feedBack-desktop/tests/mlnotedetector/bench.cpp
2026-06-16 18:48:12 +02:00

315 lines
13 KiB
C++

// Controlled detection benchmark for MlNoteDetector.
//
// Replays a fixed DI recording through the REAL MlNoteDetector and scores the
// detected onsets against a known note chart — so a parameter change can be
// measured (recall / timing) instead of guessed from noisy live takes.
//
// Build: see CMakeLists.txt. Run:
// ./mlnd_bench <model.onnx> <di-take.wav> <chart.txt> [channel]
// chart.txt : one "<chartTimeSec> <midi>" per line (jq-extracted from a
// note_detect diagnostic: .events[] | "\(.t) \(.ex)").
// channel : mix (default) | left | right
//
// The recording and the chart start at unknown relative offsets, so the
// harness searches for the time offset that best aligns them, then reports
// recall and the timing-error distribution at that offset.
#include "MlNoteDetector.h"
#include <algorithm>
#include <chrono>
#include <cmath>
#include <cstdint>
#include <cstring>
#include <fstream>
#include <iostream>
#include <map>
#include <sstream>
#include <string>
#include <thread>
#include <vector>
namespace
{
uint32_t rdU32(const uint8_t* p) { return p[0] | (p[1]<<8) | (p[2]<<16) | (uint32_t(p[3])<<24); }
uint16_t rdU16(const uint8_t* p) { return uint16_t(p[0] | (p[1]<<8)); }
// Read a 16-bit PCM WAV to mono float. channel: 0 = mix, 1 = left, 2 = right.
bool readWav(const std::string& path, int channelMode, std::vector<float>& out, int& sampleRate)
{
std::ifstream f(path, std::ios::binary);
if (!f) return false;
std::vector<uint8_t> buf((std::istreambuf_iterator<char>(f)), std::istreambuf_iterator<char>());
if (buf.size() < 44 || std::memcmp(buf.data(), "RIFF", 4) || std::memcmp(buf.data()+8, "WAVE", 4))
return false;
uint16_t fmt = 0, channels = 0, bits = 0;
uint32_t rate = 0, dataLen = 0;
const uint8_t* data = nullptr;
size_t pos = 12;
while (pos + 8 <= buf.size())
{
const char* id = reinterpret_cast<const char*>(buf.data() + pos);
const uint32_t sz = rdU32(buf.data() + pos + 4);
const uint8_t* body = buf.data() + pos + 8;
// Guard the fmt-body reads (up to body+14, i.e. 16 bytes) against a
// truncated file: a declared sz >= 16 doesn't mean 16 bytes are
// actually present.
if (!std::memcmp(id, "fmt ", 4) && sz >= 16 && pos + 8 + 16 <= buf.size())
{ fmt = rdU16(body); channels = rdU16(body+2); rate = rdU32(body+4); bits = rdU16(body+14); }
else if (!std::memcmp(id, "data", 4))
{ data = body; dataLen = std::min<uint32_t>(sz, uint32_t(buf.size() - (pos + 8))); }
pos += 8 + sz + (sz & 1);
}
if (!data || channels == 0 || rate == 0 || fmt != 1 || bits != 16) return false;
sampleRate = int(rate);
const size_t frames = dataLen / (size_t(2) * channels);
out.resize(frames);
for (size_t i = 0; i < frames; ++i)
{
auto sample = [&](int c) -> double {
return int16_t(rdU16(data + (i * channels + c) * 2)) / 32768.0;
};
double v;
if (channelMode == 1) v = sample(0);
else if (channelMode == 2 && channels > 1) v = sample(1);
else { double a = 0; for (int c = 0; c < channels; ++c) a += sample(c); v = a / channels; }
out[i] = float(v);
}
return true;
}
struct ChartNote { double t; int midi; };
struct Onset { double t; int midi; float conf; };
// One bridge poll: the full active-note set, mirroring audio.detectNotes().
