File size: 10,197 Bytes
c413e6f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Statistics over all voice profiles, all 40 emotions and all 57 VoiceNet dimensions.

  python build_stats.py --src <RELEASE_ROOT>/index --dst <RELEASE_ROOT>/stats [--workers 32]

Writes three CSVs, each split by `origin`:
  stats_profiles.csv    one row per (voice, origin)     -- counts, hours, duration distribution
  stats_emotions.csv    one row per (emotion, origin)   -- intensity distribution + argmax counts
  stats_dimensions.csv  one row per (vn dim, origin)    -- reg distribution + bucket histogram

Percentiles are computed from fixed-width histograms (4096 bins over a per-column fixed range),
which is exact in mean/count/min/max and accurate to one bin width in the percentiles -- about
0.002 on the 0-7 emotion scale and 0.003 on the VoiceNet reg scale. This is what lets 26 M rows be
summarised in bounded memory in a single streaming pass. The bin width is reported in each file.
"""
import argparse, os, glob, csv, math
import numpy as np, pyarrow.parquet as pq
from concurrent.futures import ProcessPoolExecutor

EMO = ["Affection","Amusement","Anger","Astonishment_Surprise","Awe","Bitterness","Concentration",
 "Confusion","Contemplation","Contempt","Contentment","Disappointment","Disgust","Distress","Doubt",
 "Elation","Embarrassment","Emotional_Numbness","Fatigue_Exhaustion","Fear","Helplessness",
 "Hope_Enthusiasm_Optimism","Impatience_and_Irritability","Infatuation","Interest",
 "Intoxication_Altered_States_of_Consciousness","Longing","Malevolence_Malice","Pain",
 "Pleasure_Ecstasy","Pride","Relief","Sadness","Sexual_Lust","Shame","Sourness","Teasing",
 "Thankfulness_Gratitude","Triumph","Jealousy_and_Envy"]
VN = ["AGEV","AROU","ARSH","ATCK","BKGN","BRGT","CHNK","CLRT","COGL","DARC","DFLU","EMPH","ESTH",
 "EXPL","FOCS","FULL","GEND","HARM","METL","RANG","RCQL","REGS","RESP","ROUG","R_CHST","R_HEAD",
 "R_MASK","R_MIXD","R_NASL","R_ORAL","R_THRT","SMTH","STNC","STRU","S_ASMR","S_AUTH","S_CART",
 "S_CASU","S_CONV","S_DRAM","S_FORM","S_MONO","S_NARR","S_NEWS","S_PLAY","S_RANT","S_STRY",
 "S_TECH","S_WHIS","TEMP","TENS","VALN","VALS","VFLX","VOLT","VULN","WARM"]
NB = 4096
RANGES = {"emo": (0.0, 8.0), "vn": (-6.0, 10.0), "dur": (0.0, 120.0)}

class Hist:
    __slots__ = ("lo","hi","h","n","s","mn","mx")
    def __init__(self, lo, hi):
        self.lo, self.hi = lo, hi
        self.h = np.zeros(NB, np.int64); self.n = 0; self.s = 0.0
        self.mn = math.inf; self.mx = -math.inf
    def add(self, x):
        x = x[np.isfinite(x)]
        if x.size == 0: return
        self.n += x.size; self.s += float(x.sum())
        self.mn = min(self.mn, float(x.min())); self.mx = max(self.mx, float(x.max()))
        i = np.clip(((x - self.lo) / (self.hi - self.lo) * NB).astype(np.int64), 0, NB - 1)
        self.h += np.bincount(i, minlength=NB)
    def merge(self, o):
        self.h += o.h; self.n += o.n; self.s += o.s
        self.mn = min(self.mn, o.mn); self.mx = max(self.mx, o.mx); return self
    def pct(self, qs):
        if self.n == 0: return [None] * len(qs)
        c = np.cumsum(self.h); out = []
        for q in qs:
            k = np.searchsorted(c, q / 100.0 * self.n)
            out.append(self.lo + (min(k, NB - 1) + 0.5) * (self.hi - self.lo) / NB)
        return out
    def row(self):
        p = self.pct([1, 10, 25, 50, 75, 90, 99])
        return dict(count=self.n, mean=(self.s / self.n if self.n else None),
                    min=(self.mn if self.n else None), max=(self.mx if self.n else None),
                    p1=p[0], p10=p[1], p25=p[2], median=p[3], p75=p[4], p90=p[5], p99=p[6])

