upload legacy/final

legacy/final/README.md

@@ -0,0 +1,174 @@
+# CD-MSC — Best Ensemble (techniques, losses, and how to run)
+
+BioDCASE 2026 Task 5 — Cross-Domain Mosquito Species Classification. Optimize **BA_unseen** = mean over the 9
+species of per-species recall on each species' **held-out unseen recording domain** (true leave-one-domain-out).
+
+This folder is the self-contained final model: every model/loss/feature definition, all member checkpoints,
+the gate manifest, and a one-command inference script that reproduces test BA_unseen.
+
+---
+
+## 1. The problem the techniques are designed around
+
+- 9 species; 5 recording **domains** D1–D5 (devices). **D5 is ~98% of training** (clean studio); D1–D4 are
+  small and shifted. Each species is **tested on a domain it was never trained on** (LODO).
+- The domain shift is a **spectral-envelope EQ tilt**: it re-colors the overall spectrum but **preserves the
+  harmonic comb / wingbeat fundamental f0**. No information is destroyed — it is a *style* shift.
+- Failure mode without DG: the model latches onto the device "style" as a shortcut → collapses on unseen
+  domains (especially the studio→D1 transfer, where the envelope inverts).
+
+Two structural walls cap any model near ~0.40–0.45 (oracle): (a) the **movers** gambiae/arabiensis/quinque are
+100% D5-trained but tested on D1 (no in-domain anchor); (b) **dirus** has 0 D5 clips; (c) validation is 99.4%
+D5, so you cannot model-select for the unseen regime.
+
+---
+
+## 2. Representations (frozen embeddings — no encoder is fine-tuned)
+
+| member | embedding | dim | why |
+|---|---|---|---|
+| `base_perch` | **Perch v2** (Google bird-vocalization model, frozen) | 1536 | strongest general bioacoustic encoder here |
+| `arm_harmonic` | **physics harmonic feature** (`harmonic_features.py`) | 102 | device-INVARIANT by construction (comb/f0) |
+| `arm_birdmae` | **BirdMAE** (frozen) | 1024 | independent FM; agrees with harmonic on Aedes |
+| `arm_bgwhiten` | **background-whitened spectrum** (`bgwhiten_features.py`) | 257 | device channel estimated from the clip's background frames and subtracted; 3rd independent device-invariant voter |
+
+The harmonic feature = liftered cepstrum (drops the low-quefrency EQ envelope, keeps the pitch band) + f0/
+harmonic ratios on the whitened spectrum. Perch's fine-tuning is impossible (it ships as a jax2tf SavedModel
+with gradients disabled), so all members are **frozen embedding + trainable probe**.
+
+---
+
+## 3. Probe architectures (`model.py`)
+
+- **RP** (Perch, BirdMAE): input LayerNorm → Linear(d→256) → 2× Pre-LN residual block (LN→Linear(4×,GELU)→
+  Dropout→Linear→+residual) → LayerNorm → species head (9) + auxiliary domain head (5).
+- **HarmNet** (harmonic): LN → Linear(102→128) → GELU → Dropout → Linear(128→64) → LN → species + domain heads.
+
+---
+
+## 4. Training techniques & losses (`losses.py`)
+
+Per-member objective (every member, same recipe):
+
+```
+L = CE(species)
+  + 0.01 * CE(domain)            # auxiliary domain head (plain aux, not adversarial)
+  + 0.50 * genus_aux             # hierarchical Aedes/Culex/Anopheles consistency
+  + 0.50 * supcon(T=0.10)        # supervised contrastive on the L2-normalized embedding
+  + 1.00 * coral                 # Deep CORAL — align per-domain covariances
+  + 0.25 * mmd                   # multi-kernel Gaussian MMD — align per-domain distributions  <-- THE lever
+```
+
+- **MMD (0.25)** is the single most important knob: aligning the per-domain embedding distributions lifted
+  BA_unseen ~0.25 → ~0.33. The weight 0.25 is the measured optimum (a sweep peak); 0.0→0.245, 0.1→0.31,
+  0.25→0.33, 0.5→0.30, 0.75→0.28.
+- **CORAL** complements MMD (second-order alignment).
+- **genus_aux** keeps cross-genus structure correct (so Aedes/Culex aren't mislabeled Anopheles).
+- **SupCon** (L2-normalized; an earlier L1 bug made it inert) — small positive.
+- Domain head at 0.01 is a *plain* classifier (DANN/gradient-reversal was tried, ~neutral).
+
+### Augmentations (in `model.py`, active only in `.train()`)
+- **Embedding Gaussian noise** σ=0.05 (RP) / 0.03 (HarmNet).
+- **Feature channel-dropout** p=0.1 (RP).
+- *Tested and dropped (within noise / negative):* class-mixup (slightly hurts), domain-mixup (≈neutral),
+  DicL contrastive (hurts), forward EQ-augmentation (neutral), MixStyle, genus-gated MoE (hurts).
+
+### Optimization / selection
+- AdamW, lr 3e-4, cosine schedule + 5% warmup, weight decay 1e-4, batch 512.
+- Base: **30 epochs**; arms: 20 epochs. **Uniform** sampling.
+- Checkpoint selection on **seen-val species balanced accuracy** (never the test set).
+- Per member: ensemble over seeds by **softmax averaging** (base 8 seeds, each arm 3 seeds).
+
+---
+
+## 5. The ensemble + the UNANIMOUS-3 AGREEMENT GATE (the headline technique)
+
+Each member is a seed-ensemble. The final prediction:
+
+```
+pred = base_perch_ensemble
+EXCEPT on clips where  harmonic.argmax == birdmae.argmax == bgwhiten.argmax == c  AND  min(all 3 confs) > tau:
