File size: 10,465 Bytes
a3206bd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Emit the two ADDITIONAL submissions on the eval set, reusing one feature pass:
  architecture_2 = FG gate    : unanimous-3 with the harmonic voter swapped for the
                                FOREGROUND-harmonic feature (dev BA_unseen 0.365, tau=0.0).
  architecture_3 = 2-voter gate: override base where harmonic & birdmae agree (dev 0.3411, tau=0.4).

Both reuse the deployed bundle's base/harmonic/birdmae/bgwhiten members; FG adds the saved
5-seed fg-harmonic arm (data/perch/harmonicfg_arm.pt). Run with the sibling TF venv python:
  .../.venv/bin/python final/predict_eval_extra.py [--self-test]
"""
import os, sys, json, argparse
import numpy as np
import torch, torch.nn as nn, torch.nn.functional as F
import librosa

HERE = os.path.dirname(os.path.abspath(__file__))
ROOT = os.path.dirname(HERE)
sys.path.insert(0, HERE)
sys.path.insert(0, ROOT)

from framework.metadata import SPECIES_NAMES, SPECIES_ID_TO_NAME, DOMAIN_NAMES  # noqa: E402
import harmonic_features as HF      # noqa: E402  (load_wav, harmonic_feature, HARM_DIM)
import bgwhiten_features as BG      # noqa: E402  (bgwhiten_feature, BGW_DIM)
from predict_eval import read_ids, load_parquet_emb, IDX_TO_SPECIES_ID, EVAL_PARQUET_DIR, DEV_BIRDMAE_PARQUET  # noqa: E402

FG_ARM_PT = os.path.join(ROOT, "data/perch/harmonicfg_arm.pt")
# val-selected gate taus that reproduce the published dev BA_unseen numbers
TAU_GATE3, TAU_FG, TAU_2VOTER = 0.3, 0.0, 0.4


# --- foreground-harmonic feature: deployed harmonic feature but on the loudest-50% frames ---
def fg_harmonic_feature(y):
    Q_LO, Q_HI, N_FFT, HOP, SR = HF.Q_LO, HF.Q_HI, HF.N_FFT, HF.HOP, HF.SR
    S = np.abs(librosa.stft(y, n_fft=N_FFT, hop_length=HOP)) + 1e-8
    logS = np.log(S); e = S.sum(0); T = S.shape[1]
    fg = np.argsort(e)[::-1][:max(3, int(0.50 * T))]
    meanlog = logS[:, fg].mean(1)                       # FOREGROUND mean spectrum (vs whole-clip)
    cep = np.fft.irfft(meanlog, n=N_FFT)
    liftered = cep[Q_LO:Q_HI].copy()
    try:
        f0 = librosa.yin(y, fmin=150, fmax=1200, sr=SR, frame_length=N_FFT); f0 = f0[np.isfinite(f0)]
    except Exception:
        f0 = np.array([])
    if len(f0):
        f0med, f0std, voiced = float(np.median(f0)), float(np.std(f0)), len(f0) / max(1, len(y) // HOP)
    else:
        f0med = f0std = voiced = 0.0
    env = np.fft.rfft(np.concatenate([cep[:Q_LO], np.zeros(N_FFT - Q_LO)]))[:len(meanlog)].real
    white = meanlog - env
    freqs = np.fft.rfftfreq(N_FFT, 1 / SR); hr = []
    if f0med > 0:
        for k in range(1, 7):
            fk = k * f0med
            if fk < freqs[-1]: hr.append(white[np.argmin(np.abs(freqs - fk))])
    hr = np.array(hr + [0.0] * (6 - len(hr)))
    f0harm = np.array([f0med / 600.0, f0std / 200.0, voiced] + list(hr / (np.abs(hr).max() + 1e-6)), np.float32)
    return np.concatenate([liftered.astype(np.float32), f0harm]).astype(np.float32)


