fairtalking-second-work / scripts /analysis /trace_cta_forward.py
huahua123313's picture
Add files using upload-large-folder tool
eea47ad verified
Raw
History Blame Contribute Delete
14.6 kB
"""Step-by-step forward trace of CTA — shows every tensor shape and key stat
as data flows from raw mp4/wav to the final fake-probability.
This is a DEBUG / DOCUMENTATION script:
* NOT meant to be run as part of training/evaluation
* Purpose: when you draw the framework figure, run this once and you'll
see exactly what each box should contain (shape, dtype, range, semantic)
* Picks two paired samples (1 real + 1 fake, same identity from HDTF-paird
so they share driving audio) and runs CTA through them, printing every step
Usage:
/opt/conda/envs/pytorch/bin/python3 scripts/analysis/trace_cta_forward.py \\
--ckpt outputs/cta_diffusion_combined_20260604_205145/checkpoints/epoch16-valauc1.0000.ckpt
Optional:
--hdtf_root /path/to/HDTF-paird # default: gy5 path
--num 023 # which basename_num to pick
--fake_generator AniPortrait # which fake gen to pair with the real
Output:
* console: structured step-by-step printout (recommend `... | tee trace.log`)
* NO files written
"""
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.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 = "─" * 90
DHRULE = "═" * 90
def banner(title: str, level: int = 0):
bar = DHRULE if level == 0 else HRULE
print()
print(bar)
print(f" {title}")
print(bar)
def tprint(name: str, x, more: str = "", indent: int = 2):
"""Print a tensor's key stats: shape, dtype, range, mean, std."""
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}] "
f"mean={x.mean().item():+.4f} std={x.std().item():+.4f}")
elif isinstance(x, np.ndarray):
info = (f"shape={x.shape} dtype={x.dtype} "
f"range=[{x.min():+.4f}, {x.max():+.4f}]")
elif isinstance(x, (int, float)):
info = f"value={x}"
else:
info = f"{type(x).__name__}"
head = f"{sp}{name:30s}"
print(f"{head} {info}")
if more:
print(f"{sp}{'':30s}{more}")
def step(n: int, title: str):
print(f"\n ── Step {n}: {title} ".ljust(92, "─"))
def explain(text: str, indent: int = 4):
sp = " " * indent
wrapped = textwrap.fill(text, width=90 - indent,
initial_indent=sp, subsequent_indent=sp)
print(f"\033[2m{wrapped}\033[0m") # dim text
# ============================================================================
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",
help="basename_num under HDTF-paird Real/<num>_Fake_HDTF.mp4")
p.add_argument("--fake_generator", default="AniPortrait",
help="which generator to pair with the real video (must "
"exist under <hdtf_root>/<gen>/)")
return p.parse_args()
def load_paired_inputs(hdtf_root: str, num: str, fake_gen: str,
num_frames: int = 16, frame_stride: int = 2,
frame_size: int = 224,
audio_seconds: float = 2.56,
audio_sample_rate: int = 16000):
"""Load (real_video, real_audio) and (fake_video, real_audio) tensors.
HDTF-paired: real and fake share the SAME driving audio (key for CTA's
'audio is real' story)."""
root = Path(hdtf_root)
real_vid = root / "Real" / f"{num}_Fake_HDTF.mp4"
fake_vid = root / fake_gen / f"{num}_Fake_HDTF_{fake_gen}.mp4"
audio_w = root / "_audio" / "Real" / f"{num}_Fake_HDTF.wav"
if not real_vid.exists():
raise SystemExit(f"missing real video: {real_vid}")
if not fake_vid.exists():
raise SystemExit(f"missing fake video: {fake_vid}")
if not audio_w.exists():
raise SystemExit(f"missing audio: {audio_w}")
rv = load_video_clip(str(real_vid), num_frames, frame_stride, frame_size)
fv = load_video_clip(str(fake_vid), num_frames, frame_stride, frame_size)
a = load_audio_clip(str(audio_w), audio_seconds, audio_sample_rate)
return (real_vid, rv), (fake_vid, fv), (audio_w, a)
def trace_one(label: str, video_tensor, audio_tensor, model, eval_transform, device):
"""Run CTA forward on a single sample and print every step."""
banner(f" T R A C I N G '{label.upper()}' S A M P L E", level=0)
# ---- Step 1: raw input ------------------------------------------------
step(1, "RAW INPUT (load_video_clip / load_audio_clip)")
explain("Video comes out of decord/pyav as a (T, 3, H, W) float tensor in [0, 1]. "
"Audio is a 1-D waveform at 16 kHz, 2.56 s long → 40960 samples.")
