void extractBinaryLayerParams(const caffe::LayerParameter& layer, LayerParams& layerParams)
{
const std::string &name = layer.name();
int li;
for (li = 0; li != netBinary.layer_size(); li++)
{
const caffe::LayerParameter& binLayer = netBinary.layer(li);
// Break if the layer name is the same and the blobs are not cleared
if (binLayer.name() == name && binLayer.blobs_size() != 0)
break;
}
if (li == netBinary.layer_size())
return;
caffe::LayerParameter* binLayer = netBinary.mutable_layer(li);
const int numBlobs = binLayer->blobs_size();
layerParams.blobs.resize(numBlobs);
for (int bi = 0; bi < numBlobs; bi++)
{
blobFromProto(binLayer->blobs(bi), layerParams.blobs[bi]);
}
binLayer->clear_blobs();
CV_Assert(numBlobs == binLayer->blobs().ClearedCount());
for (int bi = 0; bi < numBlobs; bi++)
{
delete binLayer->mutable_blobs()->ReleaseCleared();
}
}
void populateNet(Net dstNet)
{
CV_TRACE_FUNCTION();
int layersSize = net.layer_size();
layerCounter.clear();
addedBlobs.clear();
addedBlobs.reserve(layersSize + 1);
//setup input layer names
std::vector<String> netInputs(net.input_size());
{
for (int inNum = 0; inNum < net.input_size(); inNum++)
{
addedBlobs.push_back(BlobNote(net.input(inNum), 0, inNum));
netInputs[inNum] = net.input(inNum);
}
}
for (int li = 0; li < layersSize; li++)
{
const caffe::LayerParameter &layer = net.layer(li);
String name = layer.name();
String type = layer.type();
LayerParams layerParams;
extractLayerParams(layer, layerParams);
extractBinaryLayerParms(layer, layerParams);
int repetitions = layerCounter[name]++;
if (repetitions)
name += String("_") + toString(repetitions);
if (type == "Input")
{
for (int outNum = 0; outNum < layer.top_size(); outNum++)
{
addOutput(layer, 0, outNum);
addedBlobs.back().outNum = netInputs.size();
netInputs.push_back(addedBlobs.back().name);
}
continue;
}
int id = dstNet.addLayer(name, type, layerParams);
for (int inNum = 0; inNum < layer.bottom_size(); inNum++)
addInput(layer.bottom(inNum), id, inNum, dstNet);
for (int outNum = 0; outNum < layer.top_size(); outNum++)
addOutput(layer, id, outNum);
}
dstNet.setInputsNames(netInputs);
addedBlobs.clear();
}
void extractBinaryLayerParms(const caffe::LayerParameter& layer, LayerParams& layerParams)
{
const std::string &name = layer.name();
int li;
for (li = 0; li != netBinary.layer_size(); li++)
{
if (netBinary.layer(li).name() == name)
break;
}
if (li == netBinary.layer_size() || netBinary.layer(li).blobs_size() == 0)
return;
const caffe::LayerParameter &binLayer = netBinary.layer(li);
layerParams.blobs.resize(binLayer.blobs_size());
for (int bi = 0; bi < binLayer.blobs_size(); bi++)
{
blobFromProto(binLayer.blobs(bi), layerParams.blobs[bi]);
}
}
void populateNet(Net dstNet)
{
CV_TRACE_FUNCTION();
int layersSize = net.layer_size();
layerCounter.clear();
addedBlobs.clear();
addedBlobs.reserve(layersSize + 1);
//setup input layer names
std::vector<String> netInputs(net.input_size());
{
for (int inNum = 0; inNum < net.input_size(); inNum++)
{
addedBlobs.push_back(BlobNote(net.input(inNum), 0, inNum));
netInputs[inNum] = net.input(inNum);
}
}
for (int li = 0; li < layersSize; li++)
{
const caffe::LayerParameter &layer = net.layer(li);
String name = layer.name();
String type = layer.type();
LayerParams layerParams;
extractLayerParams(layer, layerParams);
extractBinaryLayerParams(layer, layerParams);
int repetitions = layerCounter[name]++;
if (repetitions)
name += String("_") + toString(repetitions);
if (type == "Input")
{
for (int outNum = 0; outNum < layer.top_size(); outNum++)
{
addOutput(layer, 0, outNum);
addedBlobs.back().outNum = netInputs.size();
netInputs.push_back(addedBlobs.back().name);
