File size: 45,625 Bytes
4dad82a 6534798 4dad82a 896c974 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 4dad82a 6534798 896c974 6534798 896c974 4dad82a 896c974 4dad82a 896c974 4dad82a 896c974 4dad82a 896c974 4dad82a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 |
# =====================================================================================
# SD1.5 Flow-Matching Trainer — David-Driven Adaptive Timestep Sampling
# Quartermaster: Mirel
# FIXED: David nested output handling + reliability filtering + clean checkpoint loading
# =====================================================================================
from __future__ import annotations
import os, json, math, random, re, shutil
from dataclasses import dataclass, asdict
from pathlib import Path
from typing import Dict, List, Tuple, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
# Diffusers
from diffusers import StableDiffusionPipeline, DDPMScheduler
from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
# Repo deps
from geovocab2.train.model.core.geo_david_collective import GeoDavidCollective
from geovocab2.data.prompt.symbolic_tree import SynthesisSystem
# HF / safetensors
from huggingface_hub import snapshot_download, HfApi, create_repo, hf_hub_download
from safetensors.torch import load_file
# =====================================================================================
# 1) CONFIG
# =====================================================================================
@dataclass
class BaseConfig:
run_name: str = "sd15_flowmatch_david_weighted"
out_dir: str = "./runs/sd15_flowmatch_david_weighted"
ckpt_dir: str = "./checkpoints_sd15_flow_david_weighted"
save_every: int = 1
# Data
num_samples: int = 200_000
batch_size: int = 32
num_workers: int = 2
seed: int = 42
# Models / Blocks
model_id: str = "runwayml/stable-diffusion-v1-5"
active_blocks: Tuple[str, ...] = ("down_0","down_1","down_2","down_3","mid","up_0","up_1","up_2","up_3")
pooling: str = "mean"
# Flow training
epochs: int = 20
lr: float = 1e-4
weight_decay: float = 1e-3
grad_clip: float = 1.0
amp: bool = True
global_flow_weight: float = 1.0
block_penalty_weight: float = 0.2
use_local_flow_heads: bool = False
local_flow_weight: float = 1.0
# KD
use_kd: bool = True
kd_weight: float = 0.25
# David
david_repo_id: str = "AbstractPhil/geo-david-collective-sd15-base-e40"
david_cache_dir: str = "./_hf_david_cache"
david_state_key: Optional[str] = None
# Fusion
alpha_timestep: float = 0.5
beta_pattern: float = 0.25
delta_incoherence: float = 0.25
lambda_min: float = 0.5
lambda_max: float = 3.0
block_weights: Dict[str, float] = None
# Timestep Weighting (David-guided adaptive sampling)
use_timestep_weighting: bool = True
use_david_weights: bool = True
timestep_shift: float = 3.0 # SD3-style shift (higher = bias toward clean)
base_jitter: int = 5 # Base ±jitter around bin center
adaptive_chaos: bool = True # Scale jitter by pattern difficulty
profile_samples: int = 2500 # Samples to profile David's difficulty
reliability_threshold: float = 0.15 # Minimum accuracy to trust David's guidance
# Scheduler
num_train_timesteps: int = 1000
# Inference
sample_steps: int = 30
guidance_scale: float = 7.5
# HuggingFace
hf_repo_id: Optional[str] = "AbstractPhil/sd15-flow-matching"
upload_every_epoch: bool = True
continue_training: bool = True
def __post_init__(self):
Path(self.out_dir).mkdir(parents=True, exist_ok=True)
Path(self.ckpt_dir).mkdir(parents=True, exist_ok=True)
Path(self.david_cache_dir).mkdir(parents=True, exist_ok=True)
if self.block_weights is None:
self.block_weights = {'down_0':0.7,'down_1':0.9,'down_2':1.0,'down_3':1.1,'mid':1.2,'up_0':1.1,'up_1':1.0,'up_2':0.9,'up_3':0.7}
# =====================================================================================
# 2) DAVID-WEIGHTED TIMESTEP SAMPLER
# =====================================================================================
class DavidWeightedTimestepSampler:
"""
Samples timesteps weighted by David's inherent difficulty + SD3 shift + adaptive chaos.
FIXED: Properly handles nested GeoDavidCollective output structure.
FIXED: Filters out unreliable bins (accuracy < threshold).
"""
def __init__(self, num_timesteps=1000, num_bins=100, shift=3.0, base_jitter=5, adaptive_chaos=True, reliability_threshold=0.15):
self.num_timesteps = num_timesteps
self.num_bins = num_bins
self.shift = shift
self.base_jitter = base_jitter
self.adaptive_chaos = adaptive_chaos
self.reliability_threshold = reliability_threshold
self.difficulty_weights = None # Timestep difficulty
self.pattern_difficulty = None # Pattern confusion per bin
def _apply_shift(self, t: float) -> float:
"""Apply SD3-style timestep shift (operates on normalized t ∈ [0,1])."""
if self.shift <= 0:
return t
return self.shift * t / (1.0 + (self.shift - 1.0) * t)
def compute_difficulty_from_david(self, david, teacher, device, num_samples=500):
"""Profile David's confusion patterns to create difficulty map."""