# --- HarmNet identical to analysis/add_harmonicfg.py (so the saved state_dicts load) ---
class HarmNet(nn.Module):
    def __init__(s, d):
        super().__init__()
        s.net = nn.Sequential(nn.LayerNorm(d, eps=1e-6), nn.Linear(d, 128), nn.GELU(),
                              nn.Dropout(0.2), nn.Linear(128, 64), nn.LayerNorm(64, eps=1e-6))
        s.sp = nn.Linear(64, 9); s.dm = nn.Linear(64, 5)

    def forward(s, x):
        e = s.net(x); return s.sp(e), s.dm(e), e


def _feat_one(args):
    audio_root, fid = args
    try:
        y = HF.load_wav(os.path.join(audio_root, f"{fid}.wav"))
        return HF.harmonic_feature(y), BG.bgwhiten_feature(y), fg_harmonic_feature(y), True
    except Exception:
        return (np.zeros(HF.HARM_DIM, np.float32), np.zeros(BG.BGW_DIM, np.float32),
                np.zeros(HF.HARM_DIM, np.float32), False)


def compute_features(ids, audio_root, workers):
    harm = np.zeros((len(ids), HF.HARM_DIM), np.float32)
    bgw = np.zeros((len(ids), BG.BGW_DIM), np.float32)
    fg = np.zeros((len(ids), HF.HARM_DIM), np.float32)
    failed = 0
    work = [(audio_root, f) for f in ids]
    from concurrent.futures import ProcessPoolExecutor
    with ProcessPoolExecutor(max_workers=workers) as ex:
        for j, (h, b, g, ok) in enumerate(ex.map(_feat_one, work, chunksize=32)):
            harm[j], bgw[j], fg[j] = h, b, g
            if not ok: failed += 1
            if (j + 1) % 2000 == 0: print(f"  features {j+1}/{len(ids)}", flush=True)
    return harm, bgw, fg, failed


def fg_arm_probs(fg_feat, device):
    """Mean softmax over the 5 saved fg-harmonic HarmNet seeds."""
    arm = torch.load(FG_ARM_PT, map_location=device, weights_only=False)
    mean, std = arm["means_stds"][0]
    X = torch.tensor((fg_feat - mean) / std, dtype=torch.float32, device=device)
    probs = []
    for entry in arm["states"]:
        state = entry[1]                       # (seed, state_dict, vba, fba)
        mo = HarmNet(arm["feature_dim"]).to(device); mo.load_state_dict(state); mo.eval()
        with torch.no_grad():
            probs.append(F.softmax(mo(X)[0], 1).cpu().numpy())
    return np.mean(probs, 0)


def unaniN(base, voters, tau):
    A = np.stack([v.argmax(1) for v in voters], 1)
    M = np.stack([v.max(1) for v in voters], 1)
    fire = np.all(A == A[:, :1], 1) & (M.min(1) > tau)
    out = base.copy(); out[fire] = np.mean([v[fire] for v in voters], 0)
    return out, int(fire.sum())


def _ba_unseen(pred, y, unseen):
    yy = y[unseen]; pp = pred[unseen]
    r = [(pp[yy == c] == c).mean() for c in range(9) if (yy == c).any()]
    return float(np.mean(r))


def write_submission(ids, probs, out_path, title):
    idx = probs.argmax(1)
    sid = np.array([IDX_TO_SPECIES_ID[i] for i in idx], dtype=int)
    os.makedirs(os.path.dirname(out_path), exist_ok=True)
    with open(out_path, "w") as fh:
        fh.write("file_id,predicted_species_id\n")
        for f, s in zip(ids, sid):
            fh.write(f"{f},{s}\n")
    print(f"\n{title}: wrote {len(ids)} rows -> {out_path}")
    for s in range(1, 10):
        c = int((sid == s).sum())
        print(f"  {s:>2} {SPECIES_ID_TO_NAME[str(s)]:<26} {c:>6}  ({100*c/len(ids):.1f}%)")


def self_test(device):
    from ensemble_model import GatedEnsemble
    P = os.path.join(ROOT, "data/perch")
    d = np.load(f"{P}/test.npz", allow_pickle=True)
    perch = d["emb"].astype(np.float32); yte = d["species"].astype(int); dte = d["domain"].astype(int)
    fte = np.array([str(f) for f in d["file_id"]])

    def by_fid(npz, key):
        h = np.load(npz, allow_pickle=True); ix = {str(f): i for i, f in enumerate(h["file_id"])}
        return h[key].astype(np.float32)[np.array([ix[f] for f in fte])]

    harm = by_fid(f"{P}/harmonic_test.npz", "harm")
    bgw = by_fid(f"{P}/bgwhiten_test.npz", "bgw")
    bird, _ = load_parquet_emb(DEV_BIRDMAE_PARQUET, list(fte), 1024)
    sm = json.load(open(f"{ROOT}/data/metadata/split_summary.json"))
    ud = {SPECIES_NAMES.index(k): DOMAIN_NAMES.index(v) for k, v in sm["unseen_domain_by_species"].items()}
    unseen = np.array([dte[i] == ud.get(int(yte[i]), -1) for i in range(len(yte))])

    ens = GatedEnsemble(os.path.join(HERE, "bundle.pt"), device=device)
    mp = ens.member_probs(perch, harm, bird, bgw)
    base, harmp, birdp, bgwp = mp["base_perch"], mp["arm_harmonic"], mp["arm_birdmae"], mp["arm_bgwhiten"]