struct PollRec { double t; std::vector<MlNoteDetector::ActiveNote> notes; };
} // namespace
int main(int argc, char** argv)
{
if (argc < 4)
{
std::cerr << "usage: mlnd_bench <model.onnx> <di-take.wav> <chart.txt> [mix|left|right]\n";
return 2;
}
int channelMode = 0;
if (argc > 4)
{
const std::string c = argv[4];
channelMode = (c == "left") ? 1 : (c == "right") ? 2 : 0;
}
std::vector<float> wav;
int sampleRate = 0;
if (!readWav(argv[2], channelMode, wav, sampleRate))
{ std::cerr << "FAIL: cannot read 16-bit WAV " << argv[2] << "\n"; return 1; }
const double wavSec = double(wav.size()) / sampleRate;
std::cout << "WAV: " << wav.size() << " samples @ " << sampleRate << " Hz ("
<< wavSec << " s), channel=" << (channelMode==1?"left":channelMode==2?"right":"mix") << "\n";
std::vector<ChartNote> chart;
{
std::ifstream cf(argv[3]);
if (!cf) { std::cerr << "FAIL: cannot read chart " << argv[3] << "\n"; return 1; }
std::string line;
while (std::getline(cf, line))
{
std::istringstream is(line);
ChartNote n{};
if (is >> n.t >> n.midi) chart.push_back(n);
}
}
std::cout << "chart: " << chart.size() << " notes\n";
MlNoteDetector det;
if (!det.loadModel(juce::File(juce::String(argv[1]))))
{ std::cerr << "FAIL: loadModel returned false\n"; return 1; }
det.prepare((double) sampleRate, 256);
// Feed the WAV at real time (1.0x) and poll the active set every 50 ms,
// the same cadence as the plugin. A rising per-pitch onsetSeq is a
// detected onset; back-date it by onsetAgeMs to its true time.
//
// feedRate MUST stay 1.0: onsetAgeMs is measured in wall-clock time inside
// MlNoteDetector, while fedSec is the fed-audio timeline. Feeding faster
// than real time desynchronises the two — a wall-clock age would map to a
// larger span of fed audio — so back-dated onset times would drift later
// than the chart over a long recording. At 1.0x the two timelines agree.
const int block = 256;
const double feedRate = 1.0;
std::map<int, int> lastSeq;
std::vector<Onset> onsets;
std::vector<PollRec> polls; // full detectNotes() stream, one entry per poll
double nextPollSec = 0.0;
for (size_t i = 0; i < wav.size(); i += block)
{
const int n = (int) std::min<size_t>(block, wav.size() - i);
det.pushSamples(wav.data() + i, n);
std::this_thread::sleep_for(std::chrono::microseconds(
(long long) (1e6 * n / sampleRate / feedRate)));
const double fedSec = double(i + n) / sampleRate;
if (fedSec >= nextPollSec)
{
nextPollSec += 0.050;
auto active = det.getActiveNotes();
polls.push_back({ fedSec, active });
for (const auto& a : active)
{
auto it = lastSeq.find(a.midi);
const int prevSeq = (it == lastSeq.end()) ? 0 : it->second;
// onsetSeq == 0 means "no detected onset" (sustained activity),
// so only a strictly-advancing, non-zero counter is a real new
// onset — otherwise back-dating by the sentinel age would forge
// an onset at poll time. Still track the pitch either way.
if (a.onsetSeq > prevSeq && a.onsetSeq > 0)
{
const double age = (a.onsetAgeMs < 1.0e6f) ? a.onsetAgeMs / 1000.0 : 0.0;
onsets.push_back({ fedSec - age, a.midi, a.confidence });
}
lastSeq[a.midi] = a.onsetSeq;
}
}
}
constexpr double kDrainSec = 0.400;
std::this_thread::sleep_for(std::chrono::milliseconds(
(int) (kDrainSec * 1000)));
// One final poll after the drain delay: late onsets / active notes the
// detector only resolved during the trailing inference would otherwise be
// omitted from the metrics and detectstream.json. Wall time has advanced
// by the drain sleep, so the poll timestamp is wavSec + kDrainSec — using
// a bare wavSec would back-date trailing onsets by the drain duration,
// since onsetAgeMs is measured at this (later) wall-clock instant.
{
const double finalPollSec = wavSec + kDrainSec;
auto active = det.getActiveNotes();
polls.push_back({ finalPollSec, active });
for (const auto& a : active)
{
auto it = lastSeq.find(a.midi);
const int prevSeq = (it == lastSeq.end()) ? 0 : it->second;
// See the in-loop poll above: onsetSeq 0 is "no onset", not a hit.