}
continue;
}
else if (type == "BatchNorm")
{
if (!layerParams.get<bool>("use_global_stats", true))
{
CV_Assert_N(layer.bottom_size() == 1, layer.top_size() == 1);
LayerParams mvnParams;
mvnParams.set("eps", layerParams.get<float>("eps", 1e-5));
std::string mvnName = name + "/mvn";
int repetitions = layerCounter[mvnName]++;
if (repetitions)
mvnName += String("_") + toString(repetitions);
int mvnId = dstNet.addLayer(mvnName, "MVN", mvnParams);
addInput(layer.bottom(0), mvnId, 0, dstNet);
addOutput(layer, mvnId, 0);
net.mutable_layer(li)->set_bottom(0, layer.top(0));
layerParams.blobs[0].setTo(0); // mean
layerParams.blobs[1].setTo(1); // std
}
}
else if ("ConvolutionDepthwise" == type)
{
type = "Convolution";
}
int id = dstNet.addLayer(name, type, layerParams);
for (int inNum = 0; inNum < layer.bottom_size(); inNum++)
addInput(layer.bottom(inNum), id, inNum, dstNet);
for (int outNum = 0; outNum < layer.top_size(); outNum++)
addOutput(layer, id, outNum);
}
dstNet.setInputsNames(netInputs);
addedBlobs.clear();
}
void convertProtoToLuaV2(const caffe::NetParameter &netparam, const char* lua_name, const char* cuda_package)
{
PACKAGE_TYPE cuda_package_type = CCN2;
if(std::string(cuda_package) == "ccn2")
cuda_package_type = CCN2;
else if(std::string(cuda_package) == "nn")
cuda_package_type = NN;
else if(std::string(cuda_package) == "cudnn")
cuda_package_type = CUDNN;
std::ofstream ofs (lua_name);
ofs << "require '" << cuda_package << "'\n";
ofs << "require 'cunn'\n";
ofs << "local model = {}\n";
if(std::string(cuda_package)=="ccn2")
ofs<< "table.insert(model, {'torch_transpose_dwhb', nn.Transpose({1,4},{1,3},{1,2})})\n";
else if(std::string(cuda_package)=="nn" || std::string(cuda_package)=="cudnn")
ofs<< "require 'inn'\n";
int num_output = netparam.input_shape_size() * 4;
for (int i=0; i<netparam.layer_size(); ++i)
{
std::vector<std::pair<std::string, std::string>> lines;
auto& layer = netparam.layer(i);
if(layer.type() == "Convolution")
{
auto ¶m = layer.convolution_param();
int groups = param.group() == 0 ? 1 : param.group();
int nInputPlane = layer.blobs(0).shape().dim(1)*groups;
int nOutputPlane = layer.blobs(0).shape().dim(0);
//int nOutputPlane = param.num_output();
num_output = nOutputPlane;
int kW = param.kernel_w();
int kH = param.kernel_h();
int dW = param.stride_w();
int dH = param.stride_h();
if(kW==0 || kH==0)
{
kW = param.kernel_size();
kH = kW;
}
if(dW==0 || dH==0)
{
dW = param.stride();
dH = dW;
}
int pad_w = param.pad_w();
int pad_h = param.pad_h();
if(pad_w==0 || pad_h==0)
{
pad_w = param.pad();
pad_h = pad_w;
}
if(cuda_package_type == CCN2)
{
if(kW != kH || dW != dH || pad_w != pad_h)
{
std::cout << "ccn2 only supports square images!\n";
break;
}
char buf[1024];
sprintf(buf, "ccn2.SpatialConvolution(%d, %d, %d, %d, %d, %d)",
nInputPlane, nOutputPlane, kW, dW, pad_w, groups);
lines.emplace_back(layer.name(), buf);
}
else if(cuda_package_type == NN)
{
if(groups != 1)
{
std::cout << "nn supports no groups!\n";
break;
}
char buf[1024];
sprintf(buf, "nn.SpatialConvolutionMM(%d, %d, %d, %d, %d, %d, %d, %d)",
nInputPlane, nOutputPlane, kW, kH, dW, dH, pad_w, pad_h);
lines.emplace_back(layer.name(), buf);
}
else
{
char buf[1024];
sprintf(buf, "cudnn.SpatialConvolution(%d, %d, %d, %d, %d, %d, %d, %d, %d)",
nInputPlane, nOutputPlane, kW, kH, dW, dH, pad_w, pad_h, groups);
lines.emplace_back(layer.name(), buf);
}
}
if(layer.type() == "Pooling")
{
auto ¶m = layer.pooling_param();
std::string ptype = param.pool() == caffe::PoolingParameter::MAX ? "Max" : "Avg";
int kW = param.kernel_w();
int kH = param.kernel_h();
int dW = param.stride_w();
