# /// script # requires-python = ">=3.11" # dependencies = [ # "torch>=2.4", # "torchvision>=0.19", # "timm>=1.0", # "datasets>=3.0", # "pyarrow", # "onnx>=1.16", # "onnxruntime>=1.18", # "pillow", # "numpy", # "huggingface_hub>=0.30", # ] # /// """Challenge-5 acoustic wake model: train on the unified dataset, evaluate, export ONNX, publish to HF. Runs as a Hugging Face Job (uv). One arch x one seed per job. Usage: job_train.py --arch mobilenet_v3_small --seed 42 [--epochs 10] [--batch 64] """ import argparse, io, json, math, os, random, time from pathlib import Path import numpy as np ARCHES = ("mobilenet_v3_small", "shufflenet_v2_x0_5", "squeezenet1_1", "efficientnet_lite0", "ghostnet_050", "cnn6_custom") SOUND_CLASSES = ("drone_buzz", "vehicle_mechanical", "explosion", "ambient_outdoor") CLASS_NAMES_UNIFIED = ("drone_buzz", "vehicle_mechanical", "explosion", "ambient_outdoor", "unknown") DATASET = "Tonic/unified-drone-acoustic-cuas" RESULTS_REPO = "Tonic/challenge5-acoustic-results" def build_backbone(arch: str, pretrained: bool): """Returns (backbone_module, feat_dim). Backbone maps [B,3,224,224] -> [B,feat_dim].""" import torch.nn as nn from torchvision import models if arch == "mobilenet_v3_small": m = models.mobilenet_v3_small(weights="IMAGENET1K_V1" if pretrained else None) feat = m.classifier[0].in_features m.classifier = nn.Identity() return m, feat if arch == "shufflenet_v2_x0_5": m = models.shufflenet_v2_x0_5(weights="IMAGENET1K_V1" if pretrained else None) feat = m.fc.in_features m.fc = nn.Identity() return m, feat if arch == "squeezenet1_1": m = models.squeezenet1_1(weights="IMAGENET1K_V1" if pretrained else None) feat = 512 class Sq(nn.Module): def __init__(self, m): super().__init__() self.features = m.features def forward(self, x): return self.features(x).mean(dim=(2, 3)) return Sq(m), feat if arch in ("efficientnet_lite0", "ghostnet_050"): import timm names = {"efficientnet_lite0": ["tf_efficientnet_lite0", "efficientnet_lite0"], "ghostnet_050": ["ghostnet_050", "ghostnet_100"]}[arch] last = None import torch for n in names: try: m = timm.create_model(n, pretrained=pretrained, num_classes=0) m.eval() with torch.no_grad(): feat = int(m(torch.zeros(1, 3, 224, 224)).shape[1]) return m, feat except Exception as e: last = e raise RuntimeError(f"timm model for {arch} failed: {last}") if arch == "cnn6_custom": import torch.nn as nn ch = (32, 64, 128, 256, 256, 512) blocks = [] prev = 3 for c in ch: blocks += [nn.Conv2d(prev, c, 3, padding=1, bias=False), nn.BatchNorm2d(c), nn.ReLU(inplace=True), nn.Conv2d(c, c, 3, padding=1, groups=c, bias=False), # depthwise nn.BatchNorm2d(c), nn.ReLU(inplace=True), nn.Conv2d(c, c, 1, bias=False), nn.BatchNorm2d(c), nn.ReLU(inplace=True), nn.MaxPool2d(2)] prev = c class CNN6(nn.Module): def __init__(self): super().