_coherence
|
||
|
|
candidate = Candidate(self.knowns[:])
|
||
|
|
|
||
|
|
for i in range(0, self.actual_N):
|
||
|
|
candidate.indices.append(i)
|
||
|
|
try:
|
||
|
|
count_0, subspace_coherence_0 = self.get_distribution(candidate, 0)
|
||
|
|
# count_1, subspace_coherence_1 = self.get_distribution(candidate, 1)
|
||
|
|
# delta = (subspace_coherence_0 - base_coherence) * count_0 / self.sample_size
|
||
|
|
# delta = subspace_coherence_0 - subspace_coherence_1
|
||
|
|
self.subspace_uplift_samples[i][sample] = subspace_coherence_0 - base_coherence
|
||
|
|
self.subspace_uplift_weights[i][sample] = count_0 / self.sample_size
|
||
|
|
# self.subspace_uplift_left_samples[i][sample] = subspace_coherence_0
|
||
|
|
# self.subspace_uplift_right_samples[i][sample] = subspace_coherence_1 - base_coherence
|
||
|
|
|
||
|
|
# if index_hash(candidate.indices) in self.stops:
|
||
|
|
# continue
|
||
|
|
|
||
|
|
for j in range(0, len(self.inputs)):
|
||
|
|
self.outputs[j] = self.base_outputs[j] ^ candidate.evaluate(self.inputs[j])
|
||
|
|
|
||
|
|
coherence = self.coherence()
|
||
|
|
self.coherence_samples[i][sample] = coherence - base_coherence
|
||
|
|
# self.coherence_samples[i][sample] = coherence
|
||
|
|
finally:
|
||
|
|
candidate.indices.pop()
|
||
|
|
|
||
|
|
if self.epoch >= self.samples:
|
||
|
|
# for i in range(0, self.actual_N):
|
||
|
|
# parameters = stats.norm.fit(self.uplift_samples[i])
|
||
|
|
# print(i, parameters)
|
||
|
|
# print(i, stats.kstest(self.uplift_samples[i], "norm", parameters))
|
||
|
|
|
||
|
|
added = False
|
||
|
|
# parameters = stats.norm.fit(self.base_coherence_samples)
|
||
|
|
# (base_mu, _) = parameters
|
||
|
|
|
||
|
|
# (hist, bins) = np.histogram(self.base_coherence_samples, self.num_bins, density=True)
|
||
|
|
# fitter = modeling.fitting.LevMarLSQFitter()
|
||
|
|
# model = modeling.models.Gaussian1D()
|
||
|
|
# fitted_model = fitter(model, bins[1:], hist)
|
||
|
|
# print('Base', fitted_model.mean.value, self.err(fitted_model, bins, hist))
|
||
|
|
|
||
|
|
# x = np.linspace(0, 1.0, 10000)
|
||
|
|
# density = stats.gaussian_kde(self.base_coherence_samples)(x)
|
||
|
|
# mode = x[np.argsort(density)[-1]]
|
||
|
|
# print(mode)
|
||
|
|
|
||
|
|
# for i in range(0, self.actual_N):
|
||
|
|
# count = 0
|
||
|
|
# for j in range(0, self.samples):
|
||
|
|
# for k in range(0, self.samples):
|
||
|
|
# if self.coherence_samples[i][j] > self.base_coherence_samples[k]:
|
||
|
|
# count += 1
|
||
|
|
# print(i, count)
|
||
|
|
|
||
|
|
try:
|
||
|
|
index = -1
|
||
|
|
lowest_index = -1
|
||
|
|
lowest_pvalue = -1
|
||
|
|
highest_index = -1
|
||
|
|
highest_pvalue = -1
|
||
|
|
best_pvalue = -1
|
||
|
|
pvalue_sum = 0
|
||
|
|
pvalue_denom = 0
|
||
|
|
is_subspace = False
|
||
|
|
|
||
|
|
for i in range(0, self.actual_N):
|
||
|
|
if i in self.knowns:
|
||
|
|
continue
|
||
|
|
try:
|
||
|
|
result = stats.ttest_1samp(self.coherence_samples[i], 0, alternative='greater')
|
||
|
|
print(i, result)
|
||
|
|