tprint("video_raw (T,3,H,W) [0,1]", video_tensor)
tprint("audio_raw (S,) float", audio_tensor)
# ---- Step 2: eval transform (normalize) --------------------------------
step(2, "EVAL TRANSFORM (ImageNet-mean/std normalization, no augmentation)")
explain("VideoTransform with training=False only normalizes; it does NOT crop or jitter. "
"Output is (T, 3, H, W) with ~zero mean per channel.")
video_norm = eval_transform(video_tensor)
tprint("video_norm (T,3,H,W)", video_norm)
# batch axis
video_b = video_norm.unsqueeze(0).to(device)
audio_b = audio_tensor.unsqueeze(0).to(device)
# ---- Step 3: VideoMAE backbone ----------------------------------------
step(3, "VIDEO BACKBONE (VideoMAE-base, frozen 70%)")
explain("VideoMAE patchifies the 16-frame clip with patch=16, tubelet=2 → "
"(16/2) × (224/16) × (224/16) = 8 × 14 × 14 = 1568 tokens, each 768-d. "
"Output: pooled (mean over tokens) and full token sequence.")
with torch.no_grad():
v = model.model.video(video_b)
tprint("v.tokens (B,N,768)", v["tokens"])
tprint("v.pooled (B,768)", v["pooled"])
# ---- Step 4: Wav2Vec2 backbone ----------------------------------------
step(4, "AUDIO BACKBONE (Wav2Vec2-base, frozen 80%)")
explain("Wav2Vec2 outputs ~50 Hz tokens. 2.56 s → ~127 frames at 50 Hz → "
"wav2vec2 hidden seq ~80 (subsampled). Each token is 768-d.")
with torch.no_grad():
a = model.model.audio(audio_b)
tprint("a.tokens (B,T_a,768)", a["tokens"])
tprint("a.pooled (B,768)", a["pooled"])
# ---- Step 5: A→V predictor --------------------------------------------
step(5, "CROSS-MODAL PREDICTOR f_{A→V} (Transformer Decoder × 4)")
explain("src = a.tokens (audio); tgt_query = v.tokens (video). Predicts "
"video token embeddings from audio. Teacher-forcing target = v.tokens. "
"Output v_pred has SAME shape as v.tokens.")
with torch.no_grad():
v_pred = model.model.av_pred(src_tokens=a["tokens"], tgt_query=v["tokens"])
tprint("v_pred (B,N,768)", v_pred)
L_AV = F.mse_loss(v_pred, v["tokens"], reduction="none").mean(dim=[1, 2])
tprint("L_AV (B,)", L_AV,
more="= mean of (v_pred - v.tokens)^2 over tokens & channels")
# ---- Step 6: V→A predictor --------------------------------------------
step(6, "CROSS-MODAL PREDICTOR f_{V→A} (Transformer Decoder × 4)")
explain("src = v.tokens; tgt_query = a.tokens. Predicts audio tokens from video. "
"Output a_pred has SAME shape as a.tokens.")
with torch.no_grad():
a_pred = model.model.va_pred(src_tokens=v["tokens"], tgt_query=a["tokens"])
tprint("a_pred (B,T_a,768)", a_pred)
L_VA = F.mse_loss(a_pred, a["tokens"], reduction="none").mean(dim=[1, 2])
tprint("L_VA (B,)", L_VA,
more="= mean of (a_pred - a.tokens)^2 over tokens & channels")
# ---- Step 7: asymmetry score ------------------------------------------
step(7, "ASYMMETRY SCORE")
s_asym = L_VA - L_AV
L_total = L_VA + L_AV
tprint("s_asym = L_VA - L_AV", s_asym,
more="positive ⇒ A→V easier (V→A harder); near-zero or negative ⇒ flat OOD")
tprint("L_total = L_VA + L_AV", L_total,
more="total prediction difficulty; helps classifier disambiguate easy/hard clips")
# ---- Step 8: classifier head ------------------------------------------
step(8, "CLASSIFIER HEAD (MLP, 1538 → 256 → 1)")
explain("Input is the concatenation of [v.pooled, a.pooled, s_asym, L_total] = "
"(B, 768+768+1+1) = (B, 1538). The two scalars are detach()ed during "
"training so BCE gradient cannot leak back into the predictors.")
