File size: 17,009 Bytes
95db09d eb37c53 95db09d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 |
"""
GEOMETRIC BASIN CLASSIFIER - CIFAR-100 [PROPER STRUCTURE]
----------------------------------------------------------
Meant to replace the need for cross-entropy with cantor stairs and produce a more solid form of loss. The experiment was successful.
Requires additional testing with alternative systems and accessors.
Author: AbstractPhil + Claude Sonnet 4.5
License: MIT
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import torchvision
import torchvision.transforms as transforms
from tqdm import tqdm
import math
import numpy as np
import os
import json
from datetime import datetime
from pathlib import Path
import csv
# Hugging Face Hub integration
try:
from huggingface_hub import HfApi, create_repo
HF_AVAILABLE = True
except ImportError:
print("β οΈ huggingface_hub not installed. Run: pip install huggingface_hub")
HF_AVAILABLE = False
# Safetensors integration
try:
from safetensors.torch import save_file as save_safetensors
from safetensors.torch import load_file as load_safetensors
SAFETENSORS_AVAILABLE = True
except ImportError:
print("β οΈ safetensors not installed. Run: pip install safetensors")
SAFETENSORS_AVAILABLE = False
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# MIXING AUGMENTATIONS
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def alphamix_data(x, y, alpha_range=(0.3, 0.7), spatial_ratio=0.25):
"""AlphaMix: Spatially localized transparent overlay."""
batch_size = x.size(0)
index = torch.randperm(batch_size, device=x.device)
y_a, y_b = y, y[index]
alpha_min, alpha_max = alpha_range
beta_sample = np.random.beta(2, 2)
alpha = alpha_min + (alpha_max - alpha_min) * beta_sample
_, _, H, W = x.shape
overlay_ratio = np.sqrt(spatial_ratio)
overlay_h = int(H * overlay_ratio)
overlay_w = int(W * overlay_ratio)
top = np.random.randint(0, H - overlay_h + 1)
left = np.random.randint(0, W - overlay_w + 1)
composited_x = x.clone()
overlay_region = alpha * x[:, :, top:top+overlay_h, left:left+overlay_w]
background_region = (1 - alpha) * x[index, :, top:top+overlay_h, left:left+overlay_w]
composited_x[:, :, top:top+overlay_h, left:left+overlay_w] = overlay_region + background_region
return composited_x, y_a, y_b, alpha
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# DEVIL'S STAIRCASE PE
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class DevilStaircasePE(nn.Module):
"""Devil's Staircase PE - let alpha float naturally."""
def __init__(self, levels=20, features_per_level=4, smooth_tau=0.25, base=3):
super().__init__()
self.levels = levels
self.features_per_level = features_per_level
self.tau = smooth_tau
self.base = base
self.alpha = nn.Parameter(torch.tensor(0.1))
self.base_features = 2
if features_per_level > 2:
self.feature_expansion = nn.Linear(self.base_features, features_per_level)
else:
self.feature_expansion = None
def forward(self, positions, seq_len):
x = positions.float() / max(1, (seq_len - 1))
x = x.clamp(1e-6, 1.0 - 1e-6)
feats = []
Cx = torch.zeros_like(x)
for k in range(1, self.levels + 1):
scale = self.base ** k
y = (x * scale) % self.base
centers = torch.tensor([0.5, 1.5, 2.5], device=x.device, dtype=x.dtype)
d2 = (y.unsqueeze(-1) - centers) ** 2
logits = -d2 / (self.tau + 1e-8)
p = F.softmax(logits, dim=-1)
bit_k = p[..., 2] + self.alpha * p[..., 1]
Cx = Cx + bit_k * (0.5 ** k)
ent = -(p * p.clamp_min(1e-8).log()).sum(dim=-1)
pdf_proxy = 1.1 - ent / math.log(3.0)
base_feat = torch.stack([bit_k, pdf_proxy], dim=-1)
if self.feature_expansion is not None:
level_feat = self.feature_expansion(base_feat)
else:
level_feat = base_feat
feats.append(level_feat)
pe_levels = torch.stack(feats, dim=1)
return pe_levels, Cx
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# RESIDUAL BLOCK
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class ResidualBlock(nn.Module):
"""Basic residual block with skip connection."""