+       there, override with mean(harmonic, birdmae, bgwhiten)
+```
+
+- **Why it works:** the studio→D1 EQ shift inverts Perch's envelope (Perch lands *below chance* on D5→D1),
+  while the device-invariant reps survive. Both harmonic and BirdMAE independently nail **aegypti** (Aedes,
+  distinct wingbeat: harmonic 0.90, BirdMAE 0.93 vs Perch 0.43), so the agreement captures aegypti (0.43→~0.81).
+- **The 3rd voter is a precision VETO (this is what raised the score).** A *two*-voter gate (harmonic+birdmae)
+  over-fires: it overrides true **mover** clips (arabiensis/quinque) toward aegypti — the metric-aware cost.
+  Adding the **background-whitened** spectrum as a *third independent device-invariant voter*, and requiring
+  **unanimous** agreement, makes the gate fire only when it is really safe. It keeps aegypti high **and stops
+  sacrificing the movers**: arabiensis 0.148→0.192, quinque 0.140→0.196. The bgwhiten voter need not predict
+  the movers itself — it simply *vetoes the bad overrides*. Net **+0.02 over the 2-voter gate**, and the gain
+  is on the LARGE-n unseen partitions (arabiensis 1820, quinque 672 clips), LOSO-stable across seeds.
+- `tau` is the only knob, selected on seen-val → deployable, no test-set tuning.
+- **Honest caveat:** it still cannot *separate* the *An. gambiae* complex (gambiae/arabiensis share wingbeat
+  f0); gambiae stays floored. The win is recovering the movers the 2-voter gate was throwing away.
+
+The exact numbers (base, 2-voter, unanimous-3 BA_unseen, tau, per-species) are in `ensemble.json` and printed
+by `infer.py`. The previous 2-voter package is preserved in `checkpoints_backup_2voter/`.
+
+---
+
+## 6. Files
+
+```
+final/
+  README.md              <- this file
+  model.py               <- RP, HarmNet, ResidualBlock
+  losses.py              <- CE/genus/SupCon/MMD/CORAL + total_loss (the recipe)
+  harmonic_features.py   <- raw audio -> 102-d device-invariant harmonic feature
+  bgwhiten_features.py   <- raw audio -> 257-d background-whitened spectrum (3rd voter)
+  infer.py               <- load checkpoints + cached embeddings -> unanimous-3 gated BA_unseen
+  ensemble_model.py      <- GatedEnsemble: one object loads bundle.pt and applies the 3-voter gate
+  bundle_ensemble.py     <- packs all 19 checkpoints + gate tau into bundle.pt
+  ensemble.json          <- member list, gate tau, results
+  checkpoints/
+    base_perch_s{1234,42,2023,7,99,123,256,777}.pt   (8)
+    arm_harmonic_s{1234,42,2023}.pt                   (3)
+    arm_birdmae_s{1234,42,2023}.pt                    (3)
+    arm_bgwhiten_s{1234,42,2023,7,99}.pt              (5)
+  checkpoints_backup_2voter/   <- the previous 2-voter package (revertable)
+```
+
+Each `.pt` holds: `model_state_dict`, standardization `mean`/`std`, `modality`, `arch`, `feature_dim`,
+`seed`, `val_ba`, `test_ba_unseen`.
+
+## 7. Run
+
+```bash
+# reproduces test BA_unseen from the per-member checkpoints + cached embeddings
+python final/infer.py
+```
+
+### Single-file ensemble checkpoint (all 14 members + gate in one object)
+```bash
+python final/bundle_ensemble.py          # packs checkpoints/ -> final/bundle.pt (~85 MB, 19 members)
+```
+```python
+from ensemble_model import GatedEnsemble
+ens = GatedEnsemble("final/bundle.pt")                                       # one object = whole ensemble + gate
+probs = ens.predict_proba(perch_emb, harmonic_feat, birdmae_emb, bgwhiten_feat)   # (N,9) gated probabilities
+labels = probs.argmax(1)                                                          # species 0..8
+# ens.member_probs(...) returns the 4 pre-gate group probabilities if you want to inspect the gate.
+```
+`bundle.pt` holds every member's weights + standardization stats + the gate tau, so it is the single
+"checkpoint with the ensemble." It reproduces gated BA_unseen = 0.3616.
+
+For new raw audio: produce Perch (1536-d) and BirdMAE (1024-d) embeddings with those frozen models, the 102-d
+harmonic feature via `harmonic_features.harmonic_feature_from_path(wav)`, and the 257-d background-whitened
+spectrum via `bgwhiten_features.bgwhiten_feature_from_path(wav)`, then feed each to its member checkpoints and
+apply the unanimous-3 gate (see `infer.py`).
+
+---
+
+## 8. Results & honesty
+
+- Best deployable: **unanimous-3 gated ensemble = 0.3616** (base 0.3178; 2-voter 0.3411; +0.020 from the
+  background-whitened 3rd voter). LOSO-stable; gains on the large-n movers (arabiensis 0.148→0.192, quinque
+  0.140→0.196) plus aegypti (0.81).
+- Two levers found beyond the MMD/CORAL recipe: (1) the agreement gate, (2) the **background-whitened 3rd
+  voter** that turns the gate unanimous and stops it sacrificing the movers. sl-BEATs, HumBug (mosquito CNN),
+  CLAP/BioLingual were all tested and did NOT help; CQT helped the kill-test but not the gated metric.
+- The remaining floor is **data/biology-bound**: nothing separates gambiae/arabiensis (acoustically identical
+  congeners) without an in-domain anchor the splits don't provide. gambiae stays ~0.05.