__init__() self.net = nn.Sequential(*blocks) def forward(self, x): return self.net(x).mean(dim=(2, 3)) return CNN6(), ch[-1] raise ValueError(arch) def main(): import torch import torch.nn as nn ap = argparse.ArgumentParser() ap.add_argument("--arch", choices=ARCHES, required=True) ap.add_argument("--seed", type=int, default=42) ap.add_argument("--epochs", type=int, default=15) ap.add_argument("--batch", type=int, default=64) ap.add_argument("--lr", type=float, default=3e-4) ap.add_argument("--dataset", type=str, default=DATASET) ap.add_argument("--max-train-rows", type=int, default=160000) ap.add_argument("--val-rows", type=int, default=4000) ap.add_argument("--pretrained", type=int, default=1) ap.add_argument("--out", type=Path, default=Path("/tmp/c5_out")) args = ap.parse_args() random.seed(args.seed); np.random.seed(args.seed); torch.manual_seed(args.seed) if torch.cuda.is_available(): torch.cuda.manual_seed_all(args.seed) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"[job] arch={args.arch} seed={args.seed} device={device} " f"cuda={torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'none'}", flush=True) # ---------- data ---------- from datasets import load_dataset, concatenate_datasets t0 = time.time() ds = load_dataset(args.dataset, "windows_5s") print(f"[data] loaded in {time.time()-t0:.0f}s: { {k: len(v) for k, v in ds.items()} }", flush=True) unames = ds["train"].features["sound_class"].names assert unames == list(CLASS_NAMES_UNIFIED), unames def to_xy(split, source=None, labeled=True, cap=None): d = ds[split] if source is not None: d = d.filter(lambda r: r["source"] == source, num_proc=4) if labeled: d = d.filter(lambda r: r["labeled"], num_proc=4) if cap and len(d) > cap: d = d.shuffle(seed=args.seed).select(range(cap)) return d train_d = to_xy("train", cap=args.max_train_rows) # oversample the 48 challenge5 train clips so the challenge distribution is not # drowned out by the (easy) external corpora c5_train = ds["train"].filter(lambda r: r["source"] == "challenge5", num_proc=4) rep = max(1, min(200, (args.max_train_rows // 20) // max(1, len(c5_train)))) c5_up = concatenate_datasets([c5_train] * (rep + 1)) train_d = concatenate_datasets([train_d, c5_up]) print(f"[data] challenge5 train oversampled x{rep} -> +{len(c5_up)} rows; train total={len(train_d)}", flush=True) val_d = to_xy("validation", cap=args.val_rows) c5_valid = to_xy("validation", source="challenge5", labeled=True) c5_test = ds["test"].filter(lambda r: r["source"] == "challenge5", num_proc=4) print(f"[data] train={len(train_d)} val={len(val_d)} c5_valid={len(c5_valid)} c5_test={len(c5_test)}", flush=True) from PIL import Image from torch.utils.data import DataLoader def augment(img, is_c5): # img: HxWx3 float 0..1; gain + freq/time masks (stronger for challenge rows) g = random.uniform(0.4, 2.0) if is_c5 else random.uniform(0.8, 1.25) img = img * g if is_c5 and random.random() < 0.7: h, w = img.shape[0], img.shape[1] fm = random.randint(0, max(1, h // 8)); f0 = random.randint(0, h - fm) img[f0:f0 + fm] = 0 tm = random.randint(0, max(1, w // 8)); t0 = random.randint(0, w - tm) img[:, t0:t0 + tm] = 0 return np.clip(img, 0, 1) def collate(batch): xs = torch.stack([torch.from_numpy(augment(np.asarray(r["spectrogram"], dtype=np.float32) / 255.0, r["source"] == "challenge5")) .permute(2, 0, 1) for r in batch]) wake = torch.tensor([float(r["has_drone_audio"]) for r in batch]) cls = torch.tensor([SOUND_CLASSES.index(CLASS_NAMES_UNIFIED[r["sound_class"]]) if r["sound_class"] < 4 else 0 for r in batch], dtype=torch.long) ids = [r["sample_id"] for r in batch] return xs, wake, cls, ids train_loader = DataLoader(train_d, batch_size=args.batch, shuffle=True, num_workers=4, collate_fn=collate, drop_last=True) val_loader = DataLoader(val_d, batch_size=256, shuffle=False, num_workers=4, collate_fn=collate) c5v_loader = DataLoader(c5_valid, batch_size=16, shuffle=False, collate_fn=collate) c5t_loader = DataLoader(c5_test, batch_size=16, shuffle=False, collate_fn=collate) # ---------- model ---------- class WakeCNN(nn.Module): def __init__(self, arch, pretrained): super().