Download cpp/tools/check.cpp from RL10x/tiny-voca-hindi: direct link, hf CLI and curl.
- Browser
- Download file 5.25 kB
-
https://huggingface.co/RL10x/tiny-voca-hindi/resolve/main/cpp/tools/check.cpp
- Command line
-
hf download hf://RL10x/tiny-voca-hindi/cpp/tools/check.cpp
-
curl -L -o check.cpp https://huggingface.co/RL10x/tiny-voca-hindi/resolve/main/cpp/tools/check.cpp
5.25 kB
| // Compare the C++ engine with the PyTorch reference dumps (scripts/tts_hi_reference.py), stage by stage. | |
| // tts_check model.gguf tts_cpp/ref/case0 [threads] | |
| template <typename T> | |
| static std::vector<T> read_bin(const std::string& path) { | |
| std::ifstream f(path, std::ios::binary | std::ios::ate); | |
| if (!f) { std::fprintf(stderr, "missing %s\n", path.c_str()); std::exit(1); } | |
| const size_t n = size_t(f.tellg()) / sizeof(T); | |
| std::vector<T> v(n); | |
| f.seekg(0); | |
| f.read(reinterpret_cast<char*>(v.data()), std::streamsize(n * sizeof(T))); | |
| return v; | |
| } | |
| static void compare(const char* name, const std::vector<float>& ref, const std::vector<float>& got) { | |
| if (ref.size() != got.size()) { | |
| std::printf(" %-15s size differs: ref %zu vs engine %zu\n", name, ref.size(), got.size()); | |
| return; | |
| } | |
| double max_abs = 0, err = 0, pow_ = 0; | |
| for (size_t i = 0; i < ref.size(); ++i) { | |
| const double d = double(got[i]) - ref[i]; | |
| max_abs = std::max(max_abs, std::fabs(d)); | |
| err += d * d; pow_ += double(ref[i]) * ref[i]; | |
| } | |
| const double snr = 10 * std::log10(pow_ / std::max(err, 1e-30)); | |
| std::printf(" %-15s max|diff| %.3g SNR %.1f dB\n", name, max_abs, snr); | |
| } | |
| int main(int argc, char** argv) { | |
| if (argc < 3) { std::fprintf(stderr, "usage: tts_check model.gguf ref_case_dir [threads]\n"); return 2; } | |
| const std::string dir = argv[2]; | |
| auto model = TtsModel::load(argv[1], argc > 3 ? std::atoi(argv[3]) : 4); | |
| int style = 0; unsigned seed = 0; float octave = 1.6f; int steps = 0; float temperature = -1.0f, guidance = 0.0f; | |
| { std::ifstream f(dir + "/inputs.txt"); f >> style >> seed >> octave; f >> steps >> temperature; f >> guidance; } | |
| const auto ids = read_bin<int32_t>(dir + "/ids.bin"), stress = read_bin<int32_t>(dir + "/stress.bin"); | |
| const bool flow = std::ifstream(dir + "/x0.bin").good(); // scripts/tts_hi_flow_reference.py | |
| const auto noise = read_bin<float>(dir + (flow ? "/x0.bin" : "/noise.bin")); | |
| TtsOptions opt; opt.style = style; opt.octave_fix = octave; opt.noise = noise.data(); | |
| std::vector<float> dur_noise, pe_noise; | |
| if (flow) { | |
| opt.flow_steps = steps; opt.temperature = temperature; opt.guidance = guidance; | |
| if (std::ifstream(dir + "/dur_x0.bin").good()) { // prosody-flow model | |
| dur_noise = read_bin<float>(dir + "/dur_x0.bin"); pe_noise = read_bin<float>(dir + "/pe_x0.bin"); | |
| opt.duration_noise = dur_noise.data(); opt.pitch_energy_noise = pe_noise.data(); | |
| } | |
| } | |
| TtsDebug dbg; | |
| auto t0 = std::chrono::steady_clock::now(); | |
| const std::vector<float> wave = model->synthesize(ids, stress, opt, &dbg); | |
| const double ms = std::chrono::duration<double, std::milli>(std::chrono::steady_clock::now() - t0).count(); | |
| const auto ref_dur = read_bin<int32_t>(dir + "/durations.bin"); | |
| int mismatch = 0; | |
| for (size_t i = 0; i < ref_dur.size() && i < dbg.durations.size(); ++i) mismatch += ref_dur[i] != dbg.durations[i]; | |
| std::printf("%s: %zu tokens, style %s, %d frames (%.2f s audio) in %.1f ms\n", dir.c_str(), ids.size(), | |
| model->styles()[style].c_str(), dbg.frames, double(wave.size()) / model->sample_rate(), ms); | |
| std::printf(" durations %d of %zu tokens differ\n", mismatch, ref_dur.size()); | |
| if (flow) { | |
| if (mismatch == 0) { | |
| compare("flow_cond", read_bin<float>(dir + "/cond.bin"), dbg.flow_cond); | |
| compare("mel", read_bin<float>(dir + "/mel.bin"), dbg.mel); | |
| compare("wave", read_bin<float>(dir + "/wave.bin"), wave); | |
| } else { | |
| std::printf(" (later stages not comparable: different frame counts)\n"); | |
| } | |
| } else { | |
| compare("encoder", read_bin<float>(dir + "/encoder.bin"), dbg.encoder); | |
| compare("duration_raw", read_bin<float>(dir + "/duration_raw.bin"), dbg.duration_raw); | |
| } | |
| if (!flow && mismatch == 0) { | |
| compare("prosody_hidden", read_bin<float>(dir + "/prosody_hidden.bin"), dbg.prosody_hidden); | |
| compare("prosody_raw", read_bin<float>(dir + "/prosody_raw.bin"), dbg.prosody_raw); | |
| compare("prosody", read_bin<float>(dir + "/prosody.bin"), dbg.prosody); | |
| compare("acoustic", read_bin<float>(dir + "/acoustic.bin"), dbg.acoustic); | |
| compare("harmonic_head", read_bin<float>(dir + "/harmonic_head.bin"), dbg.harmonic_head); | |
| compare("noise_head", read_bin<float>(dir + "/noise_head.bin"), dbg.noise_head); | |
| compare("f0", read_bin<float>(dir + "/f0.bin"), dbg.f0); | |
| compare("wave", read_bin<float>(dir + "/wave.bin"), wave); | |
| } else if (!flow) { | |
| std::printf(" (later stages not comparable: different frame counts)\n"); | |
| } | |
| PROCESS_MEMORY_COUNTERS pmc{}; | |
| if (GetProcessMemoryInfo(GetCurrentProcess(), &pmc, sizeof(pmc))) | |
| std::printf(" RAM: peak working set %.1f MB (weights %.1f MB)\n", pmc.PeakWorkingSetSize / 1048576.0, | |
| model->weight_bytes() / 1048576.0); | |
| return 0; | |
| } | |