legacy/final/bgwhiten_features.py

@@ -0,0 +1,42 @@
+"""Background-whitened spectrum feature (257-d) — the 3rd gate voter's input.
+
+Idea: a recording's DEVICE channel (EQ tilt + noise floor) is revealed by its non-mosquito BACKGROUND frames.
+Estimate that channel from the low-energy frames and SUBTRACT it from the mean mosquito-frame log-spectrum ->
+a device-invariant 'clean' spectrum. This is the strongest device-invariant species signal we measured on the
+catastrophic studio->field (D5->D1) transfer (cross-domain 0.530 vs Perch 0.140), and as a 3rd unanimous-gate
+voter it vetoes bad overrides and recovers the movers.
+
+    from bgwhiten_features import bgwhiten_feature, bgwhiten_feature_from_path
+    w = bgwhiten_feature(y)                 # y: mono waveform; -> (257,) float32
+    w = bgwhiten_feature_from_path("x.wav") # load + extract
+"""
+import numpy as np
+import librosa, soundfile as sf
+
+SR = 16000; DUR = 1.0; NS = int(SR * DUR); N_FFT = 512; HOP = 128
+BG_Q = 0.30   # background = lowest 30% energy frames (device channel estimate)
+FG_Q = 0.50   # foreground = top 50% energy frames (mosquito present)
+BGW_DIM = N_FFT // 2 + 1   # 257
+
+
+def load_wav(path):
+    y, sr = sf.read(path)
+    if y.ndim > 1: y = y.mean(1)
+    y = y.astype(np.float32)
+    if sr != SR: y = librosa.resample(y, orig_sr=sr, target_sr=SR)
+    y = y[:NS] if len(y) >= NS else np.pad(y, (0, NS - len(y)))
+    m = np.abs(y).max(); return y / m if m > 0 else y
+
+
+def bgwhiten_feature(y):
+    """Mono waveform (any length; trimmed/padded to 1 s) -> 257-d background-whitened log-spectrum."""
+    if len(y) < N_FFT: y = np.pad(y, (0, N_FFT - len(y)))
+    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]
+    bg = np.argsort(e)[:max(3, int(BG_Q * T))]            # quietest frames = device channel
+    fg = np.argsort(e)[::-1][:max(3, int(FG_Q * T))]      # loudest frames = mosquito
+    return (logS[:, fg].mean(1) - logS[:, bg].mean(1)).astype(np.float32)
+
+
+def bgwhiten_feature_from_path(path):
+    return bgwhiten_feature(load_wav(path))