print("🔍 Profiling David's timestep & pattern difficulty...")
david.eval()
teacher.eval()
# Track David's accuracy and pattern entropy per bin
correct_per_bin = torch.zeros(self.num_bins)
total_per_bin = torch.zeros(self.num_bins)
entropy_per_bin = torch.zeros(self.num_bins)
entropy_count_per_bin = torch.zeros(self.num_bins)
with torch.no_grad():
for _ in tqdm(range(num_samples // 32), desc="Profiling David", leave=False):
# Random latents and timesteps
x = torch.randn(32, 4, 64, 64, device=device, dtype=torch.float16)
t = torch.randint(0, self.num_timesteps, (32,), device=device)
t_bins = (t // 10)
# Dummy conditioning
ehs = torch.randn(32, 77, 768, device=device, dtype=torch.float16)
# Get teacher features
teacher.hooks.clear()
_ = teacher.unet(x, t, encoder_hidden_states=ehs)
feats = {k: v.float() for k, v in teacher.hooks.bank.items()}
# Pool features
pooled = {name: f.mean(dim=(2, 3)) for name, f in feats.items()}
# Get David's outputs (NESTED STRUCTURE!)
outputs = david(pooled, t.float())
# ================================================================
# FIXED: Aggregate across blocks
# ================================================================
# 1. Timestep difficulty (from classification error)
timestep_logits_list = []
for block_name, block_out in outputs.items():
if 'timestep_logits' in block_out:
timestep_logits_list.append(block_out['timestep_logits'])
if timestep_logits_list:
# Average predictions across blocks
ts_logits = torch.stack(timestep_logits_list).mean(0)
preds = ts_logits.argmax(dim=-1)
for pred, true_bin in zip(preds, t_bins):
bin_idx = true_bin.item()
correct_per_bin[bin_idx] += (pred == true_bin).float().item()
total_per_bin[bin_idx] += 1
# 2. Pattern difficulty (from entropy)
pattern_logits_list = []
for block_name, block_out in outputs.items():
if 'pattern_logits' in block_out:
pattern_logits_list.append(block_out['pattern_logits'])
if pattern_logits_list:
# Average predictions across blocks
pt_logits = torch.stack(pattern_logits_list).mean(0)
P = pt_logits.softmax(-1)
ent = -(P * P.clamp_min(1e-9).log()).sum(-1)
norm_ent = ent / math.log(P.shape[-1]) # Normalize by max entropy
for i, true_bin in enumerate(t_bins):
bin_idx = true_bin.item()
entropy_per_bin[bin_idx] += norm_ent[i].item()
entropy_count_per_bin[bin_idx] += 1
# Compute accuracy per bin
accuracy_per_bin = correct_per_bin / (total_per_bin.clamp(min=1))
# ========================================================================
# RELIABILITY FILTERING: Disable bins with accuracy < threshold
# ========================================================================
reliable_mask = accuracy_per_bin >= self.reliability_threshold
num_reliable = reliable_mask.sum().item()
num_disabled = self.num_bins - num_reliable
print(f"\n🎯 Reliability Analysis:")
print(f" Threshold: {self.reliability_threshold:.0%}")
print(f" Reliable bins: {num_reliable}/{self.num_bins}")
print(f" Disabled bins: {num_disabled}/{self.num_bins}")
if num_disabled > 0:
disabled_bins = torch.where(~reliable_mask)[0].tolist()
disabled_accs = [accuracy_per_bin[i].item() for i in disabled_bins]
print(f" Disabled: {disabled_bins[:10]}{'...' if len(disabled_bins) > 10 else ''}")
print(f" (accuracies: {[f'{a:.1%}' for a in disabled_accs[:10]]})")
# Create difficulty weights ONLY for reliable bins
if num_reliable == 0:
print("\n⚠️ WARNING: No reliable bins found! Falling back to uniform sampling.")