# (hist, bins) = np.histogram(self.coherence_samples[i], 20, range=(-0.01, 0.01))
|
||
|
|
# total = 0
|
||
|
|
# for j in range(0, 20):
|
||
|
|
# total += hist[j] * (bins[j] + bins[j + 1]) / 2
|
||
|
|
# mode = total / sum(hist)
|
||
|
|
|
||
|
|
# fitter = modeling.fitting.LevMarLSQFitter()
|
||
|
|
# model = modeling.models.Gaussian1D()
|
||
|
|
# fitted_model = fitter(model, bins[1:], hist)
|
||
|
|
# mode = fitted_model.mean.value
|
||
|
|
# print(i, total)
|
||
|
|
|
||
|
|
# result = stats.kstest(self.base_coherence_samples, self.coherence_samples[i], alternative='greater')
|
||
|
|
# print(i, result)
|
||
|
|
# value = result.pvalue * (1 - result.statistic)
|
||
|
|
# parameters = stats.norm.fit(self.coherence_samples[i])
|
||
|
|
# (mu, _) = parameters
|
||
|
|
# density = stats.gaussian_kde(self.coherence_samples[i])(x)
|
||
|
|
# mode = x[np.argsort(density)[-1]]
|
||
|
|
# print(i, mode)
|
||
|
|
# print(i, mu)
|
||
|
|
if not isnan(result.pvalue):
|
||
|
|
if i == self.last_index:
|
||
|
|
delta = abs(result.pvalue - self.last_pvalue)
|
||
|
|
if delta < 0.1:
|
||
|
|
print('Low delta!')
|
||
|
|
print(self.last_index, delta)
|
||
|
|
# self.last_index = -1
|
||
|
|
self.left_half = not self.left_half
|
||
|
|
# self.layers.pop()
|
||
|
|
# self.base = self.cache_layers()
|
||
|
|
# return
|
||
|
|
|
||
|
|
pvalue_sum += result.pvalue
|
||
|
|
pvalue_denom += 1
|
||
|
|
if lowest_index < 0 or result.pvalue < lowest_pvalue:
|
||
|
|
lowest_index = i
|
||
|
|
lowest_pvalue = result.pvalue
|
||
|
|
if highest_index < 0 or result.pvalue > highest_pvalue:
|
||
|
|
highest_index = i
|
||
|
|
highest_pvalue = result.pvalue
|
||
|
|
except Exception as e:
|
||
|
|
print(e)
|
||
|
|
pass
|
||
|
|
average_pvalue = pvalue_sum / pvalue_denom
|
||
|
|
print(average_pvalue)
|
||
|
|
index = highest_index if self.left_half else lowest_index
|
||
|
|
best_pvalue = highest_pvalue if self.left_half else lowest_pvalue
|
||
|
|
|
||
|
|
self.last_index = index
|
||
|
|
self.last_pvalue = best_pvalue
|
||
|
|
# if average_pvalue < 0.5:
|
||
|
|
# index = lowest_index
|
||
|
|
# best_pvalue = lowest_pvalue
|
||
|
|
# else:
|
||
|
|
# index = highest_index
|
||
|
|
# best_pvalue = highest_pvalue
|
||
|
|
# print(e)
|
||
|
|
|
||
|
|
# for i in range(0, self.actual_N):
|
||
|
|
# if i in self.knowns:
|
||
|
|
# continue
|
||
|
|
# # result = stats.kstest(self.base_coherence_samples, self.subspace_uplift_left_samples[i], alternative='greater')
|
||
|
|
# # # result = stats.kstest(self.subspace_uplift_left_samples[i], self.subspace_uplift_right_samples[i], alternative='greater')
|
||
|
|
# # print(i, result)
|
||
|
|
# # value = result.pvalue * (1 - result.statistic)
|
||
|
|
# # parameters = stats.norm.fit(self.subspace_uplift_left_samples[i])
|
||
|
|
# # (mu, _) = parameters
|
||
|
|
# try:
|
||
|
|
# result = stats.ttest_1samp(self.subspace_uplift_samples[i], 0, alternative='greater')
|
||
|
|
# print(i, result)
|
||
|
|
