File size: 8,374 Bytes
6fdd8c7 | 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 | """Validation suite v4 (6.25 Hz, new intensity target).
Checks: 6.25 Hz rate contract, interface contract, teacher fidelity on held-out
RAVDESS actors 19-24, intensity corr vs the NEW salience-gated absolute-energy
target, and — the acceptance gate — separation of RAVDESS ground-truth intensity
labels (normal 01 vs strong 02) over ALL 336 emotional held-out files.
Writes /tmp/work/surgical_out_v4/eval_report_v4.json + dynamics_plot_v4.png
"""
from __future__ import annotations
import glob, json, os, random, sys
import numpy as np
import torch
import torch.nn.functional as F
import torchaudio
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from surgical_emotion import SurgicalEmotionEncoder
FUNASR_DIR = "/tmp/work/emotion2vec_plus_base"
OUT_DIR = "/tmp/work/surgical_out_v4"
CKPT = os.path.join(OUT_DIR, "surgical_emotion_module.pt")
RAVDESS_DIR = "/tmp/work/ravdess_kag"
EMO_IDX = [0, 1, 2, 3, 6, 7]
GATE_FLOOR = 0.02
def rms_raw(wav, T):
L = wav.numel(); n_full = (L + 319) // 320
w = torch.zeros(T * 320); w[: min(L, T * 320)] = wav[: min(L, T * 320)]
ms = w.view(T, 320).pow(2).mean(1).sqrt()
if T > n_full: ms[n_full:] = ms[n_full - 1]
if T >= 9:
ms = F.conv1d(ms[None, None], torch.ones(9)[None, None] / 9.0, padding=4)[0, 0]
return ms
def js_div(p, q, eps=1e-8):
p = p.clamp(min=eps); q = q.clamp(min=eps)
m = 0.5 * (p + q)
kl = lambda a, b: float((a * (a.log() - b.log())).sum(-1))
return 0.5 * kl(p, m) + 0.5 * kl(q, m)
def load_wav(path, max_s=12.0):
wav, sr = torchaudio.load(path)
wav = wav.mean(0)
if sr != 16000:
wav = torchaudio.functional.resample(wav, sr, 16000)
wav = wav[: int(max_s * 16000)]
return wav / wav.abs().max().clamp(min=1e-4)
def new_target(enc, wav, states):
probs = torch.softmax(enc.teacher.proj(states), -1)[0] # (T,9)
p6 = probs[:, EMO_IDX].max(-1).values.clamp(min=GATE_FLOOR).cpu()
env = (rms_raw(wav, states.size(1)) / META["global_ref"]).clamp(0, 1)
return p6 * env, probs
def main():
device = "cuda:0"
enc = SurgicalEmotionEncoder.from_pretrained(FUNASR_DIR, CKPT, device=device)
enc.eval()
report = {"frame_rate_hz": enc.cfg.frame_rate_hz, "checks": {}, "intensity": {}}
# 1. rate & interface contract at 6.25 Hz
ok = True
for sec in [1.0, 2.5, 5.0, 10.0]:
n = int(sec * 16000)
wav = (torch.randn(1, n, device=device)); wav = wav / wav.abs().max().clamp(min=1e-3)
out = enc(wav, torch.tensor([n], device=device))
expect = int(np.ceil(sec * enc.cfg.frame_rate_hz))
got = int(out["frame_embs"].shape[1])
good = abs(got - expect) <= 1
ok &= good
report["checks"][f"{sec}s"] = {"expected~": expect, "got": got, "ok": good}
assert out["frame_embs"].shape[-1] == enc.cfg.out_dim
assert 0.0 <= float(out["frame_intensity"].min()) and float(out["frame_intensity"].max()) <= 1.0
report["rate_contract_ok"] = bool(ok)
print("rate contract ok:", ok)
files = sorted(glob.glob(os.path.join(RAVDESS_DIR, "**", "*.wav"), recursive=True))
speech = [f for f in files if os.path.basename(f).split("-")[1] == "01"
and 19 <= int(os.path.basename(f).split("-")[-1].split(".")[0]) <= 24]
emotional = [f for f in speech if os.path.basename(f).split("-")[2] != "01"]
random.seed(0)
sel = random.sample(speech, min(60, len(speech)))
agree, jss, corrs = [], [], []
plot_sample = None
with torch.no_grad():
for f in sel:
wav = load_wav(f)
if wav.numel() < 8000: continue
out = enc(wav[None].to(device), torch.tensor([wav.numel()], device=device))
states = enc.teacher.extract_features(
F.layer_norm(wav[None].to(device), wav[None].to(device).shape),
padding_mask=None, mask=False)["x"]
tp = torch.softmax(enc.teacher.proj(states.mean(1)), -1)[0]
sp = out["utt_probs"][0]