legacy/final/bundle.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02f9e6d23f20eb4e43a777aa6acf9a17e1209d0452fe51bf9d44c56ec005cbf1
+size 85342253

legacy/final/bundle_ensemble.py

@@ -0,0 +1,26 @@
+"""Pack the 14 per-member checkpoints + gate config into ONE file: final/bundle.pt.
+Load it with: GatedEnsemble("final/bundle.pt"). Run: python final/bundle_ensemble.py
+"""
+import os, sys, glob, json
+import torch
+HERE = os.path.dirname(os.path.abspath(__file__)); CKPT = os.path.join(HERE, "checkpoints")
+man = json.load(open(os.path.join(HERE, "ensemble.json")))
+
+bundle = {"tau": man["gate"]["tau"], "members": {},
+          "meta": {"pipeline": man["pipeline"], "description": man["description"],
+                   "results": man["results"], "feature_dims": man["feature_dims"]}}
+n = 0
+for group in ("base_perch", "arm_harmonic", "arm_birdmae", "arm_bgwhiten"):
+    bundle["members"][group] = []
+    for cp in sorted(glob.glob(os.path.join(CKPT, f"{group}_s*.pt"))):
+        st = torch.load(cp, map_location="cpu", weights_only=False)
+        bundle["members"][group].append({
+            "state": st["model_state_dict"], "mean": st["mean"], "std": st["std"],
+            "standardize": st["standardize"], "arch": st["arch"], "feature_dim": st["feature_dim"],
+            "seed": st["seed"]})
+        n += 1
+out = os.path.join(HERE, "bundle.pt")
+torch.save(bundle, out)
+mb = os.path.getsize(out) / 1e6
+print(f"bundled {n} member checkpoints + gate(tau={bundle['tau']}) -> {out}  ({mb:.1f} MB)")
+print("groups:", {g: len(v) for g, v in bundle["members"].items()})

legacy/final/checkpoints/arm_bgwhiten_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ef53039e3e4b73780ea26cdcd1f1c396885df707c45c05f83d187a73380545c
+size 4502101

legacy/final/checkpoints/arm_bgwhiten_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1e88991ab12b71159ea222bcfe0cf824b8c226fd53471f461c2df8cb2feac14
+size 4502101

legacy/final/checkpoints/arm_bgwhiten_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c9dde138a978758e481511c5ef817bd53da87b81ff5e02f490ab5cd1ac59d53
+size 4502045

legacy/final/checkpoints/arm_bgwhiten_s7.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e799f1073610ca7de3f118b357912efb742c50930dc87b1456215fe54a992114
+size 4502017

legacy/final/checkpoints/arm_bgwhiten_s99.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47e7b8a57616a9dc4988a4e6e7256ac117b4c39e14c54dc798cab8da50dcde89
+size 4502045

legacy/final/checkpoints/arm_birdmae_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96068fbe76aae1f47fcf0a631f9d2317e9b1f31441f4a6ebfb50affe251ff712
+size 5302713

legacy/final/checkpoints/arm_birdmae_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c305e2bb41333897e8be6b43fa8430982d6b9f1ff327ebdcd34814656451830
+size 5302713

