Method prepare_chunk

src/models/demucs/frontend.cpp:244–283

}

void HTDemucsFrontend::prepare_chunk(std::vector<float> & chunk_planar) {
    if (chunk_planar.empty() || static_cast<int64_t>(chunk_planar.size()) % config_.audio_channels != 0) {
        throw std::runtime_error("HTDemucs chunk size mismatch");
    }
    input_samples_ = static_cast<int64_t>(chunk_planar.size()) / config_.audio_channels;
    if (input_samples_ > config_.segment_samples) {
        throw std::runtime_error("HTDemucs chunk length exceeds training segment length");
    }
    if (input_samples_ != config_.segment_samples) {
        throw std::runtime_error("HTDemucs frontend expects session chunks padded to segment length");
    }
    time_input_ = &chunk_planar;

    pad1d_reflect_fast(
        padded_,
        *time_input_,
        config_.audio_channels,
        config_.segment_samples,
        pad_indices_);
    compute_stft_complex_normalized(
        stft_framed_,
        stft_spectrum_,
        padded_,
        stft_window_,
        config_.audio_channels,
        padded_samples_,
        stft_full_frames_,
        config_.n_fft,
        stft_frame_indices_,
        fft_threads_);
    const auto [freq_sum, freq_sumsq] = build_demucs_complex_input(
        freq_input_,
        stft_spectrum_,
        config_.audio_channels,
        config_.stft_freq_bins,
        config_.stft_frames,
        stft_full_frames_);
    std::tie(freq_mean_, freq_std_) = normalize_in_place_with_stats(freq_input_, freq_sum, freq_sumsq);
    std::tie(time_mean_, time_std_) = normalize_in_place(*time_input_);
}

const std::vector<float> & HTDemucsFrontend::freq_input() const noexcept { return freq_input_; }
const std::vector<float> & HTDemucsFrontend::time_input() const noexcept { return *time_input_; }