# # (hist, bins) = np.histogram(self.subspace_uplift_samples[i], 20, range=(-0.01, 0.01))
|
||
|
|
# # bin_index = np.argsort(hist)[-1]
|
||
|
|
# # mode = (bins[bin_index] + bins[bin_index + 1]) / 2
|
||
|
|
# # fitter = modeling.fitting.LevMarLSQFitter()
|
||
|
|
# # model = modeling.models.Gaussian1D()
|
||
|
|
# # fitted_model = fitter(model, bins[1:], hist)
|
||
|
|
# # mode = fitted_model.mean.value
|
||
|
|
# # print(i, mode)
|
||
|
|
# # density = stats.gaussian_kde(self.subspace_uplift_samples[i], weights=self.subspace_uplift_weights[i])(x)
|
||
|
|
# # density = stats.gaussian_kde(self.subspace_uplift_samples[i])(x)
|
||
|
|
# # mode = x[np.argsort(density)[-1]]
|
||
|
|
# # print(i, mode)
|
||
|
|
# # print(i, mu)
|
||
|
|
# if (index < 0 or result.pvalue < lowest_pvalue) and not isnan(result.pvalue):
|
||
|
|
# # if index < 0 or value < lowest_pvalue:
|
||
|
|
# index = i
|
||
|
|
# lowest_pvalue = result.pvalue
|
||
|
|
# is_subspace = True
|
||
|
|
|
||
|
|
# # if result.pvalue > 0.95:
|
||
|
|
# # index = i
|
||
|
|
# # parameters = stats.norm.fit(self.subspace_uplift_samples[i])
|
||
|
|
# # (mu, _) = parameters
|
||
|
|
# # if mu > base_mu:
|
||
|
|
# # if index < 0 or mu > highest_mu:
|
||
|
|
# # index = i
|
||
|
|
# # highest_mu = mu
|
||
|
|
# except Exception as e:
|
||
|
|
# print(e)
|
||
|
|
# pass
|
||
|
|
# # print(e)
|
||
|
|
|
||
|
|
if index >= 0:
|
||
|
|
if is_subspace:
|
||
|
|
# print('subspace')
|
||
|
|
self.knowns.append(index)
|
||
|
|
print(self.knowns, best_pvalue)
|
||
|
|
else:
|
||
|
|
# print('flat')
|
||
|
|
self.knowns.append(index)
|
||
|
|
# self.layer_confidence[index_hash(self.knowns)] = confidence
|
||
|
|
# num_terms = len(self.knowns)
|
||
|
|
print(self.knowns, best_pvalue)
|
||
|
|
print(base_coherence)
|
||
|
|
self.add_layer()
|
||
|
|
# if num_terms > self.num_terms:
|
||
|
|
# self.stops = set()
|
||
|
|
# self.num_terms = num_terms
|
||
|
|
self.knowns = []
|
||
|
|
return
|
||
|
|
else:
|
||
|
|
self.knowns = []
|
||
|
|
# else:
|
||
|
|
# self.knowns = []
|
||
|
|
|
||
|
|
# if len(self.knowns) > 0:
|
||
|
|
# # self.add_stop()
|
||
|
|
# self.knowns = []
|
||
|
|
finally:
|
||
|
|
# fig, axs = plt.subplots(int(self.actual_N / 4), 4)
|
||
|
|
# x_eval = np.linspace(-1.0, 1.0, num=1000)
|
||
|
|
# for i in range(0, int(self.actual_N / 4)):
|
||
|
|
# for j in range(0, 4):
|
||
|
|
# # (hist, bins) = np.histogram(self.base_coherence_samples, self.num_bins, density=True)
|
||
|
|
# # fitter = modeling.fitting.LevMarLSQFitter()
|
||
|
|
# # model = modeling.models.Gaussian1D()
|
||
|
|
# # fitted_model = fitter(model, bins[1:], hist)
|
||
|
|
# # axs[i][j].scatter(bins[1:], hist, s=1, color='r', alpha=0.5)
|
||
|
|
# # axs[i][j].plot(x_eval, fitted_model(x_eval), color='r')
|
||
|
|
|
||
|
|
# (hist, bins) = np.histogram(self.coherence_samples[i * 4 + j], self.num_bins, density=True)
|
||
|
|
# # fitter = modeling.fitting.LevMarLSQFitter()
|
||
|
|
# # model = modeling.models.Gaussian1D()
|
||
|
|
# # fitted_model = fitter(model, bins[1:], hist)
|
