File size: 19,368 Bytes
eea47ad | 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 | """Full CTA TRAINING-step trace — covers everything that's MISSING from
trace_cta_forward.py:
1. Constructs a mini-batch: 1 real + 1 fake from HDTF-paired (so they share
the same driving audio — the structural premise of CTA).
2. Walks through `predict_pairs()` exposing the cross-attention internals
(Q/K/V shapes, attention map shape) of f_{A→V} and f_{V→A}.
3. Computes ALL FIVE losses exactly as `CTALitModule.training_step` does:
loss_av (REAL-only, drives predictors + backbones)
loss_va (REAL-only, drives predictors + backbones)
loss_asym (BOTH, margin loss on the asymmetry score)
loss_cls (BOTH, BCE on classifier; L_AV/L_VA are detach()ed)
loss_aux (FAKE-paired, cross-generator asym consistency)
Each loss's value, sample-mask, gradient destination is logged.
4. Performs ONE backward step and reports which parameters received non-zero
gradients per loss — proving "predictor is trained on real ONLY",
"detach() blocks BCE → predictor", etc.
This is a debug / paper-figure-aid script; not used during training.
Usage:
/opt/conda/envs/pytorch/bin/python3 scripts/analysis/trace_cta_training.py \\
--ckpt outputs/cta_diffusion_combined_20260604_205145/checkpoints/epoch16-valauc1.0000.ckpt
"""
from __future__ import annotations
import argparse
import os
import sys
import textwrap
from pathlib import Path
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
# silence weights_only restriction
import lightning_fabric.utilities.cloud_io as _lf_cloud_io
_orig_torch_load = torch.load
def _unsafe_torch_load(*args, **kwargs):
kwargs["weights_only"] = False
return _orig_torch_load(*args, **kwargs)
_lf_cloud_io.torch.load = _unsafe_torch_load
torch.load = _unsafe_torch_load
sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
from omegaconf import OmegaConf
from src.data.fairtalking_dataset import load_video_clip, load_audio_clip
from src.data.transforms import build_video_transform
from src.methods import build_method
# ============================================================================
# Pretty-print helpers
# ============================================================================
HRULE = "─" * 92
DHRULE = "═" * 92
def banner(title: str, level: int = 0):
print()
print(DHRULE if level == 0 else HRULE)
print(f" {title}")
print(DHRULE if level == 0 else HRULE)
def step(n, title):
print(f"\n ── Step {n}: {title} ".ljust(94, "─"))
def tprint(name, x, indent=2, more=""):
sp = " " * indent
if isinstance(x, torch.Tensor):
info = (f"shape={tuple(x.shape)} dtype={str(x.dtype).replace('torch.', '')} "
f"range=[{x.min().item():+.4f}, {x.max().item():+.4f}]")
if x.is_floating_point() or x.is_complex():
info += f" mean={x.float().mean().item():+.4f}"
else:
info += f" values={x.detach().cpu().tolist()[:8]}"
if x.requires_grad or x.grad_fn is not None:
info += " grad_fn=" + (type(x.grad_fn).__name__ if x.grad_fn else "(leaf,req)")
else:
info = f"{type(x).__name__} = {x}"
print(f"{sp}{name:30s} {info}")
if more:
print(f"{sp}{'':30s} ↳ {more}")
def explain(text, indent=4):
sp = " " * indent
wrapped = textwrap.fill(text, width=92 - indent,
initial_indent=sp, subsequent_indent=sp)
print(f"\033[2m{wrapped}\033[0m")
# ============================================================================
# Cross-attention introspection
# ============================================================================
def expose_cross_attention(predictor, src_tokens, tgt_tokens, name="f_AV"):
"""Run the predictor with a forward hook that captures the FIRST
decoder layer's cross-attention input/output. Prints Q/K/V shapes
and attention shape so the figure can show 'cross-attention' explicitly.
"""
captured = {}
layer0 = predictor.decoder.layers[0]
def hook(module, args, kwargs, output):