self.difficulty_weights = torch.ones(self.num_bins) / self.num_bins
self.pattern_difficulty = torch.ones(self.num_bins) * 0.5
return self.difficulty_weights
# Compute difficulty (inverse accuracy) for reliable bins
timestep_difficulty = torch.zeros(self.num_bins)
timestep_difficulty[reliable_mask] = (1.0 - accuracy_per_bin[reliable_mask]) + 0.1
# Zero out unreliable bins (won't be sampled)
timestep_difficulty[~reliable_mask] = 0.0
# Normalize weights over reliable bins only
self.difficulty_weights = timestep_difficulty / timestep_difficulty.sum()
# Compute pattern difficulty (average entropy per bin)
self.pattern_difficulty = entropy_per_bin / (entropy_count_per_bin.clamp(min=1))
self.pattern_difficulty = self.pattern_difficulty.clamp(min=0.1, max=1.0)
# Set entropy to 0.5 (neutral) for disabled bins
self.pattern_difficulty[~reliable_mask] = 0.5
# ========================================================================
# REPORT
# ========================================================================
print(f"\n✓ David difficulty map computed (filtered):")
print(f" Avg timestep accuracy (all bins): {accuracy_per_bin.mean():.2%}")
print(f" Avg timestep accuracy (reliable): {accuracy_per_bin[reliable_mask].mean():.2%}")
# Find hardest/easiest among reliable bins
reliable_indices = torch.where(reliable_mask)[0]
if len(reliable_indices) > 0:
hardest_idx = reliable_indices[accuracy_per_bin[reliable_mask].argmin()].item()
easiest_idx = reliable_indices[accuracy_per_bin[reliable_mask].argmax()].item()
print(f" Hardest reliable bin: {hardest_idx} ({accuracy_per_bin[hardest_idx]:.2%} acc)")
print(f" Easiest reliable bin: {easiest_idx} ({accuracy_per_bin[easiest_idx]:.2%} acc)")
print(f" Avg pattern entropy (reliable): {self.pattern_difficulty[reliable_mask].mean():.3f}")
# Show sampling distribution (top 10 weighted bins)
top_weights, top_bins = self.difficulty_weights.topk(10)
print(f"\n📊 Top 10 sampled bins (by difficulty weight):")
for i, (bin_idx, weight) in enumerate(zip(top_bins.tolist(), top_weights.tolist())):
acc = accuracy_per_bin[bin_idx].item()
print(f" {i+1}. Bin {bin_idx:2d}: weight={weight:.3f} (acc={acc:.1%})")
return self.difficulty_weights
def sample(self, batch_size: int) -> List[int]:
"""Sample timesteps with David weighting + shift + adaptive chaos."""
if self.difficulty_weights is None:
# Fallback to uniform
return [random.randint(0, self.num_timesteps - 1) for _ in range(batch_size)]
timesteps = []
for _ in range(batch_size):
# 1. Sample bin weighted by David's difficulty
bin_idx = torch.multinomial(self.difficulty_weights, 1).item()
# 2. Get bin center as normalized t
bin_center_raw = bin_idx * (self.num_timesteps // self.num_bins) + (self.num_timesteps // self.num_bins) // 2
t_normalized = bin_center_raw / self.num_timesteps
# 3. Apply SD3 shift
t_shifted = self._apply_shift(t_normalized)
# 4. Add adaptive chaos (jitter scaled by pattern difficulty)
if self.adaptive_chaos:
chaos_scale = self.pattern_difficulty[bin_idx].item()
jitter = int(self.base_jitter * (0.5 + chaos_scale)) # 0.5-1.5x base jitter
else:
jitter = self.base_jitter
# 5. Convert back to raw timestep with jitter
t_raw = int(t_shifted * self.num_timesteps)
t_raw += random.randint(-jitter, jitter)
t_raw = max(0, min(self.num_timesteps - 1, t_raw))
timesteps.append(t_raw)
return timesteps
# =====================================================================================
# 3) DATA
# =====================================================================================
class SymbolicPromptDataset(Dataset):
def __init__(self, n:int, seed:int=42, timestep_sampler=None):
self.n = n
self.timestep_sampler = timestep_sampler
random.seed(seed)
self.sys = SynthesisSystem(seed=seed)
def __len__(self): return self.n
def __getitem__(self, idx):
r = self.sys.synthesize(complexity=random.choice([1,2,3,4,5]))
prompt = r['text']
if self.timestep_sampler:
t = self.timestep_sampler.sample(1)[0]
else:
t = random.randint(0, 999)
return {"prompt": prompt, "t": t}
def collate(batch: List[dict]):
prompts = [b["prompt"] for b in batch]
t = torch.tensor([b["t"] for b in batch], dtype=torch.long)
t_bins = t // 10
return {"prompts": prompts, "t": t, "t_bins": t_bins}
# =====================================================================================
# 4) HOOKS + POOLING
# =====================================================================================
class HookBank:
def __init__(self, unet: UNet2DConditionModel, active: Tuple[str, ...]):
self.active = set(active)
self.bank: Dict[str, torch.Tensor] = {}
self.hooks: List[torch.utils.hooks.RemovableHandle] = []
self._register(unet)
def _register(self, unet: UNet2DConditionModel):
def mk(name):
def h(m, i, o):
out = o[0] if isinstance(o,(tuple,list)) else o
self.bank[name] = out
return h
for i, blk in enumerate(unet.down_blocks):
nm = f"down_{i}"
if nm in self.active: self.hooks.append(blk.register_forward_hook(mk(nm)))
if "mid" in self.active:
self.hooks.append(unet.mid_block.register_forward_hook(mk("mid")))