def __init__(self, in_channels, out_channels, stride=1):
super().__init__()
self.conv1 = nn.Conv2d(in_channels, out_channels, 3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_channels)
self.shortcut = nn.Sequential()
if stride != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1, stride=stride, bias=False),
nn.BatchNorm2d(out_channels)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = F.relu(out)
return out
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GEOMETRIC BASIN COMPATIBILITY
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GeometricBasinCompatibility(nn.Module):
"""Compute geometric compatibility scores - FULLY BATCHED."""
def __init__(self, num_classes=100, pe_levels=20, features_per_level=4):
super().__init__()
self.num_classes = num_classes
self.pe_levels = pe_levels
self.features_per_level = features_per_level
self.class_signatures = nn.Parameter(
torch.randn(num_classes, pe_levels, features_per_level) * 0.1
)
self.cantor_prototypes = nn.Parameter(
torch.linspace(0.0, 1.0, num_classes)
)
self.level_resonance = nn.Parameter(
torch.ones(num_classes, pe_levels) / pe_levels
)
def forward(self, pe_levels, cantor_measures):
B = pe_levels.shape[0]
# 1. TRIADIC COMPATIBILITY
pe_norm = F.normalize(pe_levels, p=2, dim=-1)
sig_norm = F.normalize(self.class_signatures, p=2, dim=-1)
similarities = torch.einsum('blf,clf->bcl', pe_norm, sig_norm)
similarities = (similarities + 1) / 2
resonance = F.softmax(self.level_resonance, dim=-1)
triadic_compat = (similarities * resonance.unsqueeze(0)).sum(dim=-1)
# 2. SELF-SIMILARITY
level_pairs = []
for k in range(self.pe_levels - 1):
level_k = pe_levels[:, k, :]
level_k1 = pe_levels[:, k+1, :]
sim = F.cosine_similarity(level_k, level_k1, dim=-1, eps=1e-8)
sim = (sim + 1) / 2
level_pairs.append(sim)
self_sim_pattern = torch.stack(level_pairs, dim=1)
expected_patterns = torch.sigmoid(
self.level_resonance[:, :-1] - self.level_resonance[:, 1:]
)
pattern_diff = torch.abs(
self_sim_pattern.unsqueeze(1) - expected_patterns.unsqueeze(0)
)
self_sim_compat = 1 - pattern_diff.mean(dim=-1)
self_sim_compat = torch.clamp(self_sim_compat, 0.0, 1.0)
# 3. CANTOR COHERENCE
distances = torch.abs(
cantor_measures.unsqueeze(1) - self.cantor_prototypes.unsqueeze(0)
)
cantor_compat = torch.exp(-distances ** 2 / 0.1) + 1e-8
# 4. HIERARCHICAL CHECK
split_point = self.pe_levels // 2
early_levels = pe_levels[:, :split_point, :].mean(dim=1)
late_levels = pe_levels[:, split_point:, :].mean(dim=1)
early_targets = self.class_signatures[:, :split_point, :].mean(dim=1)
late_targets = self.class_signatures[:, split_point:, :].mean(dim=1)
early_levels_norm = F.normalize(early_levels, p=2, dim=-1)
late_levels_norm = F.normalize(late_levels, p=2, dim=-1)
early_targets_norm = F.normalize(early_targets, p=2, dim=-1)
late_targets_norm = F.normalize(late_targets, p=2, dim=-1)
early_compat = torch.matmul(early_levels_norm, early_targets_norm.t())
late_compat = torch.matmul(late_levels_norm, late_targets_norm.t())
early_compat = (early_compat + 1) / 2
late_compat = (late_compat + 1) / 2
hier_compat = (early_compat + late_compat) / 2
# 5. COMBINE
eps = 1e-6
triadic_compat = torch.clamp(triadic_compat, eps, 1.0)
self_sim_compat = torch.clamp(self_sim_compat, eps, 1.0)
cantor_compat = torch.clamp(cantor_compat, eps, 1.0)
hier_compat = torch.clamp(hier_compat, eps, 1.0)
compatibility_scores = (
triadic_compat *
self_sim_compat *
cantor_compat *
hier_compat
) ** 0.25
return compatibility_scores
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GEOMETRIC BASIN LOSS
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GeometricBasinLoss(nn.Module):
"""Loss based on geometric basin compatibility."""