__init__() self.register_buffer("mean", torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1)) self.register_buffer("std", torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)) self.backbone, feat = build_backbone(arch, pretrained) self.head_wake = nn.Linear(feat, 1) self.head_class = nn.Linear(feat, 4) def forward(self, x): x = (x - self.mean) / self.std f = self.backbone(x) return self.head_wake(f).squeeze(-1), self.head_class(f) model = WakeCNN(args.arch, bool(args.pretrained)).to(device) n_params = sum(p.numel() for p in model.parameters()) print(f"[model] {args.arch} params={n_params/1e6:.2f}M", flush=True) opt = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=1e-4) sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=args.epochs) pos = float(np.mean([r["has_drone_audio"] for r in train_d.select(range(2000))])) pos_weight = torch.tensor([min(2.0, max(1.0, ((1 - pos) / max(pos, 1e-3)) ** 0.5))], device=device) print(f"[train] pos_rate~{pos:.3f} pos_weight={float(pos_weight):.3f}", flush=True) bce = nn.BCEWithLogitsLoss(pos_weight=pos_weight) ce = nn.CrossEntropyLoss() def eval_loader(loader, model): model.eval() dev = next(model.parameters()).device W, Y, C, P, ids = [], [], [], [], [] with torch.no_grad(): for xs, wake, cls, sids in loader: wl, cl = model(xs.to(dev)) W.append(torch.sigmoid(wl).cpu()); Y.append(wake); C.append(cls) P.append(cl.argmax(1).cpu()); ids += sids return (torch.cat(W).numpy(), torch.cat(Y).numpy(), torch.cat(C).numpy(), torch.cat(P).numpy(), ids) def sweep(W, Y, grid=None): grid = grid or [round(0.05 + 0.05 * i, 2) for i in range(18)] best = (-1, 0.5, None) for t in grid: p = W >= t tp = int((p & (Y > 0.5)).sum()); fp = int((p & (Y <= 0.5)).sum()); fn = int((~p & (Y > 0.5)).sum()) f1 = 2 * tp / max(1, 2 * tp + fp + fn) u = f1 - 0.15 * float(p.mean()) if u > best[0]: best = (u, t, (f1, float(p.mean()), tp, fp, fn)) return best def sweep_with_c5(W, Y, Wv, Yv): """Pick threshold maximizing unified-val utility; among thresholds within 0.03 utility of the best, tie-break by challenge5-valid F1 (8 clips).""" grid = [round(0.05 + 0.05 * i, 2) for i in range(18)] scored = [] for t in grid: p = W >= t tp = int((p & (Y > 0.5)).sum()); fp = int((p & (Y <= 0.5)).sum()); fn = int((~p & (Y > 0.5)).sum()) f1 = 2 * tp / max(1, 2 * tp + fp + fn) u = f1 - 0.15 * float(p.mean()) pv = Wv >= t tpv = int((pv & (Yv > 0.5)).sum()); fpv = int((pv & (Yv <= 0.5)).sum()); fnv = int((~pv & (Yv > 0.5)).sum()) f1v = 2 * tpv / max(1, 2 * tpv + fpv + fnv) scored.append((u, t, f1, float(p.mean()), f1v, pv.mean())) best_u = max(x[0] for x in scored) cands = [x for x in scored if x[0] >= best_u - 0.03] cands.sort(key=lambda x: (x[4], -x[5])) # best c5 F1, then lowest duty u, t, f1, duty, f1v, dutyv = cands[0] return