Function effective_weight_norm_conv1d

src/framework/modules/hubert_encoder.cpp:127–157

    const std::string & prefix,
    int64_t out_channels,
    int64_t in_channels,
    int64_t kernel_size) {
    const auto g = source.require_f32(prefix + ".weight_g", {1, 1, kernel_size});
    const auto v = source.require_f32(prefix + ".weight_v", {out_channels, in_channels, kernel_size});
    std::vector<float> weight(v.size());
    for (int64_t k = 0; k < kernel_size; ++k) {
        double sum = 0.0;
        for (int64_t out = 0; out < out_channels; ++out) {
            for (int64_t in = 0; in < in_channels; ++in) {
                const size_t index = static_cast<size_t>((out * in_channels + in) * kernel_size + k);
                sum += static_cast<double>(v[index]) * static_cast<double>(v[index]);
            }
        }
        const double norm = std::sqrt(sum);
        if (norm == 0.0) {
            throw std::runtime_error("HuBERT positional-conv weight norm is zero");
        }
        const float scale_value = static_cast<float>(static_cast<double>(g[static_cast<size_t>(k)]) / norm);
        for (int64_t out = 0; out < out_channels; ++out) {
            for (int64_t in = 0; in < in_channels; ++in) {
                const size_t index = static_cast<size_t>((out * in_channels + in) * kernel_size + k);
                weight[index] = v[index] * scale_value;
            }
        }
    }
    return weight;
}

core::TensorValue grouped_pos_conv(
    core::ModuleBuildContext & ctx,
    const core::TensorValue & input_bct,
    const Conv1dWeights & weights,
    const HubertEncoderConfig & config) {
    const int64_t groups = config.num_conv_pos_embedding_groups;