int dH = param.stride_h();
int padW = param.pad_w();
int padH = param.pad_h();
if(kW==0 || kH==0)
{
kW = param.kernel_size();
kH = kW;
}
if(dW==0 || dH==0)
{
dW = param.stride();
dH = dW;
}
char buf[1024];
switch(cuda_package_type)
{
case CCN2: // ceil mode by default
if(param.pool() == caffe::PoolingParameter::MAX)
sprintf(buf, "ccn2.SpatialMaxPooling(%d, %d)", kW, dW);
else if(param.pool() == caffe::PoolingParameter::AVE)
sprintf(buf, "ccn2.SpatialAvgPooling(%d, %d)", kW, dW);
else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
THError("Stochastic pooling is not implemented in DHWB format");
break;
case CUDNN:
if(param.pool() == caffe::PoolingParameter::MAX)
sprintf(buf, "cudnn.SpatialMaxPooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
else if(param.pool() == caffe::PoolingParameter::AVE)
sprintf(buf, "cudnn.SpatialAveragePooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
sprintf(buf, "inn.SpatialStochasticPooling(%d, %d, %d, %d)", kW, kH, dW, dH);
break;
case NN:
if(param.pool() == caffe::PoolingParameter::MAX)
sprintf(buf, "nn.SpatialMaxPooling(%d, %d, %d, %d, %d, %d):ceil()", kW, kH, dW, dH, padW, padH);
else if(param.pool() == caffe::PoolingParameter::AVE)
sprintf(buf, "inn.SpatialAveragePooling(%d, %d, %d, %d)", kW, kH, dW, dH); // padding is not supported yet
else if(param.pool() == caffe::PoolingParameter::STOCHASTIC)
sprintf(buf, "inn.SpatialStochasticPooling(%d, %d, %d, %d)", kW, kH, dW, dH);
break;
}
lines.emplace_back(layer.name(), buf);
}
if(layer.type() == "ReLU")
{
if(cuda_package_type == CUDNN)
lines.emplace_back(layer.name(), "cudnn.ReLU(true)");
else
lines.emplace_back(layer.name(), "nn.ReLU(true)");
}
if(layer.type() == "Sigmoid")
{
if(cuda_package_type == CUDNN)
lines.emplace_back(layer.name(), "cudnn.Sigmoid(true)");
else
lines.emplace_back(layer.name(), "nn.Sigmoid()");
}
if(layer.type() == "Tanh")
{
if(cuda_package_type == CUDNN)
lines.emplace_back(layer.name(), "cudnn.Tanh(true)");
else
lines.emplace_back(layer.name(), "nn.Tanh()");
}
if(layer.type() == "LRN")
{
auto ¶m = layer.lrn_param();
int local_size = param.local_size();
float alpha = param.alpha();
float beta = param.beta();
float k = param.k();
char buf[1024];
if(std::string(cuda_package) == "ccn2")
sprintf(buf, "ccn2.SpatialCrossResponseNormalization(%d, %.6f, %.4f, %f)", local_size, alpha, beta, k);
else
sprintf(buf, "inn.SpatialCrossResponseNormalization(%d, %.6f, %.4f, %f)", local_size, alpha, beta, k);
lines.emplace_back(layer.name(), buf);
}
if(layer.type() == "InnerProduct")
{
auto ¶m = layer.inner_product_param();
int nInputPlane = layer.blobs(0).shape().dim(1);
int nOutputPlane = param.num_output();
char buf[1024];
sprintf(buf, "nn.Linear(%d, %d)", nInputPlane, nOutputPlane);
if(num_output != nInputPlane)
{
if(std::string(cuda_package) == "ccn2")
lines.emplace_back("torch_transpose_bdwh", "nn.Transpose({4,1},{4,2},{4,3})");
lines.emplace_back("torch_view", "nn.View(-1):setNumInputDims(3)");
}
lines.emplace_back(layer.name(), buf);
num_output = nOutputPlane;
}
if(layer.type() == "Dropout")
{
char buf[1024];
sprintf(buf, "nn.Dropout(%f)", layer.dropout_param().dropout_ratio());
lines.emplace_back(layer.name(), buf);
}
if(layer.type()=="SoftmaxWithLoss")
{
lines.emplace_back(layer.name(), "nn.SoftMax()");
}
if(layer.type()=="Softmax")
{
lines.emplace_back(layer.name(), "nn.SoftMax()");
}
if(!lines.empty())
for(auto& it: lines)
ofs << "table.insert(model, {'" << it.first << "', " << it.second << "})\n";
else
{
ofs << "-- warning: module '" << layer.name() << "' not found\n";
std::cout << "warning: module '" << layer.name() << "' not found\n";
}
}
ofs << "return model";
}