||
|
|
# axs[i][j].scatter(bins[1:], hist, s=1, color='g', alpha=0.5)
|
||
|
|
# # axs[i][j].plot(x_eval, fitted_model(x_eval), color='g')
|
||
|
|
|
||
|
|
# (hist, bins) = np.histogram(self.subspace_uplift_samples[i * 4 + j], self.num_bins, density=True)
|
||
|
|
# # fitter = modeling.fitting.LevMarLSQFitter()
|
||
|
|
# # model = modeling.models.Gaussian1D()
|
||
|
|
# # fitted_model = fitter(model, bins[1:], hist)
|
||
|
|
# axs[i][j].scatter(bins[1:], hist, s=1, color='b', alpha=0.5)
|
||
|
|
# # axs[i][j].plot(x_eval, fitted_model(x_eval), color='b')
|
||
|
|
|
||
|
|
# # kde0 = stats.gaussian_kde(self.base_coherence_samples)
|
||
|
|
# kde1 = stats.gaussian_kde(self.coherence_samples[i * 4 + j])
|
||
|
|
# # kde2 = stats.gaussian_kde(self.subspace_uplift_samples[i * 4 + j], weights=self.subspace_uplift_weights[i])
|
||
|
|
# kde2 = stats.gaussian_kde(self.subspace_uplift_samples[i * 4 + j])
|
||
|
|
# # axs[i][j].plot(x_eval, kde0(x_eval), color='r')
|
||
|
|
# axs[i][j].plot(x_eval, kde1(x_eval), color='g')
|
||
|
|
# axs[i][j].plot(x_eval, kde2(x_eval), color='b')
|
||
|
|
# # n, bins, patches = axs[i][j].hist(self.base_coherence_samples, 50, density=True, facecolor='r', alpha=0.5)
|
||
|
|
# # n, bins, patches = axs[i][j].hist(self.coherence_samples[i * 4 + j], 50, density=True, facecolor='g', alpha=0.5)
|
||
|
|
# # n, bins, patches = axs[i][j].hist(self.subspace_uplift_samples[i * 4 + j], 50, density=True, facecolor='b', alpha=0.5)
|
||
|
|
# plt.show()
|
||
|
|
self.epoch = 0
|
||
|
|
|
||
|
|
return
|
||
|
|
|
||
|
|
# print('=====' + str(base_coherence))
|
||
|
|
# print(self.uplifts)
|
||
|
|
# print(self.uplift_means)
|
||
|
|
# print(self.uplift_medians)
|
||
|
|
# print(self.uplift_stddevs)
|
||
|
|
# print(self.uplift_ranges)
|
||
|
|
# print(self.uplift_convergences)
|
||
|
|
# print(self.subspace_uplifts)
|
||
|
|
|
||
|
|
if index >= 0:
|
||
|
|
self.knowns.append(index)
|
||
|
|
print(base_coherence)
|
||
|
|
print(self.knowns, self.epoch)
|
||
|
|
# print(self.uplift_medians)
|
||
|
|
# print(self.uplifts)
|
||
|
|
# print(self.subspace_uplifts)
|
||
|
|
self.add_layer()
|
||
|
|
self.uplifts.fill(0)
|
||
|
|
self.subspace_uplifts.fill(0)
|
||
|
|
self.uplift_medians.fill(0)
|
||
|
|
self.uplift_convergences.fill(0)
|
||
|
|
self.uplift_samples = [[] for _ in range(0, self.actual_N)]
|
||
|
|
self.epoch = 0
|
||
|
|
return
|
||
|
|
|
||
|
|
if subspace_index >= 0:
|
||
|
|
self.knowns.append(subspace_index)
|
||
|
|
print(self.knowns, self.epoch)
|
||
|
|
# print(self.uplifts)
|
||
|
|
# print(self.subspace_uplifts)
|
||
|
|
self.uplifts.fill(0)
|
||
|
|
self.subspace_uplifts.fill(0)
|
||
|
|
self.uplift_medians.fill(0)
|
||
|
|
self.uplift_convergences.fill(0)
|
||
|
|
self.uplift_samples = [[] for _ in range(0, self.actual_N)]
|
||
|
|
self.epoch = 0
|
||
|
|
return
|
||
|
|
|
||
|
|
# print('======')
|
||
|
|
# print(self.epoch, base_coherence)
|
||
|
|
# print('======')
|
||
|
|
|
||
|
|