agree.append(int(sp.argmax().item() == tp.argmax().item()))
jss.append(js_div(sp[None], tp[None]))
tgt50, _ = new_target(enc, wav, states)
si = out["frame_intensity"][0, :, 0].cpu()
tgt = F.interpolate(tgt50[None, None], size=si.size(0), mode="area")[0, 0]
if si.var() > 1e-8 and tgt.var() > 1e-8:
corrs.append(float(np.corrcoef(si.numpy(), tgt.numpy())[0, 1]))
if plot_sample is None and "-05-" in f:
plot_sample = (os.path.basename(f), out, states, tgt50)
report["intensity"]["fidelity_agreement"] = float(np.mean(agree))
report["intensity"]["fidelity_mean_js"] = float(np.mean(jss))
report["intensity"]["mean_corr_new_target"] = float(np.mean(corrs))
print(f"agreement={report['intensity']['fidelity_agreement']:.3f} "
f"JS={report['intensity']['fidelity_mean_js']:.4f} "
f"int_corr={report['intensity']['mean_corr_new_target']:.3f}")
# 2. ground-truth intensity separation over ALL emotional held-out files
groups = {"01": [], "02": []}
groups_tgt = {"01": [], "02": []}
with torch.no_grad():
for f in emotional:
wav = load_wav(f)
if wav.numel() < 8000: continue
out = enc(wav[None].to(device), torch.tensor([wav.numel()], device=device))
states = enc.teacher.extract_features(
F.layer_norm(wav[None].to(device), wav[None].to(device).shape),
padding_mask=None, mask=False)["x"]
tgt50, _ = new_target(enc, wav, states)
lbl = os.path.basename(f).split("-")[3]
groups[lbl].append(float(out["frame_intensity"][0, :, 0].mean()))
groups_tgt[lbl].append(float(tgt50.mean()))
a, b = np.array(groups["01"]), np.array(groups["02"])
d = (b.mean() - a.mean()) / np.sqrt((a.var() + b.var()) / 2 + 1e-9)
report["intensity"]["sep_normal_mean"] = float(a.mean())
report["intensity"]["sep_strong_mean"] = float(b.mean())
report["intensity"]["sep_diff"] = float(b.mean() - a.mean())
report["intensity"]["sep_cohen_d"] = float(d)
print(f"intensity separation: normal {a.mean():.3f} vs strong {b.mean():.3f} "
f"diff {b.mean()-a.mean():+.3f} Cohen-d {d:.2f} (n={len(a)}/{len(b)})")
ta, tb = np.array(groups_tgt["01"]), np.array(groups_tgt["02"])
td = (tb.mean() - ta.mean()) / np.sqrt((ta.var() + tb.var()) / 2 + 1e-9)
report["intensity"]["target_sep_normal_mean"] = float(ta.mean())
report["intensity"]["target_sep_strong_mean"] = float(tb.mean())
report["intensity"]["target_sep_cohen_d"] = float(td)
print(f"target itself : normal {ta.mean():.3f} vs strong {tb.mean():.3f} "
f"Cohen-d {td:.2f} (ceiling diagnostic)")
# 3. dynamics plot
try:
import matplotlib; matplotlib.use("Agg")
import matplotlib.pyplot as plt
name, out, states, tgt50 = plot_sample
fp = out["frame_probs"][0].cpu().numpy()
si = out["frame_intensity"][0, :, 0].cpu()
tgt = F.interpolate(tgt50[None, None], size=si.size(0), mode="area")[0, 0]
t = np.arange(si.size(0)) * 0.16
fig, ax = plt.subplots(2, 1, figsize=(11, 6))
for k, lab in enumerate(["angry", "disgusted", "fearful", "happy",
"neutral", "other", "sad", "surprised"]):
ax[0].plot(t, fp[:, k], label=lab, lw=1.2)
ax[0].set_title(f"v4 frame emotion probs @6.25Hz — {name}")
ax[0].legend(fontsize=7, ncol=4); ax[0].set_ylabel("p")
ax[1].plot(t, si, color="crimson", lw=1.8, label="v4 intensity")
ax[1].plot(t, tgt, color="steelblue", lw=1.4, ls="--", label="target")
ax[1].set_ylabel("intensity"); ax[1].set_ylim(0, 1.05); ax[1].legend(fontsize=8)
ax[1].set_xlabel("time (s) — frame = 160 ms | 6.25 Hz (1 frame / dots.tts token)")
fig.tight_layout()
fig.savefig(os.path.join(OUT_DIR, "dynamics_plot_v4.png"), dpi=110)
print("plot saved")
except Exception as e:
print("plot skipped:", e)
with open(os.path.join(OUT_DIR, "eval_report_v4.json"), "w") as f:
json.dump(report, f, indent=1)
print("EVAL_DONE")
if __name__ == "__main__":
with open(os.path.join(OUT_DIR, "target_meta.json")) as f:
META = json.load(f)
main()
|