    # fg dev-test probs: recompute from the arm on harmonicfg_test.npz, and check vs the saved `ft`
    fg_feat = by_fid(f"{P}/harmonicfg_test.npz", "harm")
    fgp = fg_arm_probs(fg_feat, device)
    saved_ft = torch.load(FG_ARM_PT, map_location="cpu", weights_only=False)["ft"]
    drift = float(np.abs(fgp - saved_ft).max())
    print(f"[self-test] fg-arm reproduce vs saved ft: max|delta|={drift:.4g}")
    assert drift < 1e-4, "fg-arm inference does not reproduce saved probs"

    g3, _ = unaniN(base, [harmp, birdp, bgwp], TAU_GATE3)
    fg, _ = unaniN(base, [fgp, birdp, bgwp], TAU_FG)
    tv, _ = unaniN(base, [harmp, birdp], TAU_2VOTER)
    for name, pred, exp in [("gate3", g3, 0.3616), ("FG", fg, 0.365), ("2voter", tv, 0.3411)]:
        ba = _ba_unseen(pred.argmax(1), yte, unseen)
        print(f"[self-test] {name:7s} dev BA_unseen={ba:.4f} (expect {exp})")
        assert abs(ba - exp) < 0.001, f"{name} mismatch {ba} != {exp}"
    print("[self-test] OK\n")


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--audio-root", default=os.path.join(ROOT, "eval"))
    ap.add_argument("--ids", default=os.path.join(ROOT, "data/metadata_eval/Test_ids.txt"))
    ap.add_argument("--parquet-dir", default=EVAL_PARQUET_DIR)
    ap.add_argument("--bundle", default=os.path.join(HERE, "bundle.pt"))
    ap.add_argument("--sub-dir", default=os.path.join(ROOT, "final_submission"))
    ap.add_argument("--workers", type=int, default=max(1, (os.cpu_count() or 2) - 1))
    ap.add_argument("--self-test", action="store_true")
    args = ap.parse_args()
    device = "cuda" if torch.cuda.is_available() else "cpu"

    if args.self_test:
        self_test(device)

    ids = read_ids(args.ids)
    print(f"eval clips: {len(ids)}")
    print("loading Perch + BirdMAE eval embeddings ...")
    perch, mp_ = load_parquet_emb(os.path.join(args.parquet_dir, "perch.parquet"), ids, 1536)
    bird, mb_ = load_parquet_emb(os.path.join(args.parquet_dir, "birdmae.parquet"), ids, 1024)
    print(f"  perch missing={mp_}  birdmae missing={mb_}")
    print(f"computing harmonic + bgwhiten + fg-harmonic ({args.workers} workers) ...")
    harm, bgw, fgf, failed = compute_features(ids, args.audio_root, args.workers)
    print(f"  features done; unreadable clips (zero-filled)={failed}")

    print("running member probes ...")
    from ensemble_model import GatedEnsemble
    ens = GatedEnsemble(args.bundle, device=device)
    mp = ens.member_probs(perch, harm, bird, bgw)
    base, harmp, birdp, bgwp = mp["base_perch"], mp["arm_harmonic"], mp["arm_birdmae"], mp["arm_bgwhiten"]
    fgp = fg_arm_probs(fgf, device)

    fg_pred, n_fg = unaniN(base, [fgp, birdp, bgwp], TAU_FG)
    tv_pred, n_tv = unaniN(base, [harmp, birdp], TAU_2VOTER)
    print(f"FG gate fired on {n_fg}/{len(ids)} clips; 2-voter gate fired on {n_tv}/{len(ids)}")

    write_submission(ids, fg_pred, os.path.join(args.sub_dir, "architecture_2/predictions.txt"),
                     "architecture_2 (FG gate, dev 0.365)")
    write_submission(ids, tv_pred, os.path.join(args.sub_dir, "architecture_3/predictions.txt"),
                     "architecture_3 (2-voter gate, dev 0.3411)")


if __name__ == "__main__":
    main()