Function sparsify

samples/common/sampleUtils.cpp:329–355

    IFullyConnectedLayer& l, int32_t k, int32_t trs, std::vector<int8_t>& sparseWeights);

void sparsify(nvinfer1::INetworkDefinition& network, std::vector<std::vector<int8_t>>& sparseWeights)
{
    for (int32_t l = 0; l < network.getNbLayers(); ++l)
    {
        auto* layer = network.getLayer(l);
        auto const t = layer->getType();
        if (t == nvinfer1::LayerType::kCONVOLUTION)
        {
            auto& conv = *static_cast<IConvolutionLayer*>(layer);
            auto const& dims = conv.getKernelSizeNd();
            ASSERT(dims.nbDims == 2 || dims.nbDims == 3);
            auto const k = conv.getNbOutputMaps();
            auto const trs = std::accumulate(dims.d, dims.d + dims.nbDims, 1, std::multiplies<int32_t>());
            sparseWeights.emplace_back();
            setSparseWeights(conv, k, trs, sparseWeights.back());
        }
        else if (t == nvinfer1::LayerType::kFULLY_CONNECTED)
        {
            auto& fc = *static_cast<nvinfer1::IFullyConnectedLayer*>(layer);
            auto const k = fc.getNbOutputChannels();
            sparseWeights.emplace_back();
            setSparseWeights(fc, k, 1, sparseWeights.back());
        }
    }

    sparsifyMatMulKernelWeights(network, sparseWeights);
}

void sparsify(Weights const& weights, int32_t k, int32_t trs, std::vector<int8_t>& sparseWeights)
{