for i, blk in enumerate(unet.up_blocks):
nm = f"up_{i}"
if nm in self.active: self.hooks.append(blk.register_forward_hook(mk(nm)))
def clear(self): self.bank.clear()
def close(self):
for h in self.hooks: h.remove()
self.hooks.clear()
def spatial_pool(x: torch.Tensor, name: str, policy: str) -> torch.Tensor:
if policy == "mean": return x.mean(dim=(2,3))
if policy == "max": return x.amax(dim=(2,3))
if policy == "adaptive":
return x.mean(dim=(2,3)) if (name.startswith("down") or name=="mid") else x.amax(dim=(2,3))
raise ValueError(f"Unknown pooling: {policy}")
# =====================================================================================
# 5) TEACHER
# =====================================================================================
class SD15Teacher(nn.Module):
def __init__(self, cfg: BaseConfig, device: str):
super().__init__()
self.pipe = StableDiffusionPipeline.from_pretrained(cfg.model_id, torch_dtype=torch.float16, safety_checker=None).to(device)
self.unet: UNet2DConditionModel = self.pipe.unet
self.text_encoder = self.pipe.text_encoder
self.tokenizer = self.pipe.tokenizer
self.hooks = HookBank(self.unet, cfg.active_blocks)
self.sched = DDPMScheduler(num_train_timesteps=cfg.num_train_timesteps)
self.device = device
for p in self.parameters(): p.requires_grad_(False)
@torch.no_grad()
def encode(self, prompts: List[str]) -> torch.Tensor:
tok = self.tokenizer(prompts, padding="max_length", max_length=self.tokenizer.model_max_length,
truncation=True, return_tensors="pt")
return self.text_encoder(tok.input_ids.to(self.device))[0]
@torch.no_grad()
def forward_eps_and_feats(self, x_t: torch.Tensor, t: torch.LongTensor, ehs: torch.Tensor):
self.hooks.clear()
eps_hat = self.unet(x_t, t, encoder_hidden_states=ehs).sample
feats = {k: v.detach().float() for k, v in self.hooks.bank.items()}
return eps_hat.float(), feats
def alpha_sigma(self, t: torch.LongTensor) -> Tuple[torch.Tensor, torch.Tensor]:
ac = self.sched.alphas_cumprod.to(self.device)[t]
alpha = ac.sqrt().view(-1,1,1,1).float()
sigma = (1.0 - ac).sqrt().view(-1,1,1,1).float()
return alpha, sigma
# =====================================================================================
# 6) STUDENT
# =====================================================================================
class StudentUNet(nn.Module):
def __init__(self, teacher_unet: UNet2DConditionModel, active_blocks: Tuple[str,...], use_local_heads: bool):
super().__init__()
self.unet = UNet2DConditionModel.from_config(teacher_unet.config)
self.unet.load_state_dict(teacher_unet.state_dict(), strict=True)
self.hooks = HookBank(self.unet, active_blocks)
self.use_local_heads = use_local_heads
self.local_heads = nn.ModuleDict()
def _ensure_heads(self, feats: Dict[str, torch.Tensor]):
if not self.use_local_heads: return
if len(self.local_heads) == len(feats): return
target_dtype = next(self.unet.parameters()).dtype
for name, f in feats.items():
c = f.shape[1]
if name not in self.local_heads:
head = nn.Conv2d(c, 4, kernel_size=1)
head = head.to(dtype=target_dtype, device=f.device)
self.local_heads[name] = head
def forward(self, x_t: torch.Tensor, t: torch.LongTensor, ehs: torch.Tensor):
self.hooks.clear()
v_hat = self.unet(x_t, t, encoder_hidden_states=ehs).sample
feats = {k: v for k, v in self.hooks.bank.items()}
self._ensure_heads(feats)
return v_hat, feats
# =====================================================================================
# 7) DAVID + ASSESSOR + FUSION
# =====================================================================================
class DavidLoader:
def __init__(self, cfg: BaseConfig, device: str):
self.cfg = cfg
self.device = device
self.repo_dir = snapshot_download(repo_id=cfg.david_repo_id, local_dir=cfg.david_cache_dir, local_dir_use_symlinks=False)
self.config_path = os.path.join(self.repo_dir, "config.json")
self.weights_path = os.path.join(self.repo_dir, "model.safetensors")
with open(self.config_path, "r") as f:
self.hf_config = json.load(f)
self.gdc = GeoDavidCollective(
block_configs=self.hf_config["block_configs"],
num_timestep_bins=int(self.hf_config["num_timestep_bins"]),
num_patterns_per_bin=int(self.hf_config["num_patterns_per_bin"]),
block_weights=self.hf_config.get("block_weights", {k:1.0 for k in self.hf_config["block_configs"].keys()}),
loss_config=self.hf_config.get("loss_config", {})
).to(device).eval()
state = load_file(self.weights_path)
self.gdc.load_state_dict(state, strict=False)
for p in self.gdc.parameters(): p.requires_grad_(False)
print(f"✓ David loaded from HF: {self.repo_dir}")
print(f" blocks={len(self.hf_config['block_configs'])} bins={self.hf_config['num_timestep_bins']} patterns={self.hf_config['num_patterns_per_bin']}")
if "block_weights" in self.hf_config:
cfg.block_weights = self.hf_config["block_weights"]
class DavidAssessor(nn.Module):
"""
CORRECTED: Properly handles GeoDavidCollective's nested multi-block output structure.
GeoDavidCollective returns: Dict[block_name, Dict[str, Tensor]]
Not a flat Dict[str, Tensor]!