def blank():
    return dict(emo={e: Hist(*RANGES["emo"]) for e in EMO},
                vn={d: Hist(*RANGES["vn"]) for d in VN},
                bucket={d: np.zeros(8, np.int64) for d in VN},
                top=({e: 0 for e in EMO}),
                voice={})

def scan(f):
    cols = (["voice","origin","dur_s","top_emotion","n_bursts","genuineness_0_6","blend_0_10"]
            + [f"emo_{e}" for e in EMO] + [f"vn_{d}_reg" for d in VN] + [f"vn_{d}_bucket" for d in VN])
    t = pq.read_table(f, columns=[c for c in cols])
    acc = {}
    origins = np.array(t["origin"].to_pylist())
    for org in np.unique(origins):
        m = origins == org
        A = acc.setdefault(str(org), blank())
        for e in EMO:
            A["emo"][e].add(np.asarray(t[f"emo_{e}"].to_numpy(zero_copy_only=False))[m])
        for d in VN:
            A["vn"][d].add(np.asarray(t[f"vn_{d}_reg"].to_numpy(zero_copy_only=False))[m])
            b = np.asarray(t[f"vn_{d}_bucket"].to_numpy(zero_copy_only=False))[m]
            b = b[np.isfinite(b)].astype(np.int64)
            A["bucket"][d] += np.bincount(np.clip(b, 0, 7), minlength=8)
        for te in np.asarray(t["top_emotion"].to_pylist())[m]:
            if te in A["top"]: A["top"][te] += 1
        vs = np.asarray(t["voice"].to_pylist())[m]
        du = np.asarray(t["dur_s"].to_numpy(zero_copy_only=False))[m]
        gn = np.asarray(t["genuineness_0_6"].to_numpy(zero_copy_only=False))[m]
        bl = np.asarray(t["blend_0_10"].to_numpy(zero_copy_only=False))[m]
        nb = np.asarray(t["n_bursts"].to_numpy(zero_copy_only=False))[m]
        for v in np.unique(vs):
            mm = vs == v
            V = A["voice"].setdefault(str(v), dict(d=Hist(*RANGES["dur"]), n=0, sg=0.0, sb=0.0, nb=0))
            V["d"].add(du[mm]); V["n"] += int(mm.sum())
            V["sg"] += float(np.nansum(gn[mm])); V["sb"] += float(np.nansum(bl[mm]))
            V["nb"] += int(np.nansum(nb[mm]))
    return acc

def merge(A, B):
    for org, b in B.items():
        a = A.setdefault(org, blank())
        for e in EMO: a["emo"][e].merge(b["emo"][e])
        for d in VN:
            a["vn"][d].merge(b["vn"][d]); a["bucket"][d] += b["bucket"][d]
        for e, v in b["top"].items(): a["top"][e] += v
        for v, s in b["voice"].items():
            t = a["voice"].setdefault(v, dict(d=Hist(*RANGES["dur"]), n=0, sg=0.0, sb=0.0, nb=0))
            t["d"].merge(s["d"]); t["n"] += s["n"]; t["sg"] += s["sg"]; t["sb"] += s["sb"]; t["nb"] += s["nb"]
    return A

