inference.py

"""Standalone inference for the LAION VocalBurst classifier (VoiceCLAP embedding -> MLP head).
    from inference import VocalBurstClassifier
    clf = VocalBurstClassifier(".")                 # a local checkout, or a HF repo id
    print(clf.predict("clip.wav"))
"""
import os, json, torch, torch.nn as nn, torchaudio
from transformers import AutoModel

class _MLP(nn.Module):
    def __init__(self, d, h, depth, out):
        super().__init__()
        L = [nn.Linear(d, h), nn.LayerNorm(h), nn.GELU(), nn.Dropout(0.0)]
        for _ in range(depth - 1):
            L += [nn.Linear(h, h), nn.LayerNorm(h), nn.GELU(), nn.Dropout(0.0)]
        L += [nn.Linear(h, out)]
        self.net = nn.Sequential(*L)
    def forward(self, x): return self.net(x)

class VocalBurstClassifier:
    def __init__(self, path_or_repo=".", device=None):
        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
        d = path_or_repo
        if not os.path.isdir(d):                       # a HF repo id -> download
            from huggingface_hub import snapshot_download
            d = snapshot_download(path_or_repo)
        self.cfg = json.load(open(f"{d}/config.json"))
        self.classes = json.load(open(f"{d}/classes.json"))
        self.groups = json.load(open(f"{d}/class_to_group.json"))
        self.nb = self.classes.index("no_burst")
        # frozen VoiceCLAP-commercial embedder (auto-downloaded from the public repo)
        self.vc = AutoModel.from_pretrained(self.cfg["embedder"], trust_remote_code=True).to(self.device).eval()
        self.mlp = _MLP(self.cfg["input_dim"], self.cfg["hidden"], self.cfg["depth"], self.cfg["num_classes"]).to(self.device).eval()
        self.mlp.load_state_dict(torch.load(f"{d}/model.pt", map_location=self.device))

    @torch.no_grad()
    def probs(self, audio_path):
        w, sr = torchaudio.load(audio_path); w = w.mean(0) if w.dim() == 2 else w
        w16 = torchaudio.functional.resample(w, sr, 16000) if sr != 16000 else w
        if len(w16) > 16000 * 30: w16 = w16[:16000 * 30]      # VoiceCLAP/Whisper sees first 30 s
        e = self.vc.encode_waveform(w16.to(self.device))
        logit = self.mlp(e).squeeze(0)
        p = torch.sigmoid(logit) if self.cfg["activation"] == "sigmoid" else torch.softmax(logit, -1)
        return p.cpu().numpy()

    def predict(self, audio_path, topk=5, no_burst_gate=0.5):
        p = self.probs(audio_path)
        if float(p[self.nb]) >= no_burst_gate:            # no-burst gate
            return {"no_burst": True, "p_no_burst": float(p[self.nb]), "predictions": []}
        order = p.argsort()[::-1]
        preds = [(self.classes[i], float(p[i])) for i in order if self.classes[i] != "no_burst"][:topk]
        return {"no_burst": False, "p_no_burst": float(p[self.nb]),
                "top1": preds[0][0], "predictions": preds,
                "group": self.groups[preds[0][0]]}