"""
def __init__(self, gdc: GeoDavidCollective, pooling: str):
super().__init__()
self.gdc = gdc
self.pooling = pooling
def _pool(self, feats: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
return {k: spatial_pool(v, k, self.pooling) for k, v in feats.items()}
@torch.no_grad()
def forward(self, feats_student: Dict[str, torch.Tensor], t: torch.LongTensor
) -> Tuple[Dict[str,float], Dict[str,float], Dict[str,float]]:
"""
Assess student features using David's geometric knowledge.
Returns:
e_t: Dict[block_name, timestep_error] - classification error per block
e_p: Dict[block_name, pattern_entropy] - normalized entropy per block
coh: Dict[block_name, coherence] - geometric coherence per block
"""
# Pool spatial features
Zs = self._pool(feats_student)
# Forward through GeoDavidCollective
# Returns: Dict[block_name, Dict[str, Tensor]]
outs = self.gdc(Zs, t.float())
# Initialize output dicts
e_t, e_p, coh = {}, {}, {}
# Compute timestep bins for targets
t_bins = (t // 10).to(next(self.gdc.parameters()).device)
# ====================================================================
# TIMESTEP ERROR - Per-block
# ====================================================================
for block_name, block_out in outs.items():
if 'timestep_logits' in block_out:
ts_logits = block_out['timestep_logits']
ce = F.cross_entropy(ts_logits, t_bins, reduction="mean")
e_t[block_name] = float(ce.item())
# If no timestep predictions, set all errors to 0
if not e_t:
for name in Zs.keys():
e_t[name] = 0.0
# ====================================================================
# PATTERN ENTROPY - Per-block
# ====================================================================
for block_name, block_out in outs.items():
if 'pattern_logits' in block_out:
pt_logits = block_out['pattern_logits']
# Compute normalized entropy
P = pt_logits.softmax(-1)
ent = -(P * (P.clamp_min(1e-9)).log()).sum(-1).mean()
norm_ent = ent / math.log(P.shape[-1]) # Normalize by max entropy
e_p[block_name] = float(norm_ent.item())
# If no pattern predictions, set all entropies to 0
if not e_p:
for name in Zs.keys():
e_p[name] = 0.0
# ====================================================================
# COHERENCE (from Cantor alphas)
# ====================================================================
try:
alphas = self.gdc.get_cantor_alphas()
# Alphas should be close to 0.5 for good coherence
# Map to coherence: 1.0 = perfect (alpha=0.5), lower = worse
for name, alpha in alphas.items():
# Coherence = 1 - 2*|alpha - 0.5|
# When alpha=0.5: coherence=1.0
# When alpha=0 or 1: coherence=0.0
coherence = 1.0 - 2.0 * abs(alpha - 0.5)
coh[name] = max(0.0, min(1.0, coherence))
except Exception:
# Fallback: assume perfect coherence
for name in Zs.keys():
coh[name] = 1.0
# Ensure all input blocks have values (fill missing with block averages)
for name in Zs.keys():
if name not in e_t:
# Use average of available blocks
e_t[name] = sum(e_t.values()) / max(len(e_t), 1) if e_t else 0.0
if name not in e_p:
e_p[name] = sum(e_p.values()) / max(len(e_p), 1) if e_p else 0.0
if name not in coh:
coh[name] = sum(coh.values()) / max(len(coh), 1) if coh else 1.0
return e_t, e_p, coh
class BlockPenaltyFusion:
def __init__(self, cfg: BaseConfig): self.cfg = cfg
def lambdas(self, e_t:Dict[str,float], e_p:Dict[str,float], coh:Dict[str,float]) -> Dict[str,float]:
lam = {}
for name, base in self.cfg.block_weights.items():
val = base * (1.0
+ self.cfg.alpha_timestep * float(e_t.get(name,0.0))
+ self.cfg.beta_pattern * float(e_p.get(name,0.0))
+ self.cfg.delta_incoherence * (1.0 - float(coh.get(name,1.0))))
lam[name] = float(max(self.cfg.lambda_min, min(self.cfg.lambda_max, val)))
return lam
# =====================================================================================
# 8) TRAINER
# =====================================================================================
class FlowMatchDavidTrainer:
def __init__(self, cfg: BaseConfig, device: str = "cuda"):
self.cfg = cfg
self.device = device
self.start_epoch = 0
self.start_gstep = 0
# Initialize David first (needed for timestep sampler)
self.david_loader = DavidLoader(cfg, device)
self.david = self.david_loader.gdc
self.assessor = DavidAssessor(self.david, cfg.pooling)
self.fusion = BlockPenaltyFusion(cfg)
# Initialize teacher (needed for David profiling)
self.teacher = SD15Teacher(cfg, device).eval()
# Initialize timestep sampler
self.timestep_sampler = None
if cfg.use_timestep_weighting:
print("\n" + "="*70)
print("🎯 ADAPTIVE TIMESTEP SAMPLING ENABLED")
print(f" David weighting: {cfg.use_david_weights}")
print(f" SD3 shift: {cfg.timestep_shift}")
print(f" Base jitter: ±{cfg.base_jitter}")
print(f" Adaptive chaos: {cfg.adaptive_chaos}")
print(f" Reliability threshold: {cfg.reliability_threshold:.0%}")