legacy/final/checkpoints/arm_birdmae_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d72e97e52b31554cb1db42d7028af5187fbc644cbf40a934747b7f8dd4037f0a
+size 5302657

legacy/final/checkpoints/arm_harmonic_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28614528dbbe6eefbe8b6e64c6f0e0e73bda280797ce54d885d9e334e152e5ed
+size 96901

legacy/final/checkpoints/arm_harmonic_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:960ed67d5bc0fc96d49ad373020387fcd3ce248f9b2f827695f5b86cbdb5b021
+size 96901

legacy/final/checkpoints/arm_harmonic_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6e3e7c73e06808149ce226c8dd558b6f44c16b2672ad09d6cfec45ad7a22b7e
+size 96801

legacy/final/checkpoints/base_perch_s123.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2447bb132bac1e6f8e4a08791b9093ae07146521edf56fba6adf80c4327770b
+size 5832833

legacy/final/checkpoints/base_perch_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2e753aba405627f5b3413f926d096a208735699d6d3cfe4d7780ddacf6db7c1
+size 5832861

legacy/final/checkpoints/base_perch_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:166ca097b8a1dc83e8ca38dfea054d968a96a460bbbd688eeb8ec14e529a58aa
+size 5832861

legacy/final/checkpoints/base_perch_s256.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45d1a064959c8c2d5710b19066c6262d417f06361f43a49203860dbf44993409
+size 5832833

legacy/final/checkpoints/base_perch_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5256eb6cf121f5ce843cc476d87dec8bae50a7dcab17982fd51f0716e47db36c
+size 5832805

legacy/final/checkpoints/base_perch_s7.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d09ec9bb2fbdb62e13789574611f1aa34bceee461e993f42884da408f78ed50
+size 5832777

legacy/final/checkpoints/base_perch_s777.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:543ad5fdf750ad863355cbada7e83f2973c7abea58ee2db27021d2c88a4d1d51
+size 5832833

legacy/final/checkpoints/base_perch_s99.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62f361355f32327de0d1341807701e798a17df6da43e77a54f35ccf5236fa262
+size 5832805

legacy/final/checkpoints_backup_2voter/arm_birdmae_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96068fbe76aae1f47fcf0a631f9d2317e9b1f31441f4a6ebfb50affe251ff712
+size 5302713

legacy/final/checkpoints_backup_2voter/arm_birdmae_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c305e2bb41333897e8be6b43fa8430982d6b9f1ff327ebdcd34814656451830
+size 5302713

legacy/final/checkpoints_backup_2voter/arm_birdmae_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d72e97e52b31554cb1db42d7028af5187fbc644cbf40a934747b7f8dd4037f0a
+size 5302657

legacy/final/checkpoints_backup_2voter/arm_harmonic_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28614528dbbe6eefbe8b6e64c6f0e0e73bda280797ce54d885d9e334e152e5ed
+size 96901

legacy/final/checkpoints_backup_2voter/arm_harmonic_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:960ed67d5bc0fc96d49ad373020387fcd3ce248f9b2f827695f5b86cbdb5b021
+size 96901

legacy/final/checkpoints_backup_2voter/arm_harmonic_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6e3e7c73e06808149ce226c8dd558b6f44c16b2672ad09d6cfec45ad7a22b7e
+size 96801

legacy/final/checkpoints_backup_2voter/base_perch_s123.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2447bb132bac1e6f8e4a08791b9093ae07146521edf56fba6adf80c4327770b
+size 5832833

legacy/final/checkpoints_backup_2voter/base_perch_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2e753aba405627f5b3413f926d096a208735699d6d3cfe4d7780ddacf6db7c1
+size 5832861

legacy/final/checkpoints_backup_2voter/base_perch_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:166ca097b8a1dc83e8ca38dfea054d968a96a460bbbd688eeb8ec14e529a58aa
+size 5832861

legacy/final/checkpoints_backup_2voter/base_perch_s256.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45d1a064959c8c2d5710b19066c6262d417f06361f43a49203860dbf44993409
+size 5832833

legacy/final/checkpoints_backup_2voter/base_perch_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5256eb6cf121f5ce843cc476d87dec8bae50a7dcab17982fd51f0716e47db36c
+size 5832805