# nn.TransformerDecoderLayer's multihead_attn is the cross-attention.
# Its forward signature is (query, key, value, ...).
# We don't directly intercept multihead_attn here; instead, we
# capture the layer-level inputs.
captured["tgt"] = args[0] # query input to the cross-attn
captured["memory"] = args[1] # key/value memory (audio for f_AV)
if isinstance(output, tuple):
captured["out"] = output[0]
else:
captured["out"] = output
handle = layer0.register_forward_hook(hook, with_kwargs=True)
out = predictor(src_tokens=src_tokens, tgt_query=tgt_tokens)
handle.remove()
print(f" {name} cross-attention (first decoder layer):")
if "tgt" in captured:
tprint(f" query (= tgt projection)", captured["tgt"], indent=4)
tprint(f" key/value memory (= src)", captured["memory"], indent=4)
tprint(f" layer-0 output", captured["out"], indent=4)
n_heads = layer0.multihead_attn.num_heads
head_dim = layer0.multihead_attn.embed_dim // n_heads
print(f" n_heads={n_heads}, head_dim={head_dim}, depth={len(predictor.decoder.layers)} layers")
return out
# ============================================================================
# Sample loader (HDTF-paired, real & fake share audio)
# ============================================================================
def load_paired_inputs(hdtf_root, num, fake_gen,
num_frames=16, frame_stride=2, frame_size=224,
audio_seconds=2.56, audio_sample_rate=16000):
root = Path(hdtf_root)
rv = root / "Real" / f"{num}_Fake_HDTF.mp4"
fv = root / fake_gen / f"{num}_Fake_HDTF_{fake_gen}.mp4"
aw = root / "_audio" / "Real" / f"{num}_Fake_HDTF.wav"
for p in (rv, fv, aw):
if not p.exists():
raise SystemExit(f"missing: {p}")
real_video = load_video_clip(str(rv), num_frames, frame_stride, frame_size)
fake_video = load_video_clip(str(fv), num_frames, frame_stride, frame_size)
audio = load_audio_clip(str(aw), audio_seconds, audio_sample_rate)
return rv, fv, aw, real_video, fake_video, audio
# ============================================================================
# Helpers for backward-trace
# ============================================================================
def grad_summary(module, label, indent=4):
"""Walk all parameters, count how many have non-zero gradient (after
backward of one specific loss). Returns nothing; prints a compact line.
"""
nz, total, l2 = 0, 0, 0.0
for p in module.parameters():
total += 1
if p.grad is not None and p.requires_grad:
g = p.grad
if torch.any(g != 0):
nz += 1
l2 += g.detach().pow(2).sum().item()
sp = " " * indent
rms = (l2 ** 0.5)
print(f"{sp}{label:35s} {nz}/{total} params with non-zero grad "
f"||grad||₂ = {rms:8.4e}")
def zero_all_grads(model):
for p in model.parameters():
if p.grad is not None:
p.grad.zero_()
# ============================================================================
def parse_args():
p = argparse.ArgumentParser()
p.add_argument("--ckpt",
default="outputs/cta_diffusion_combined_20260604_205145/checkpoints/epoch16-valauc1.0000.ckpt")
p.add_argument("--hdtf_root", default="/apdcephfs_gy5/share_303628665/joyewu/HDTF-paird")
p.add_argument("--num", default="033")
p.add_argument("--fake_generator", default="AniPortrait")
return p.parse_args()
def main():
args = parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
banner("CTA TRAINING-STEP TRACE (forward + 5 losses + backward gradient routing)", 0)
print(f" ckpt: {args.ckpt}")
print(f" device: {device}")
# ---- model -----------------------------------------------------------
state = torch.load(args.ckpt, map_location="cpu")
hp = state.get("hyper_parameters", {})
if not hp:
raise SystemExit("ckpt missing hyper_parameters")
method_cfg = OmegaConf.create(hp["method_cfg"])
backbone_cfg = OmegaConf.create(hp["backbone_cfg"])
data_cfg = OmegaConf.create(hp["data_cfg"])
print()
print(f" loss weights: av={method_cfg.loss.av_weight} va={method_cfg.loss.va_weight} "
f"asym={method_cfg.loss.asym_weight} cls={method_cfg.loss.cls_weight} "
f"aux={method_cfg.aux_crossgen.weight} (aux enabled={method_cfg.aux_crossgen.enabled})")
model = build_method(
method_name=method_cfg.name,
method_cfg=method_cfg,
backbone_cfg=backbone_cfg,
data_cfg=data_cfg,
)
sd = state.get("state_dict", state)
model.load_state_dict(sd, strict=False)
model.to(device)