u, t, (f1, duty, f1v, dutyv) args.out.mkdir(parents=True, exist_ok=True) log_path = args.out / "train_log.txt" logf = open(log_path, "a") def log(msg): print(msg, flush=True) logf.write(msg + "\n"); logf.flush() best_u, best_t, best_state = -1, 0.5, None for ep in range(1, args.epochs + 1): model.train() t0 = time.time() n_seen = 0 scaler = torch.amp.GradScaler("cuda", enabled=device.type == "cuda") for xs, wake, cls, _ in train_loader: opt.zero_grad() with torch.amp.autocast("cuda", enabled=device.type == "cuda"): wl, cl = model(xs.to(device)) loss = bce(wl, wake.to(device)) + 0.5 * ce(cl, cls.to(device)) scaler.scale(loss).backward() scaler.step(opt) scaler.update() n_seen += len(xs) sched.step() W, Y, C, P, _ = eval_loader(val_loader, model) Wv, Yv, Cv, Pv, _ = eval_loader(c5v_loader, model) u, t, stats = sweep_with_c5(W, Y, Wv, Yv) cls_acc_v = float((Pv == Cv).mean()) log(f"[ep {ep}/{args.epochs}] {time.time()-t0:.0f}s rows={n_seen} " f"unified_val: F1={stats[0]:.4f} duty={stats[1]:.3f} @thr={t:.2f} u={u:.4f} | " f"c5_valid: F1={stats[2]:.4f} duty={stats[3]:.3f} cls_acc={cls_acc_v:.3f}") if u > best_u: best_u, best_t = u, t best_state = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()} torch.save(best_state, args.out / "best.pt") if best_state is not None: model.load_state_dict(best_state) model.eval().cpu() # final metrics W, Y, C, P, _ = eval_loader(val_loader, model) Wv, Yv, Cv, Pv, ids_v = eval_loader(c5v_loader, model) u, t, stats = sweep_with_c5(W, Y, Wv, Yv) pv = Wv >= best_t tp = int((pv & (Yv > 0.5)).sum()); fp = int((pv & (Yv <= 0.5)).sum()); fn = int((~pv & (Yv > 0.5)).sum()) f1v = 2 * tp / max(1, 2 * tp + fp + fn) dutyv = float(pv.mean()) cls_acc_v = float((Pv == Cv).mean()) log(f"[final] thr={best_t:.2f} c5_valid wake_F1={f1v:.4f} duty={dutyv:.3f} cls_acc={cls_acc_v:.4f} " f"(tp={tp} fp={fp} fn={fn}) unified_val_u={best_u:.4f}") # ---------- ONNX export with threshold metadata ---------- import onnx dummy = torch.rand(1, 3, 224, 224) onnx_path = args.out / "audio_wake_cnn.onnx" torch.onnx.export(model, (dummy,), str(onnx_path), opset_version=17, input_names=["spectrogram"], output_names=["wake_logit", "sound_class_logits"], dynamo=False) m = onnx.load(str(onnx_path)) md = {e.key: e.value for e in m.metadata_props} md["wake_threshold"] = str(best_t) md["arch"] = args.arch onnx.helper.set_model_props(m, md) onnx.save(m, str(onnx_path)) onnx_mb = onnx_path.stat().st_size / 1e6 log(f"[onnx] saved {onnx_mb:.2f} MB, threshold={best_t:.2f}") # ---------- ORT latency ---------- import onnxruntime as ort sess = ort.InferenceSession(str(onnx_path), providers=["CPUExecutionProvider"]) x = np.random.rand(1, 3, 224, 224).astype(np.float32) for _ in range(10): sess.run(None, {"spectrogram": x}) ts = [] for _ in range(100): t0 = time.time() sess.run(None, {"spectrogram": x}) ts.append((time.time() - t0) * 1000) lat_ms = float(np.mean(ts)) log(f"[ort] CPU latency mean={lat_ms:.2f} ms p50={float(np.percentile(ts,50)):.2f}") # ---------- predictions on challenge5 valid + test ---------- def predict(loader): outs = [] with torch.no_grad(): for xs, wake, cls, sids in loader: wl, cl = model(xs) wp = torch.sigmoid(wl).numpy() cp = cl.argmax(1).numpy() for sid, w, c in zip(sids, wp, cp): clip = sid.split("__")[1].replace(".wav", "") outs.append({"id": clip, "wake_triggered": bool(w >= best_t), "sound_class": SOUND_CLASSES[int(c)]}) return outs preds_valid = predict(c5v_loader) preds_test = predict(c5t_loader) (args.out / "preds_valid.jsonl").write_text( "\n".join(json.dumps(p) for p in preds_valid) + "\n") (args.out / "preds_test.jsonl").write_text( "\n".join(json.dumps(p) for p in preds_test) + "\n") # ---------- publish ---------- metrics = { "arch": args.arch, "seed": args.seed, "epochs": args.epochs, "batch": args.batch, "lr": args.lr, "params": n_params, "onnx_mb": round(onnx_mb, 2), "wake_threshold": best_t, "c5_valid": {"wake_F1": round(f1v, 4), "optical_duty_cycle": round(dutyv, 4), "sound_class_accuracy": round(cls_acc_v, 4), "optical_inactivity_pct": round((1 - dutyv) * 100, 2), "tp": tp, "fp": fp, "fn": fn}, "unified_val": {"utility": round(best_u, 4), "thr": best_t, "F1": round(stats[0], 4), "duty": round(stats[1], 4)}, "ort_cpu_latency_ms": round(lat_ms, 2), "train_rows": len(train_d), "val_rows": len(val_d), "dataset": args.dataset, } (args.out / "metrics.json").write_text(json.dumps(metrics, indent=2)) repo = f"Tonic/challenge5-acoustic-{args.arch}-s{args.seed}" from huggingface_hub import HfApi api = HfApi() api.create_repo(repo, repo_type="model", exist_ok=True, private=False) try: api.update_repo_settings(repo_id=repo, repo_type="model", gated="manual") except Exception as e: log(f"[publish] gating note: {str(e)[:80]}") card = f"""--- license: apache-2.0 tags: [acoustic-wake, c-uas, challenge-5, onnx, drone-detection] --- # Challenge-5 acoustic wake model: {args.arch} (seed {args.seed}) Two-head CNN over 224x224 log-mel windows: `wake_logit` + `sound_class_logits` (4 classes), ImageNet normalization baked in-graph, `wake_threshold` stored in ONNX metadata. ## Metrics (challenge5 valid, {len(c5_valid)} clips) ```json {json.dumps(metrics['c5_valid'], indent=2)} ``` ORT CPU latency: {lat_ms:.2f} ms/window | params {n_params/1e6:.2f}M | ONNX {onnx_mb:.2f} MB Trained on {args.dataset} `windows_5s` (train, {len(train_d)} rows) + challenge5 train clips. Predictions: `preds_valid.jsonl`, `preds_test.jsonl` (challenge submission format). """ (args.out / "README.md").write_text(card) api.upload_folder(folder_path=str(args.out), path_in_repo="", repo_id=repo, repo_type="model", commit_message=f"{args.arch} seed {args.seed}: F1={f1v:.3f} acc={cls_acc_v:.3f} thr={best_t:.2f}") log(f"[publish] https://huggingface.co/{repo}") # append to shared leaderboard dataset repo try: api.create_repo(RESULTS_REPO, repo_type="dataset", exist_ok=True, private=False) row = json.dumps(metrics) + "\n" try: old = api.hf_hub_download(RESULTS_REPO, "results.jsonl", repo_type="dataset") data = open(old, "rb").read() + row.encode() except Exception: data = row.encode() api.upload_file(path_or_fileobj=data, path_in_repo="results.jsonl", repo_id=RESULTS_REPO, repo_type="dataset", commit_message=f"add {args.arch} s{args.seed}") log(f"[publish] leaderboard {RESULTS_REPO} updated") except Exception as e: log(f"[publish] leaderboard append failed: {str(e)[:120]}") log(f"[done] arch={args.arch} seed={args.seed} F1={f1v:.4f} acc={cls_acc_v:.4f} duty={dutyv:.3f} thr={best_t:.2f} lat={lat_ms:.2f}ms") if __name__ == "__main__": main()