with torch.no_grad():
feat = torch.cat([
v["pooled"], a["pooled"],
s_asym.unsqueeze(-1), L_total.unsqueeze(-1),
], dim=-1)
tprint("classifier input (B,1538)", feat)
with torch.no_grad():
logit = model.model.cls(feat).squeeze(-1)
tprint("logit (B,)", logit)
score = torch.sigmoid(logit)
tprint("score = sigmoid(logit) ∈ [0,1]", score,
more="high score ⇒ fake")
# ---- Step 9: summary --------------------------------------------------
step(9, "SUMMARY (one-liner you'd put in a figure caption)")
print(f" [{label}] score = {score.item():.4f} "
f"| s_asym = {s_asym.item():+.4f} "
f"| L_AV = {L_AV.item():.4f} | L_VA = {L_VA.item():.4f}")
return {
"label": label,
"L_AV": float(L_AV.item()),
"L_VA": float(L_VA.item()),
"s_asym": float(s_asym.item()),
"L_total": float(L_total.item()),
"score": float(score.item()),
}
# ============================================================================
def main():
args = parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
banner("CTA forward-pass trace", level=0)
print(f" ckpt : {args.ckpt}")
print(f" HDTF : {args.hdtf_root}")
print(f" num : {args.num}")
print(f" fake : {args.fake_generator} (paired with real, SAME audio)")
print(f" device: {device}")
# ---- model ------------------------------------------------------------
state = torch.load(args.ckpt, map_location="cpu")
hp = state.get("hyper_parameters", {})
if not hp:
raise SystemExit("ckpt has no hyper_parameters; cannot rebuild model")
method_cfg = OmegaConf.create(hp["method_cfg"])
backbone_cfg = OmegaConf.create(hp["backbone_cfg"])
data_cfg = OmegaConf.create(hp["data_cfg"])
print()
print(f" method.name = {method_cfg.name}")
print(f" backbone.hf_id = {backbone_cfg.hf_id}")
print(f" av_predictor.depth = {method_cfg.av_predictor.depth}")
print(f" av_predictor.heads = {method_cfg.av_predictor.heads}")
print(f" av_predictor.dropout = {method_cfg.av_predictor.dropout}")
print(f" classifier.hidden = {method_cfg.classifier.hidden}")
print(f" classifier.dropout = {method_cfg.classifier.dropout}")
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)
missing, unexpected = model.load_state_dict(sd, strict=False)
print(f" state_dict load: missing={len(missing)} unexpected={len(unexpected)}")
model.to(device).eval()
eval_transform = build_video_transform(data_cfg.aug, training=False)
# ---- load paired (real, fake) ----------------------------------------
(real_path, real_video), (fake_path, fake_video), (audio_path, 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,
)
banner("INPUT FILES", level=1)
print(f" real video : {real_path}")
print(f" fake video : {fake_path}")
print(f" audio (real, shared by both):")
print(f" {audio_path}")
# ---- trace BOTH samples ----------------------------------------------
real_summary = trace_one("real", real_video, audio, model, eval_transform, device)
fake_summary = trace_one("fake", fake_video, audio, model, eval_transform, device)
# ---- final comparison table ------------------------------------------
banner("F I N A L C O M P A R I S O N", level=0)
print()
print(f" {'sample':<6s} {'L_AV':>10s} {'L_VA':>10s} "
f"{'s_asym':>10s} {'L_total':>10s} {'score':>8s} {'label':>6s}")
print(f" {'─'*6:<6s} {'─'*10:>10s} {'─'*10:>10s} "
f"{'─'*10:>10s} {'─'*10:>10s} {'─'*8:>8s} {'─'*6:>6s}")
for s, gt in [(real_summary, "0 (real)"), (fake_summary, "1 (fake)")]:
print(f" {s['label']:<6s} {s['L_AV']:>10.5f} {s['L_VA']:>10.5f} "
f"{s['s_asym']:>+10.5f} {s['L_total']:>10.5f} {s['score']:>8.4f} {gt:>6s}")
print()
diff_lav = fake_summary["L_AV"] / max(real_summary["L_AV"], 1e-12)
diff_lva = fake_summary["L_VA"] / max(real_summary["L_VA"], 1e-12)
print(f" fake/real ratio: L_AV ×{diff_lav:6.2f} L_VA ×{diff_lva:6.2f}")
print(f" → fake video is OFF the predictor's learned real-manifold; "
f"asymmetry score collapses toward 0 / goes negative.")
print()
if __name__ == "__main__":
main()