# IMPORTANT: use eval() so dropout is OFF (gives real numerical values
# rather than NaNs). Gradients still flow normally — eval() only changes
# the behavior of dropout/BatchNorm, not autograd.
model.eval()
# Convenient handles to sub-modules for grad inspection
sub = {
"video_backbone": model.model.video,
"audio_backbone": model.model.audio,
"f_A→V predictor": model.model.av_pred,
"f_V→A predictor": model.model.va_pred,
"classifier head": model.model.cls,
}
# ---- inputs ----------------------------------------------------------
rv_path, fv_path, aw_path, real_video, fake_video, audio = load_paired_inputs(
args.hdtf_root, args.num, args.fake_generator,
num_frames=data_cfg.num_frames,
frame_stride=data_cfg.frame_stride,
frame_size=data_cfg.frame_size,
audio_seconds=data_cfg.audio_seconds,
audio_sample_rate=data_cfg.audio_sample_rate,
)
train_transform = build_video_transform(data_cfg.aug, training=True)
real_norm = train_transform(real_video)
fake_norm = train_transform(fake_video)
banner("INPUT BATCH (1 real + 1 fake)", 1)
print(f" real video: {rv_path}")
print(f" fake video: {fv_path}")
print(f" audio (real, shared by both samples): {aw_path}")
# batch the two samples together — order [real, fake], labels [0, 1]
video_batch = torch.stack([real_norm, fake_norm], dim=0).to(device)
audio_batch = torch.stack([audio, audio ], dim=0).to(device)
labels = torch.tensor([0, 1], dtype=torch.long, device=device)
tprint("video batch (B,T,3,H,W)", video_batch)
tprint("audio batch (B,S)", audio_batch)
tprint("labels (B,)", labels)
# =====================================================================
# FORWARD
# =====================================================================
banner("FORWARD (predict_pairs + cross-attention introspection)", 0)
step(1, "video backbone (VideoMAE)")
v = model.model.video(video_batch)
tprint("v.tokens (B,N,768)", v["tokens"])
tprint("v.pooled (B,768)", v["pooled"])
explain("Token grid: (16/2)·(224/16)² = 8 × 14 × 14 = 1568 spatiotemporal tokens.")
step(2, "audio backbone (Wav2Vec2)")
a = model.model.audio(audio_batch)
tprint("a.tokens (B,T_a,768)", a["tokens"])
tprint("a.pooled (B,768)", a["pooled"])
step(3, "cross-modal predictor f_{A→V} (audio → video manifold)")
explain("Cross-attention: query=projected v.tokens (target side), "
"key/value=projected a.tokens (source memory). Each video token "
"attends to ALL audio tokens to refine its prediction.")
v_pred = expose_cross_attention(model.model.av_pred, a["tokens"], v["tokens"], name="f_AV")
tprint("v_pred (B,N,768)", v_pred)
step(4, "cross-modal predictor f_{V→A} (video → audio manifold)")
a_pred = expose_cross_attention(model.model.va_pred, v["tokens"], a["tokens"], name="f_VA")
tprint("a_pred (B,T_a,768)", a_pred)
step(5, "per-sample MSE residuals L_AV and L_VA")
explain("Computed with reduction='none' then averaged over (token, channel), "
"so each sample gets its own scalar.")
l_av = F.mse_loss(v_pred, v["tokens"], reduction="none").mean(dim=[1, 2])
l_va = F.mse_loss(a_pred, a["tokens"], reduction="none").mean(dim=[1, 2])
asym = l_va - l_av
print(f" L_AV per sample: {l_av.detach().tolist()}")
print(f" L_VA per sample: {l_va.detach().tolist()}")
print(f" asym per sample: {asym.detach().tolist()}")
print(f" ↳ index 0 = real, index 1 = fake")
# =====================================================================
# FIVE LOSSES
# =====================================================================
banner("FIVE LOSSES (mirror of CTALitModule.training_step)", 0)
is_real = (labels == 0).float() # [1, 0]
denom_r = is_real.sum().clamp(min=1.0)
print(f" is_real mask = {is_real.detach().tolist()} "
f"(1 means 'count this sample', 0 means 'mask out')")
# ---- loss_av / loss_va: REAL only, drives predictor + backbones ------
step(1, "loss_av = mean(L_AV | label==0) — REAL only")
explain("fake's contribution is multiplied by 0 ⇒ no gradient through "
"the A→V predictor for fake samples.")
loss_av = (l_av * is_real).sum() / denom_r
tprint("loss_av (scalar)", loss_av)
step(2, "loss_va = mean(L_VA | label==0) — REAL only")
loss_va = (l_va * is_real).sum() / denom_r
tprint("loss_va (scalar)", loss_va)
# ---- loss_asym: BOTH, margin pushes fake's asym below real's --------
step(3, "loss_asym = ReLU(asym_fake_mean − asym_real_mean) — BOTH (margin)")
explain("If fake's asym is already lower than real's (the desired ranking), "
"ReLU(·) = 0 ⇒ no gradient. Else it pushes them apart. "
"Weak signal (weight 0.5) so predictors don't collapse.")
asym_r = asym[labels == 0]
asym_f = asym[labels == 1]
loss_asym = F.relu(asym_f.mean() - asym_r.mean())
tprint("loss_asym (scalar)", loss_asym,
more=f"asym_r mean={asym_r.mean().item():+.4f}, asym_f mean={asym_f.mean().item():+.4f}")
# ---- loss_cls: BOTH, drives classifier + backbones (predictor detached) -----
step(4, "loss_cls = BCE(score, label) — BOTH; L_AV/L_VA detach()ed")
explain("`l_av.detach()` and `l_va.detach()` cut the gradient path "
"from BCE back into the predictors. The classifier head can use "
"the asym scalars as features but cannot rewrite them.")