self.timestep_sampler = DavidWeightedTimestepSampler(
num_timesteps=cfg.num_train_timesteps,
num_bins=100,
shift=cfg.timestep_shift if cfg.use_david_weights else 0.0,
base_jitter=cfg.base_jitter,
adaptive_chaos=cfg.adaptive_chaos,
reliability_threshold=cfg.reliability_threshold
)
if cfg.use_david_weights:
self.timestep_sampler.compute_difficulty_from_david(
david=self.david,
teacher=self.teacher,
device=device,
num_samples=cfg.profile_samples
)
print("="*70 + "\n")
# Initialize dataset with sampler
self.dataset = SymbolicPromptDataset(cfg.num_samples, cfg.seed, self.timestep_sampler)
self.loader = DataLoader(self.dataset, batch_size=cfg.batch_size, shuffle=True,
num_workers=cfg.num_workers, pin_memory=True, collate_fn=collate)
# Initialize student
self.student = StudentUNet(self.teacher.unet, cfg.active_blocks, cfg.use_local_flow_heads).to(device)
self.opt = torch.optim.AdamW(self.student.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay)
self.sched = torch.optim.lr_scheduler.CosineAnnealingLR(self.opt, T_max=cfg.epochs * len(self.loader))
self.scaler = torch.cuda.amp.GradScaler(enabled=cfg.amp)
# Load latest checkpoint from HuggingFace if continuing training
if cfg.continue_training:
self._load_latest_from_hf()
self.writer = SummaryWriter(log_dir=os.path.join(cfg.out_dir, cfg.run_name))
def _load_latest_from_hf(self):
"""Load the most recent checkpoint from HuggingFace repo."""
if not self.cfg.hf_repo_id:
print("ℹ️ No HuggingFace repo specified, starting from scratch\n")
return
try:
api = HfApi()
print(f"\n🔍 Searching for latest checkpoint in {self.cfg.hf_repo_id}...")
# List all files in the repo
files = api.list_repo_files(repo_id=self.cfg.hf_repo_id, repo_type="model")
# Find all epoch checkpoints (format: {run_name}_e{epoch}.pt)
epochs = []
for f in files:
if f.endswith('.pt') and 'final' not in f.lower():
match = re.search(r'_e(\d+)\.pt$', f)
if match:
epoch_num = int(match.group(1))
epochs.append((epoch_num, f))
if not epochs:
print("ℹ️ No previous checkpoints found, starting from scratch\n")
return
# Get the latest epoch
latest_epoch, latest_file = max(epochs, key=lambda x: x[0])
print(f"📥 Found latest checkpoint: {latest_file} (epoch {latest_epoch})")
# Download checkpoint
local_path = hf_hub_download(
repo_id=self.cfg.hf_repo_id,
filename=latest_file,
repo_type="model",
cache_dir=self.cfg.ckpt_dir
)
# Load checkpoint
print(f"📦 Loading checkpoint...")
checkpoint = torch.load(local_path, map_location='cpu')
# Load student state dict
if 'student' in checkpoint:
missing, unexpected = self.student.load_state_dict(checkpoint['student'], strict=False)
if missing:
print(f" ⚠️ Missing keys: {len(missing)}")
if unexpected:
print(f" ⚠️ Unexpected keys: {len(unexpected)}")
print(f" ✓ Loaded student model")
else:
print(f" ⚠️ Warning: 'student' key not found in checkpoint")
return
# Load optimizer state
if 'opt' in checkpoint:
try:
self.opt.load_state_dict(checkpoint['opt'])
print(" ✓ Loaded optimizer state")
except Exception as e:
print(f" ⚠️ Failed to load optimizer state: {e}")
# Load scheduler state
if 'sched' in checkpoint:
try:
self.sched.load_state_dict(checkpoint['sched'])
print(" ✓ Loaded scheduler state")
except Exception as e:
print(f" ⚠️ Failed to load scheduler state: {e}")
# Set starting epoch and global step
if 'gstep' in checkpoint:
self.start_gstep = checkpoint['gstep']
self.start_epoch = latest_epoch
print(f" ✓ Resuming from epoch {self.start_epoch + 1}, global step {self.start_gstep}")
else:
# Fallback: estimate from epoch number
self.start_epoch = latest_epoch
self.start_gstep = latest_epoch * len(self.loader)
print(f" ✓ Resuming from epoch {self.start_epoch + 1} (estimated step {self.start_gstep})")
# Cleanup
del checkpoint
torch.cuda.empty_cache()
print(f"✅ Successfully resumed from checkpoint!\n")
except Exception as e:
print(f"⚠️ Failed to load checkpoint: {e}")
print(" Starting training from scratch...\n")
def _v_star(self, x_t, t, eps_hat):
alpha, sigma = self.teacher.alpha_sigma(t)
x0_hat = (x_t - sigma * eps_hat) / (alpha + 1e-8)
return alpha * eps_hat - sigma * x0_hat
def _down_like(self, tgt: torch.Tensor, ref: torch.Tensor) -> torch.Tensor:
return F.interpolate(tgt, size=ref.shape[-2:], mode="bilinear", align_corners=False)
def _kd_cos(self, s: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
s = F.normalize(s, dim=-1); t = F.normalize(t, dim=-1)
return 1.0 - (s*t).sum(-1).mean()
def train(self):
cfg = self.cfg
gstep = self.start_gstep
# Test prompts for monitoring progress
test_prompts = [
"a castle at sunset",
"a mountain landscape with trees",
"a city street at night"
]
for ep in range(self.start_epoch, cfg.epochs):
# Sample before epoch to monitor progress
if ep > 0 or self.start_epoch > 0: # Skip first ever epoch
print(f"\n🎨 Sampling test images before epoch {ep+1}...")