if (a.onsetSeq > prevSeq && a.onsetSeq > 0)
{
const double age = (a.onsetAgeMs < 1.0e6f) ? a.onsetAgeMs / 1000.0 : 0.0;
onsets.push_back({ finalPollSec - age, a.midi, a.confidence });
}
lastSeq[a.midi] = a.onsetSeq;
}
}
det.stop();
std::cout << "detected onsets: " << onsets.size() << "\n";
// --- Align: find the chart->WAV time offset maximising matches ----------
const double tol = 0.10; // ±100 ms match window
auto countMatches = [&](double delta) {
int m = 0;
for (const auto& c : chart)
{
const double target = c.t + delta;
for (const auto& o : onsets)
if (o.midi == c.midi && std::fabs(o.t - target) <= tol) { ++m; break; }
}
return m;
};
double bestDelta = 0.0;
int bestMatches = -1;
for (double d = -2.0; d <= 30.0; d += 0.010)
{
const int m = countMatches(d);
if (m > bestMatches) { bestMatches = m; bestDelta = d; }
}
// --- Dump the detect-stream (chart-aligned) for the JS matching harness --
{
std::ofstream js("detectstream.json");
js << "{\"offset\":" << bestDelta << ",\"polls\":[";
for (size_t pi = 0; pi < polls.size(); ++pi)
{
if (pi) js << ",";
js << "{\"t\":" << (polls[pi].t - bestDelta) << ",\"notes\":[";
for (size_t ni = 0; ni < polls[pi].notes.size(); ++ni)
{
const auto& a = polls[pi].notes[ni];
if (ni) js << ",";
js << "{\"midi\":" << a.midi
<< ",\"confidence\":" << a.confidence
<< ",\"onsetMs\":" << a.onsetAgeMs
<< ",\"onsetSeq\":" << a.onsetSeq << "}";
}
js << "]}";
}
js << "]}";
}
std::cout << "detect-stream: " << polls.size() << " polls -> detectstream.json\n";
// --- Report at the best offset -----------------------------------------
std::vector<double> te; // timing errors of matched notes
int matched = 0;
for (const auto& c : chart)
{
const double target = c.t + bestDelta;
double best = 1e9;
for (const auto& o : onsets)
if (o.midi == c.midi && std::fabs(o.t - target) <= tol)
if (std::fabs(o.t - target) < std::fabs(best)) best = o.t - target;
if (best < 1e8) { ++matched; te.push_back(best); }
}
std::sort(te.begin(), te.end());
auto pct = [&](double p) {
return te.empty() ? 0.0 : te[std::min(te.size()-1, (size_t)(p * te.size()))];
};
std::cout << "\n=== alignment ===\n";
std::cout << "chart->WAV offset: " << bestDelta << " s\n";
std::cout << "\n=== detection quality (±" << (tol*1000) << " ms) ===\n";
std::cout << "recall: " << matched << " / " << chart.size()
<< " (" << (100.0 * matched / std::max<size_t>(1, chart.size())) << "%)\n";
if (!te.empty())
{
double sum = 0; for (double x : te) sum += x;
std::cout << "timing error: median " << (te[te.size()/2]*1000) << " ms"
<< " p10 " << (pct(0.10)*1000) << " p90 " << (pct(0.90)*1000)
<< " mean " << (sum/te.size()*1000) << " ms\n";
}
std::cout << "onsets/note: " << (double(onsets.size()) / std::max<size_t>(1, chart.size()))
<< " (>1 = extra detections: harmonics, noise)\n";
// --- Predicted score: one-onset-one-note greedy matching ----------------
// Mirrors the plugin's fixed matcher — each onset (earliest first) claims
// the nearest still-unclaimed chart note of its pitch within ±tol. This
// predicts the live hit rate, where the "recall" above is the loose
// upper bound (one onset allowed to satisfy many notes).
{
std::vector<Onset> sorted = onsets;
std::sort(sorted.begin(), sorted.end(),
[](const Onset& a, const Onset& b){ return a.t < b.t; });
std::vector<char> claimed(chart.size(), 0);
std::vector<double> hte;
for (const auto& o : sorted)
{
int bestIdx = -1; double bestDist = 1e9;
for (size_t ci = 0; ci < chart.size(); ++ci)
{
if (claimed[ci] || chart[ci].midi != o.midi) continue;
const double d = std::fabs(o.t - (chart[ci].t + bestDelta));
if (d <= tol && d < bestDist) { bestDist = d; bestIdx = (int) ci; }
}
if (bestIdx >= 0)
{ claimed[(size_t) bestIdx] = 1; hte.push_back(o.t - (chart[(size_t) bestIdx].t + bestDelta)); }
}
std::sort(hte.begin(), hte.end());
const size_t hits = hte.size();
std::cout << "\n=== predicted score (one-onset-one-note) ===\n";
std::cout << "hits: " << hits << " / " << chart.size()
<< " (" << (100.0 * hits / std::max<size_t>(1, chart.size())) << "%)\n";
if (!hte.empty())
std::cout << "timing error: median " << (hte[hte.size()/2]*1000) << " ms"
<< " p10 " << (hte[hte.size()/10]*1000)
<< " p90 " << (hte[hte.size()*9/10]*1000) << " ms\n";
}
return 0;
}