# if len(self.candidate_pool) == 0:
|
||
|
|
# print(self.p)
|
||
|
|
|
||
|
|
# for i in range(0, min(5, len(self.candidate_pool))):
|
||
|
|
# candidate = self.candidate_pool[i]
|
||
|
|
# print(candidate.id(), candidate.uplift)
|
||
|
|
|
||
|
|
# if self.epoch < 15:
|
||
|
|
# return
|
||
|
|
|
||
|
|
if self.candidate_pool[0].uplift > 0.3:
|
||
|
|
candidate = self.candidate_pool[0]
|
||
|
|
candidate_id = candidate.id()
|
||
|
|
self.candidate_ids.remove(candidate_id)
|
||
|
|
print(candidate_id)
|
||
|
|
self.knowns = candidate.indices
|
||
|
|
self.add_layer()
|
||
|
|
self.knowns = []
|
||
|
|
self.reset_p()
|
||
|
|
self.epoch = 0
|
||
|
|
self.candidate_pool = []
|
||
|
|
self.candidate_ids = set()
|
||
|
|
elif self.candidate_pool[0].uplift < -0.3 or self.epoch > 200:
|
||
|
|
self.epoch = 0
|
||
|
|
self.num_terms += 1
|
||
|
|
self.candidate_pool = []
|
||
|
|
self.candidate_ids = set()
|
||
|
|
self.knowns = []
|
||
|
|
self.stops = set()
|
||
|
|
self.reset_p()
|
||
|
|
return
|
||
|
|
|
||
|
|
# np.copyto(self.next_p, self.p)
|
||
|
|
for _ in range(0, self.num_candidates):
|
||
|
|
candidate = self.random_candidate()
|
||
|
|
if candidate is None:
|
||
|
|
continue
|
||
|
|
candidate_id = candidate.id()
|
||
|
|
if candidate_id in visited:
|
||
|
|
continue
|
||
|
|
visited.add(candidate_id)
|
||
|
|
if self.actual_N in candidate.indices:
|
||
|
|
continue
|
||
|
|
has_candidate = True
|
||
|
|
for i in range(0, len(self.inputs)):
|
||
|
|
self.outputs[i] = self.base_outputs[i] ^ candidate.evaluate(self.inputs[i])
|
||
|
|
# coherence = self.ring_coherence()
|
||
|
|
coherence = self.coherence()
|
||
|
|
# if coherence <= base_coherence:
|
||
|
|
# continue
|
||
|
|
# for index in candidate.indices:
|
||
|
|
# self.next_p[index] += (coherence - base_coherence) * (1 / 1000.0)
|
||
|
|
# self.p_temp[index] += 0
|
||
|
|
for index in candidate.indices:
|
||
|
|
if coherence > self.max_coherences[index]:
|
||
|
|
self.max_coherences[index] = coherence
|
||
|
|
self.max_candidates[index] = candidate
|
||
|
|
# self.max_coherences[index] = max(self.max_coherences[index], coherence)
|
||
|
|
# np.copyto(self.p, self.next_p)
|
||
|
|
|
||
|
|
# np.copyto(self.p_temp, self.p)
|
||
|
|
for i in range(0, self.actual_N):
|
||
|
|
candidate = self.max_candidates[i]
|
||
|
|
if candidate is None:
|
||
|
|
continue
|
||
|
|
for index in candidate.indices:
|
||
|
|
self.p[index] += (self.max_coherences[index] - base_coherence) * (1 / 1000.0)
|
||
|
|
# print(i, self.max_coherences[i] - base_coherence, self.max_candidates[i].id())
|
||
|
|
self.normalize_p()
|
||
|
|
# print(self.p)
|
||
|
|
|
||
|
|
# np.subtract(self.p_temp, self.p, self.p_temp)
|
||
|
|
# np.abs(self.p_temp, self.p_temp)
|
||
|
|
# delta = np.sum(self.p_temp) / len(self.p_temp)
|
||
|
|
# print(delta, np.argmax(self.p))
|
||
|
|
# np.copyto(self.p_temp, self.p)
|
||
|
|
# for i in range(0, len(self.p_temp)):
|
||
|
|
# self.p_temp[i] = round(self.p_temp[i] * 100) / 100
|
||
|
|