try:
test_imgs = self.sample(test_prompts, steps=30, guidance=7.5)
# Save individual images
sample_dir = Path(cfg.out_dir) / "samples"
sample_dir.mkdir(exist_ok=True, parents=True)
for i, (img, prompt) in enumerate(zip(test_imgs, test_prompts)):
# Convert to PIL
img_np = ((img.cpu().permute(1,2,0).numpy() + 1) / 2 * 255).astype('uint8')
from PIL import Image
pil_img = Image.fromarray(img_np)
# Save with epoch number
safe_prompt = prompt.replace(" ", "_")[:30]
img_path = sample_dir / f"e{ep}_p{i}_{safe_prompt}.png"
pil_img.save(img_path)
# Log to tensorboard
self.writer.add_image(f"samples/{safe_prompt}",
(img + 1) / 2, # Normalize to [0,1]
global_step=ep)
print(f"✓ Saved {len(test_imgs)} test images to {sample_dir}")
except Exception as e:
print(f"⚠️ Sampling failed: {e}")
self.student.train()
pbar = tqdm(self.loader, desc=f"Epoch {ep+1}/{cfg.epochs}",
dynamic_ncols=True, leave=True, position=0)
acc = {"L":0.0, "Lf":0.0, "Lb":0.0}
for it, batch in enumerate(pbar):
prompts = batch["prompts"]
t = batch["t"].to(self.device)
with torch.no_grad():
ehs = self.teacher.encode(prompts)
x_t = torch.randn(len(prompts), 4, 64, 64, device=self.device, dtype=torch.float16)
with torch.no_grad():
eps_hat, t_feats_spatial = self.teacher.forward_eps_and_feats(x_t.half(), t, ehs)
v_star = self._v_star(x_t, t, eps_hat)
with torch.cuda.amp.autocast(enabled=cfg.amp):
v_hat, s_feats_spatial = self.student(x_t, t, ehs)
L_flow = F.mse_loss(v_hat, v_star)
e_t, e_p, coh = self.assessor(s_feats_spatial, t)
lam = self.fusion.lambdas(e_t, e_p, coh)
L_blocks = torch.zeros((), device=self.device)
for name, s_feat in s_feats_spatial.items():
L_kd = torch.zeros((), device=self.device)
if cfg.use_kd:
s_pool = spatial_pool(s_feat, name, cfg.pooling)
t_pool = spatial_pool(t_feats_spatial[name], name, cfg.pooling)
L_kd = self._kd_cos(s_pool, t_pool)
L_lf = torch.zeros((), device=self.device)
if cfg.use_local_flow_heads and name in self.student.local_heads:
v_loc = self.student.local_heads[name](s_feat)
v_ds = self._down_like(v_star, v_loc)
L_lf = F.mse_loss(v_loc, v_ds)
L_blocks = L_blocks + lam.get(name,1.0) * (cfg.kd_weight * L_kd + cfg.local_flow_weight * L_lf)
L_total = cfg.global_flow_weight*L_flow + cfg.block_penalty_weight*L_blocks
self.opt.zero_grad(set_to_none=True)
if cfg.amp:
self.scaler.scale(L_total).backward()
nn.utils.clip_grad_norm_(self.student.parameters(), cfg.grad_clip)
self.scaler.step(self.opt); self.scaler.update()
else:
L_total.backward()
nn.utils.clip_grad_norm_(self.student.parameters(), cfg.grad_clip)
self.opt.step()
self.sched.step(); gstep += 1
acc["L"] += float(L_total.item())
acc["Lf"] += float(L_flow.item())
acc["Lb"] += float(L_blocks.item())
if it % 50 == 0:
self.writer.add_scalar("train/total", float(L_total.item()), gstep)
self.writer.add_scalar("train/flow", float(L_flow.item()), gstep)
self.writer.add_scalar("train/blocks",float(L_blocks.item()), gstep)
for k in list(lam.keys())[:4]:
self.writer.add_scalar(f"lambda/{k}", lam[k], gstep)
if it % 10 == 0 or it == len(self.loader) - 1:
pbar.set_postfix({
"L": f"{float(L_total.item()):.4f}",
"Lf": f"{float(L_flow.item()):.4f}",
"Lb": f"{float(L_blocks.item()):.4f}"
}, refresh=False)
del x_t, eps_hat, v_star, v_hat, s_feats_spatial, t_feats_spatial
pbar.close()
n = len(self.loader)
print(f"\n[Epoch {ep+1}] L={acc['L']/n:.4f} | L_flow={acc['Lf']/n:.4f} | L_blocks={acc['Lb']/n:.4f}")
self.writer.add_scalar("epoch/total", acc['L']/n, ep+1)
self.writer.add_scalar("epoch/flow", acc['Lf']/n, ep+1)
self.writer.add_scalar("epoch/blocks",acc['Lb']/n, ep+1)
if (ep+1) % cfg.save_every == 0:
self._save(ep+1, gstep)
self._save("final", gstep)
# Final comprehensive sampling
print("\n🎨 Generating final test samples...")