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--src", required=True); ap.add_argument("--dst", required=True)
    ap.add_argument("--workers", type=int, default=32)
    A = ap.parse_args()
    fs = sorted(glob.glob(os.path.join(A.src, "origin=*", "*.parquet")))
    print(f"{len(fs)} parquet files", flush=True)
    acc = {}
    with ProcessPoolExecutor(A.workers) as ex:
        for k, r in enumerate(ex.map(scan, fs, chunksize=2), 1):
            merge(acc, r)
            if k % 500 == 0: print(f"  {k}/{len(fs)}", flush=True)
    os.makedirs(A.dst, exist_ok=True)
    bw_e = (RANGES["emo"][1]-RANGES["emo"][0])/NB; bw_v = (RANGES["vn"][1]-RANGES["vn"][0])/NB
    F = ["count","mean","min","max","p1","p10","p25","median","p75","p90","p99"]

    with open(os.path.join(A.dst, "stats_emotions.csv"), "w", newline="") as fh:
        w = csv.writer(fh); w.writerow(["# percentile bin width", f"{bw_e:.6f}"])
        w.writerow(["emotion","origin"]+F+["n_top_emotion","frac_top_emotion"])
        for org in sorted(acc):
            tot = sum(acc[org]["top"].values()) or 1
            for e in EMO:
                r = acc[org]["emo"][e].row()
                w.writerow([e,org]+[r[k] for k in F]+[acc[org]["top"][e], acc[org]["top"][e]/tot])

    with open(os.path.join(A.dst, "stats_dimensions.csv"), "w", newline="") as fh:
        w = csv.writer(fh); w.writerow(["# percentile bin width", f"{bw_v:.6f}"])
        w.writerow(["dimension","origin"]+F+[f"bucket_{i}" for i in range(8)])
        for org in sorted(acc):
            for d in VN:
                r = acc[org]["vn"][d].row()
                w.writerow([d,org]+[r[k] for k in F]+list(acc[org]["bucket"][d]))

    with open(os.path.join(A.dst, "stats_profiles.csv"), "w", newline="") as fh:
        w = csv.writer(fh)
        w.writerow(["voice","origin","n_utterances","hours","dur_mean","dur_min","dur_max",
                    "dur_p10","dur_median","dur_p90","mean_genuineness","mean_blend","total_bursts"])
        for org in sorted(acc):
            for v in sorted(acc[org]["voice"]):
                s = acc[org]["voice"][v]; r = s["d"].row(); n = s["n"] or 1
                w.writerow([v,org,s["n"], (s["d"].s/3600.0), r["mean"], r["min"], r["max"],
                            r["p10"], r["median"], r["p90"], s["sg"]/n, s["sb"]/n, s["nb"]])
    # ---- shipped normalisation artifact -------------------------------------------------
    # Descriptive statistics of THIS population, recomputed from the re-annotated output.
    # NOTE: the captions do not consume these. `vn_*_bucket` is the argmax of an ordinal
    # classifier trained on absolute prose anchors, so caption wording is population-independent.
    # These are published so that users who WANT population-relative normalisation have a
    # reference that is (a) specific to the voice profiles and (b) free of the half-speed values.
    import json
    norm = dict(
        _about=("Descriptive statistics of the LAION voice-profile population, computed from the "
                "re-annotated (stereo-fixed) output. NOT used by the caption renderer: vn_*_bucket "
                "is an absolute ordinal-classifier argmax, not a percentile of this population. "
                "Provided for users who want population-relative normalisation."),
        _scope="voice profiles only (origin=original + origin=repair); not the wider 8-dataset corpus",
        _bin_width=dict(emotion=bw_e, voicenet_reg=bw_v),
        per_origin={})
    for org in sorted(acc):
        norm["per_origin"][org] = dict(
            emotions={e: acc[org]["emo"][e].row() for e in EMO},
            voicenet={d: dict(acc[org]["vn"][d].row(),
                              bucket_counts=[int(x) for x in acc[org]["bucket"][d]]) for d in VN},
            n_voices=len(acc[org]["voice"]),
            n_utterances=sum(v["n"] for v in acc[org]["voice"].values()))
    with open(os.path.join(A.dst, "norm_stats_vprof.json"), "w") as fh:
        json.dump(norm, fh, indent=1)
    print("wrote 3 CSVs + norm_stats_vprof.json to", A.dst)

if __name__ == "__main__":
    main()