logits = model.model.classify(
v["pooled"], a["pooled"],
l_av.detach(), # ← detach()
l_va.detach(), # ← detach()
)
print(f" logits per sample: {logits.detach().squeeze(-1).tolist()}")
print(f" scores per sample: {torch.sigmoid(logits.squeeze(-1)).detach().tolist()}")
loss_cls = F.binary_cross_entropy_with_logits(logits.squeeze(-1), labels.float())
tprint("loss_cls (scalar)", loss_cls)
# ---- loss_aux: only computable if alt_video exists in batch ----------
step(5, "loss_aux = MSE(asym, asym_alt_generator) — paired FAKEs only")
explain("This batch has no alt_video / alt_audio (HDTF-paired isn't dual-generator "
"paired). In real training a fake clip can be paired with the SAME identity's "
"fake from a different generator; loss_aux pushes their s_asym to be close, "
"yielding a generator-invariant signal. Here loss_aux = 0.")
loss_aux = asym.new_zeros([])
tprint("loss_aux (scalar; 0 because no alt batch)", loss_aux)
# ---- total ----------------------------------------------------------
step(6, "total = Σ wᵢ · lossᵢ")
total = (
method_cfg.loss.av_weight * loss_av +
method_cfg.loss.va_weight * loss_va +
method_cfg.loss.asym_weight * loss_asym +
method_cfg.loss.cls_weight * loss_cls +
method_cfg.aux_crossgen.weight * loss_aux
)
tprint("total loss (scalar)", total,
more=f"= {method_cfg.loss.av_weight}·loss_av + {method_cfg.loss.va_weight}·loss_va "
f"+ {method_cfg.loss.asym_weight}·loss_asym + {method_cfg.loss.cls_weight}·loss_cls "
f"+ {method_cfg.aux_crossgen.weight}·loss_aux")
# =====================================================================
# BACKWARD — one loss at a time, see who actually receives gradients
# =====================================================================
banner("BACKWARD GRADIENT ROUTING (one loss at a time, count non-zero gradients per submodule)", 0)
explain("For each individual loss, we zero all gradients, call .backward(retain_graph=True), "
"and report how many parameters in each submodule have non-zero grads. "
"This makes 'who is trained by what' explicit.")
individual_losses = [
("loss_av", loss_av, "real-only mse on A→V"),
("loss_va", loss_va, "real-only mse on V→A"),
("loss_asym", loss_asym, "margin between asym_fake and asym_real"),
("loss_cls", loss_cls, "BCE on classifier (detach() on l_av/l_va)"),
]
for nm, l, desc in individual_losses:
zero_all_grads(model)
if l.requires_grad and l.grad_fn is not None:
try:
l.backward(retain_graph=True)
except RuntimeError as e:
print(f"\n [{nm}] backward FAILED: {e}")
continue
print(f"\n [{nm:9s}] {desc}")
for label, m in sub.items():
grad_summary(m, label)
# =====================================================================
# FINAL SUMMARY TABLE
# =====================================================================
banner("F I N A L T A B L E (this is what you'd label on the framework figure)", 0)
print()
print(f" {'sample':<6s} {'L_AV':>10s} {'L_VA':>10s} {'asym':>10s} {'logit':>10s} {'score':>8s} {'label':>6s}")
print(f" {'──────':<6s} {'──────────':>10s} {'──────────':>10s} {'──────────':>10s} {'──────────':>10s} {'──────':>8s} {'──────':>6s}")
for i, name in enumerate(["real", "fake"]):
gt = "0 (real)" if i == 0 else "1 (fake)"
sc = torch.sigmoid(logits[i].squeeze()).item()
print(f" {name:<6s} {l_av[i].item():>10.5f} {l_va[i].item():>10.5f} "
f"{asym[i].item():>+10.5f} {logits[i].squeeze().item():>+10.4f} {sc:>8.4f} {gt:>6s}")
print()
print(f" loss decomposition this batch:")
print(f" loss_av = {loss_av.item():.6f}")
print(f" loss_va = {loss_va.item():.6f}")
print(f" loss_asym = {loss_asym.item():.6f}")
print(f" loss_cls = {loss_cls.item():.6f}")
print(f" loss_aux = {loss_aux.item():.6f}")
print(f" ─────────────────────")
print(f" total = {total.item():.6f}")
print()
if __name__ == "__main__":
main()
|