def __init__(self, temperature=0.1):
super().__init__()
self.temperature = temperature
def forward(self, compatibility_scores, labels, mixed_labels=None, lam=None):
batch_size = compatibility_scores.shape[0]
if mixed_labels is not None and lam is not None:
primary_compat = compatibility_scores[torch.arange(batch_size), labels]
secondary_compat = compatibility_scores[torch.arange(batch_size), mixed_labels]
primary_loss = F.mse_loss(primary_compat, torch.full_like(primary_compat, lam))
secondary_loss = F.mse_loss(secondary_compat, torch.full_like(secondary_compat, 1 - lam))
soft_targets = torch.zeros_like(compatibility_scores)
soft_targets[torch.arange(batch_size), labels] = lam
soft_targets[torch.arange(batch_size), mixed_labels] = 1 - lam
compat_normalized = compatibility_scores / (compatibility_scores.sum(dim=1, keepdim=True) + 1e-8)
kl_loss = F.kl_div(
compat_normalized.log(),
soft_targets,
reduction='batchmean'
)
total_loss = primary_loss + secondary_loss + 0.1 * kl_loss
else:
correct_compat = compatibility_scores[torch.arange(batch_size), labels]
correct_loss = -torch.log(correct_compat + 1e-8).mean()
mask = torch.ones_like(compatibility_scores)
mask[torch.arange(batch_size), labels] = 0
incorrect_compat = compatibility_scores * mask
incorrect_loss = torch.log(1 - incorrect_compat + 1e-8).mean()
incorrect_loss = -incorrect_loss
scaled_scores = compatibility_scores / self.temperature
log_probs = F.log_softmax(scaled_scores, dim=1)
contrastive_loss = F.nll_loss(log_probs, labels)
total_loss = correct_loss + 0.5 * incorrect_loss + 0.5 * contrastive_loss
return total_loss
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GEOMETRIC BASIN CLASSIFIER
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GeometricBasinClassifier(nn.Module):
"""BIGGER classifier with deeper ResNet-style backbone."""
def __init__(self, num_classes=100, pe_levels=20, pe_features_per_level=4, dropout=0.1):
super().__init__()
self.num_classes = num_classes
self.pe_levels = pe_levels
self.pe_features_per_level = pe_features_per_level
# Initial conv
self.conv1 = nn.Conv2d(3, 64, 3, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(64)
# Residual blocks
self.layer1 = self._make_layer(64, 128, num_blocks=2, stride=2)
self.layer2 = self._make_layer(128, 256, num_blocks=2, stride=2)
self.layer3 = self._make_layer(256, 512, num_blocks=2, stride=2)
self.layer4 = self._make_layer(512, 1024, num_blocks=2, stride=2)
self.global_pool = nn.AdaptiveAvgPool2d(1)
self.dropout = nn.Dropout(dropout)
# Devil's Staircase PE
self.pe = DevilStaircasePE(pe_levels, pe_features_per_level)
# PE modulator
self.pe_modulator = nn.Sequential(
nn.Linear(1024, 512),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(512, pe_levels * pe_features_per_level)
)
# Geometric basin
self.basin = GeometricBasinCompatibility(
num_classes,
pe_levels,
pe_features_per_level
)
def _make_layer(self, in_channels, out_channels, num_blocks, stride):
layers = []
layers.append(ResidualBlock(in_channels, out_channels, stride))
for _ in range(1, num_blocks):
layers.append(ResidualBlock(out_channels, out_channels, stride=1))
return nn.Sequential(*layers)
def forward(self, x, return_details=False):
batch_size = x.shape[0]
# CNN backbone
x = F.relu(self.bn1(self.conv1(x)))
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
cnn_features = self.global_pool(x).flatten(1)
cnn_features = self.dropout(cnn_features)
# Generate PE
positions = torch.arange(batch_size, device=x.device)
pe_levels, cantor_measures = self.pe(positions, seq_len=batch_size)
# Modulate PE with CNN features
modulation = self.pe_modulator(cnn_features)
modulation = modulation.view(batch_size, self.pe_levels, self.pe_features_per_level)
pe_levels = pe_levels + 0.1 * modulation
# Geometric basin compatibility
compatibility_scores = self.basin(pe_levels, cantor_measures)
if return_details:
return {
'compatibility_scores': compatibility_scores,
'pe_levels': pe_levels,
'cantor_measures': cantor_measures,
'cnn_features': cnn_features
}
return compatibility_scores |