const PREFIX = "[BoomCast][YAMNet]"; export class BrowserYamnetClassifier { static async classify(waveform, { modelUrl, classMapUrl, signal, loadTensorFlow } = {}) { let tf; let model; let input; let output; const started = performance.now(); try { if (signal?.aborted) throw new DOMException("Cancelled", "AbortError"); console.log(`${PREFIX} Lazy-loading TensorFlow.js.`); tf = loadTensorFlow ? await loadTensorFlow() : await import("@tensorflow/tfjs"); if (signal?.aborted) throw new DOMException("Cancelled", "AbortError"); console.log(`${PREFIX} TensorFlow.js loaded.`); await tf.ready(); console.log(`${PREFIX} Backend:`, tf.getBackend()); console.log(`${PREFIX} Loading model.`); const [loaded, mappingResponse] = await Promise.all([ tf.loadGraphModel(modelUrl), fetch(classMapUrl, { cache: "force-cache", signal }), ]); model = loaded; if (!mappingResponse.ok) throw new Error("YAMNet class map unavailable"); const classMap = await mappingResponse.json(); if (!Array.isArray(classMap) || classMap.length !== 521) throw new Error("Invalid YAMNet class map"); if (signal?.aborted) throw new DOMException("Cancelled", "AbortError"); console.log(`${PREFIX} Model ready.`); // YAMNet needs at least one 0.96 s patch, including STFT padding. const samples = waveform.length < 15600 ? Float32Array.from({ length: 15600 }, (_, i) => waveform[i] || 0) : waveform; input = tf.tensor1d(samples, "float32"); console.log(`${PREFIX} Starting inference.`); const result = await model.executeAsync(input); output = Array.isArray(result) ? result : [result]; const scores = output.find((tensor) => tensor.shape.length === 2 && tensor.shape[1] === classMap.length); if (!scores) throw new Error("YAMNet score output unavailable"); const values = await scores.data(); if (signal?.aborted) throw new DOMException("Cancelled", "AbortError"); const frames = scores.shape[0]; const predictions = classMap.map((name, index) => { let score = -1; let frame = 0; for (let row = 0; row < frames; row += 1) { const value = values[row * classMap.length + index]; if (value > score) { score = value; frame = row; } } return { class: name, score, frame, timeSeconds: Number((frame * 0.48).toFixed(2)) }; }); predictions.sort((a, b) => b.score - a.score); console.log(`${PREFIX} Inference completed in`, Math.round(performance.now() - started), "ms."); console.groupCollapsed(`${PREFIX} Top local predictions`); console.table(predictions.slice(0, 10)); console.groupEnd(); return predictions; } catch (error) { console.error(`${PREFIX} Local inference failed:`, error); throw error; } finally { if (tf) console.log(`${PREFIX} TensorFlow memory before cleanup:`, tf.memory()); console.log("[BoomCast][Cleanup] Disposing YAMNet tensors."); input?.dispose(); output?.forEach((tensor) => tensor?.dispose()); console.log("[BoomCast][Cleanup] Disposing browser YAMNet model."); model?.dispose(); console.log("[BoomCast][Cleanup] Releasing waveform buffers."); if (tf) console.log(`${PREFIX} TensorFlow memory after cleanup:`, tf.memory()); console.log("[BoomCast][Cleanup] Browser classification cleanup complete."); } } }