# print(self.p_temp)
|
||
|
|
|
||
|
|
index = np.argmax(self.p)
|
||
|
|
delta_over_null = self.p[index] - self.p[self.actual_N]
|
||
|
|
if self.epoch == 0:
|
||
|
|
self.average_delta_over_null = delta_over_null
|
||
|
|
else:
|
||
|
|
self.average_delta_over_null = 0.9 * self.average_delta_over_null + 0.1 * delta_over_null
|
||
|
|
diff = self.num_terms - len(self.knowns)
|
||
|
|
|
||
|
|
print(self.average_delta_over_null, np.argpartition(self.p, -diff)[-diff:], np.argmax(self.p))
|
||
|
|
|
||
|
|
# Always iterate for a minimum number of epochs
|
||
|
|
if self.epoch < 15:
|
||
|
|
return
|
||
|
|
if self.average_delta_over_null > 0.00001 and self.average_delta_over_null < 0.001 and self.epoch < 300:
|
||
|
|
return
|
||
|
|
if self.average_delta_over_null < 0.001:
|
||
|
|
index = self.actual_N
|
||
|
|
else:
|
||
|
|
index = np.argmax(self.p)
|
||
|
|
|
||
|
|
# index = np.argmax(self.p)
|
||
|
|
# if index == self.last_value:
|
||
|
|
# self.rounds += 1
|
||
|
|
# else:
|
||
|
|
# self.rounds = 0
|
||
|
|
# self.last_value = index
|
||
|
|
|
||
|
|
# if self.rounds < 10 and self.epoch < 100:
|
||
|
|
# return
|
||
|
|
|
||
|
|
# if self.epoch < 5 or (delta > 0.001 and self.epoch < 50):
|
||
|
|
# return
|
||
|
|
|
||
|
|
# index = np.argmax(self.p)
|
||
|
|
|
||
|
|
# print(self.p)
|
||
|
|
# print(self.threshold())
|
||
|
|
# print(self.p)
|
||
|
|
# index = self.get_converged_index()
|
||
|
|
if not index is None or not has_candidate:
|
||
|
|
# print(index, delta, np.argmax(self.p))
|
||
|
|
self.epoch = 0
|
||
|
|
if index == self.actual_N or not has_candidate:
|
||
|
|
if len(self.knowns) > 0:
|
||
|
|
self.add_stop()
|
||
|
|
self.knowns.pop()
|
||
|
|
print('Backtrack: ' + str(self.knowns))
|
||
|
|
self.reset_p()
|
||
|
|
return
|
||
|
|
self.num_terms += 1
|
||
|
|
self.knowns = []
|
||
|
|
self.stops = set()
|
||
|
|
self.reset_p()
|
||
|
|
print(self.num_terms)
|
||
|
|
return
|
||
|
|
self.knowns.append(index)
|
||
|
|
# bisect.insort(self.knowns, index)
|
||
|
|
if len(self.knowns) == self.num_terms:
|
||
|
|
print('Add layer: ' + str(self.knowns))
|
||
|
|
self.add_layer()
|
||
|
|
else:
|
||
|
|
print('Found term: ' + str(self.knowns))
|
||
|
|
self.reset_p()
|
||
|
|
print(base_coherence)
|
||
|
|
return
|
||
|
|
|
||
|
|
def cache_layers(self):
|
||
|
|
expr = 'def f(x):\n\tresult=0\n'
|
||
|
|
for layer in self.layers:
|
||
|
|
expr += '\tresult^=' + layer.eval_str() + '\n'
|
||
|
|
expr += '\treturn result\n'
|
||
|
|
scope = {}
|
||
|
|
exec(expr, scope)
|
||
|
|
return scope['f']
|
||
|
|
|
||
|
|
def main():
|
||
|
|
probabilities = Probabilities()
|
||
|
|
# probabilities.knowns = [14]
|
||
|
|
# probabilities.add_layer()
|
||
|
|
# probabilities.knowns = [8]
|
||
|
|
# probabilities.add_layer()
|
||
|
|
# probabilities.knowns = [4]
|
||
|
|
# probabilities.add_layer()
|
||
|
|
while probabilities.num_terms <= probabilities.N:
|
||
|
|
probabilities.update()
|
||
|
|
|
||
|
|
if __name__ == "__main__":
|
||
|
|
main()
|