legacy/final/checkpoints_backup_2voter/base_perch_s7.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d09ec9bb2fbdb62e13789574611f1aa34bceee461e993f42884da408f78ed50
+size 5832777

legacy/final/checkpoints_backup_2voter/base_perch_s777.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:543ad5fdf750ad863355cbada7e83f2973c7abea58ee2db27021d2c88a4d1d51
+size 5832833

legacy/final/checkpoints_backup_2voter/base_perch_s99.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62f361355f32327de0d1341807701e798a17df6da43e77a54f35ccf5236fa262
+size 5832805

legacy/final/checkpoints_backup_2voter/bundle.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e100d18ddd5d41f1062ca096b7f3feb7a673712f0b23751282b7a913bed4ce1
+size 62838583

legacy/final/checkpoints_backup_2voter/ensemble.json

@@ -0,0 +1,74 @@
+{
+  "pipeline": "agreement_gate_ensemble",
+  "description": "8-seed Perch base + harmonic arm + BirdMAE arm -> agreement gate",
+  "members": {
+    "base_perch": {
+      "checkpoints": [
+        "base_perch_s1234.pt",
+        "base_perch_s42.pt",
+        "base_perch_s2023.pt",
+        "base_perch_s7.pt",
+        "base_perch_s99.pt",
+        "base_perch_s123.pt",
+        "base_perch_s256.pt",
+        "base_perch_s777.pt"
+      ],
+      "modality": "perch",
+      "arch": "rp",
+      "standardize": false,
+      "combine": "softmax-mean",
+      "selection": "synthval"
+    },
+    "arm_harmonic": {
+      "checkpoints": [
+        "arm_harmonic_s1234.pt",
+        "arm_harmonic_s42.pt",
+        "arm_harmonic_s2023.pt"
+      ],
+      "modality": "harmonic",
+      "arch": "harm",
+      "standardize": true,
+      "combine": "softmax-mean"
+    },
+    "arm_birdmae": {
+      "checkpoints": [
+        "arm_birdmae_s1234.pt",
+        "arm_birdmae_s42.pt",
+        "arm_birdmae_s2023.pt"
+      ],
+      "modality": "birdmae",
+      "arch": "rp",
+      "standardize": true,
+      "combine": "softmax-mean"
+    }
+  },
+  "gate": {
+    "type": "agreement",
+    "rule": "override base where harm.argmax==birdmae.argmax and min(conf)>tau",
+    "tau": 0.4,
+    "override": "0.5*(harm+birdmae)"
+  },
+  "results": {
+    "base_BA_unseen": 0.3178,
+    "gated_BA_unseen": 0.3411,
+    "gate_delta": 0.0233,
+    "per_species": {
+      "aegypti": 0.7969,
+      "albopictus": 0.3675,
+      "quinquefasciatus": 0.1399,
+      "gambiae": 0.0452,
+      "arabiensis": 0.1478,
+      "dirus": 0.0,
+      "pipiens": 0.8824,
+      "minimus": 0.6364,
+      "stephensi": 0.0541
+    },
+    "selector": "synth-shift val (analysis/test_selection.py); seen-val variant in checkpoints_backup_seenval/"
+  },
+  "git_commit": "9b54ec6",
+  "feature_dims": {
+    "perch": 1536,
+    "harmonic": 102,
+    "birdmae": 1024
+  }
+}

legacy/final/checkpoints_backup_seenval/base_perch_s123.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d6f729c7caa045b2904a768cd45a9d1442e8f0dfac2b6983a52f3a32bcd5faf
+size 5832833

legacy/final/checkpoints_backup_seenval/base_perch_s1234.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aca1b63836f175527bc1203c734ef2b53c1a6ccc970308f06f5da6f5a48d3ce5
+size 5832861

legacy/final/checkpoints_backup_seenval/base_perch_s2023.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fea9f727845497c3fa2ef73fb748e47a13954afc97b8e8d1258b02af2a04868
+size 5832861

legacy/final/checkpoints_backup_seenval/base_perch_s256.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99b047654a78cf7e4d11c2bba156c3093358c15d8fddc311dc693b5ee383d615
+size 5832833

legacy/final/checkpoints_backup_seenval/base_perch_s42.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b576050da34227ffda8f8f0e36f4c99a8df6417b066aef69c96f7df7393050ab
+size 5832805