Function extractBatchNormTrainedData

software/xtract/caffe_network_parser.cpp:662–742

}

void extractBatchNormTrainedData(XGraph* graph, const string& layerName, const caffe::NetParameter* Net,
                                    const map<string, int>* layerIndex)
{
    map<string, int>::const_iterator modelLayer_it = layerIndex->find(layerName);
    if(modelLayer_it == layerIndex->end())
    {
        cerr << "[EP030] Layer " << layerName << " is not found in the caffemodel file. " << endl;
        exit(-1);
    }
    
    // Get the layer from caffemodel
    int loc = modelLayer_it->second;
    const caffe::LayerParameter& binLayer  = Net->layer(loc);
    XLayer* tmpXlayer               = graph->layers[layerName];
    string txtFileName;

    // Extract MWA
    cerr << "[IG001] Extracting " << tmpXlayer->name << " weights ... " << endl;
    vector<float> mwa = extractBlobToVector(binLayer.blobs(2));
    float scaling_factor = mwa[0] == 0 ? 0 : 1.0f/mwa[0];

    // Extract the mean, do trimming & rounding before saving to TXT file
    vector<int> weightsShape = getBlobDim(binLayer.blobs(0));
    int channels = tmpXlayer->topShape.at(0).at(1);               // Number of feature maps
    ASSERT( (weightsShape.at(0) == channels), EP056,
            "BatchNorm Layer: " << tmpXlayer->name << "mismatch in Mean shape : " 
            << TensorDimToString(weightsShape) << " != " << channels)
    vector<float> weights = extractBlobToVector(binLayer.blobs(0));
    for(int i=0; i<weights.size(); ++i)
        weights[i] *= scaling_factor;
    tmpXlayer->batchnorm_params->mean.push_back(weights);
    tmpXlayer->batchnorm_params->meanDim.push_back(weightsShape);

#if DEBUG_WEIGHT_EXTRACTION
    string tmpName(tmpXlayer->name); 
    replace(tmpName.begin(), tmpName.end(), '/', '_');
    txtFileName = graph->saveDir + tmpName + "_mean";
    tmpXlayer->batchnorm_params->meanPath.push_back(txtFileName);
    int sizeInBytes = getSize(weightsShape) * sizeof(float);
    cerr << "[IG001] Saving " << txtFileName << " (" << humanReadableSize(sizeInBytes) << ")" << "\t";
	if(sizeInBytes > (14*1024*1024))
	{
		cerr << "Parsing large data, this may take a while ...";
	}
	cerr << endl;
    SAVEDATA(weights, txtFileName);
#endif
    
    // Extract the variance, do trimming & rounding before saving to TXT file
    vector<int> biasShape = getBlobDim(binLayer.blobs(1));
    ASSERT( (biasShape.at(0) == channels), EP056,
            "BatchNorm Layer: " << tmpXlayer->name << "mismatch in variance shape : " 
            << TensorDimToString(biasShape) << " != " << channels)
    vector<float> bias = extractBlobToVector(binLayer.blobs(1));
    for(int i=0; i<bias.size(); ++i)
        bias[i] *= scaling_factor;
    tmpXlayer->batchnorm_params->variance.push_back(bias);
    tmpXlayer->batchnorm_params->varianceDim.push_back(biasShape);