final_prompts = [
"a castle at sunset",
"a mountain landscape with trees",
"a city street at night",
"a portrait of a person",
"abstract geometric shapes"
]
try:
final_imgs = self.sample(final_prompts, steps=30, guidance=7.5)
sample_dir = Path(cfg.out_dir) / "samples"
sample_dir.mkdir(exist_ok=True, parents=True)
for i, (img, prompt) in enumerate(zip(final_imgs, final_prompts)):
from PIL import Image
img_np = ((img.cpu().permute(1,2,0).numpy() + 1) / 2 * 255).astype('uint8')
pil_img = Image.fromarray(img_np)
safe_prompt = prompt.replace(" ", "_")[:30]
pil_img.save(sample_dir / f"final_{safe_prompt}.png")
print(f"✓ Saved {len(final_imgs)} final images to {sample_dir}")
except Exception as e:
print(f"⚠️ Final sampling failed: {e}")
self.writer.close()
def _save(self, tag, gstep):
"""Save checkpoint and upload to HuggingFace."""
pt_path = Path(self.cfg.ckpt_dir) / f"{self.cfg.run_name}_e{tag}.pt"
torch.save({
"cfg": asdict(self.cfg),
"student": self.student.state_dict(),
"opt": self.opt.state_dict(),
"sched": self.sched.state_dict(),
"gstep": gstep
}, pt_path)
size_mb = pt_path.stat().st_size / 1e6
print(f"✓ Saved checkpoint: {pt_path.name} ({size_mb:.1f} MB)")
if self.cfg.upload_every_epoch and self.cfg.hf_repo_id:
self._upload_to_hf(pt_path, tag)
def _upload_to_hf(self, path: Path, tag):
"""Upload checkpoint to HuggingFace."""
try:
api = HfApi()
create_repo(self.cfg.hf_repo_id, exist_ok=True, private=False, repo_type="model")
print(f"📤 Uploading {path.name} to {self.cfg.hf_repo_id}...")
api.upload_file(
path_or_fileobj=str(path),
path_in_repo=path.name,
repo_id=self.cfg.hf_repo_id,
repo_type="model",
commit_message=f"Epoch {tag}"
)
print(f"✅ Uploaded: https://huggingface.co/{self.cfg.hf_repo_id}/{path.name}")
except Exception as e:
print(f"⚠️ Upload failed: {e}")
@torch.no_grad()
def sample(self, prompts: List[str], steps: Optional[int]=None, guidance: Optional[float]=None) -> torch.Tensor:
steps = steps or self.cfg.sample_steps
guidance = guidance if guidance is not None else self.cfg.guidance_scale
# Ensure student is in eval mode
was_training = self.student.training
self.student.eval()
# Use autocast to handle dtype conversions automatically
with torch.cuda.amp.autocast(enabled=self.cfg.amp):
cond_e = self.teacher.encode(prompts)
uncond_e = self.teacher.encode([""]*len(prompts))
sched = self.teacher.sched
sched.set_timesteps(steps, device=self.device)
# Create latents (autocast will handle dtype)
x_t = torch.randn(len(prompts), 4, 64, 64, device=self.device)
for t_scalar in sched.timesteps:
t = torch.full((x_t.shape[0],), t_scalar, device=self.device, dtype=torch.long)
v_u, _ = self.student(x_t, t, uncond_e)
v_c, _ = self.student(x_t, t, cond_e)
v_hat = v_u + guidance*(v_c - v_u)
alpha, sigma = self.teacher.alpha_sigma(t)
denom = (alpha**2 + sigma**2)
x0_hat = (alpha * x_t - sigma * v_hat) / (denom + 1e-8)
eps_hat = (x_t - alpha * x0_hat) / (sigma + 1e-8)
step = sched.step(model_output=eps_hat, timestep=t_scalar, sample=x_t)
x_t = step.prev_sample
# Decode (keep x_t at current dtype for VAE)
imgs = self.teacher.pipe.vae.decode(x_t / 0.18215).sample
# Restore training mode
if was_training:
self.student.train()
return imgs.clamp(-1,1)
# =====================================================================================
# 9) MAIN
# =====================================================================================
def main():
cfg = BaseConfig()
print(json.dumps(asdict(cfg), indent=2))
device = "cuda" if torch.cuda.is_available() else "cpu"
if device != "cuda":
print("⚠️ A100 strongly recommended.")
trainer = FlowMatchDavidTrainer(cfg, device=device)
trainer.train()
_ = trainer.sample(["a castle at sunset"], steps=10, guidance=7.0)
print("✓ Training complete.")
if __name__ == "__main__":
main() |