mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Merge with master, add yolov4tiny_512

Browse Source 
Signed-off-by: tk <micaelaverucchi@gmail.com>
This commit is contained in:
tk
2020-07-06 17:57:50 +02:00
parent 8d08f5aade 7c2155decf
commit a68f45cb4e
38 changed files with 1328 additions and 361 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/tkDNN/DarknetParser.h
+15 -260
View File
@@ -1,6 +1,6 @@
#pragma once
#include <iostream>
#include "tkdnn.h"
#include "tkDNN/tkdnn.h"
namespace tk { namespace dnn {
@@ -11,6 +11,7 @@ namespace tk { namespace dnn {
int channels = 3;
int batch_normalize=0;
int groups = 1;
int group_id = 0;
int filters=1;
int size_x=1;
int size_y=1;
@@ -27,267 +28,21 @@ namespace tk { namespace dnn {
std::vector<int> layers;
std::string activation = "linear";
friend std::ostream& operator<<(std::ostream& os, const darknetFields_t& f){
os << f.width << " " << f.height << " " << f.channels << " " << f.batch_normalize<< " " << f.filters << " " << f.activation<< " " << f.scale_xy;
return os;
}
};
std::ostream& operator<<(std::ostream& os, const darknetFields_t& f){
os << f.width << " " << f.height << " " << f.channels << " " << f.batch_normalize<< " " << f.filters << " " << f.activation<< " " << f.scale_xy;
return os;
}
std::string darknetParseType(const std::string& line){
size_t start = line.find("[");
size_t end = line.find("]");
if( start == std::string::npos || end == std::string::npos)
return "";
start++;
std::string type = line.substr(start, end-start);
return type;
}
bool divideNameAndValue(const std::string& line, std::string&name, std::string& value){
size_t sep = line.find("=");
if(sep == std::string::npos)
return false;
name = line.substr(0, sep);
value = line.substr(sep+1, line.size() - (sep+1));
return true;
}
std::vector<int> fromStringToIntVec(const std::string& line, const char delimiter){
std::stringstream linestream(line);
std::string value;
std::vector<int> values;
while(getline(linestream,value,delimiter))
values.push_back(std::stoi(value));
return values;
}
bool darknetParseFields(const std::string& line, darknetFields_t& fields){
std::string name,value;
if(!divideNameAndValue(line, name, value))
return false;
if(name.find("width") != std::string::npos)
fields.width = std::stoi(value);
else if(name.find("height") != std::string::npos)
fields.height = std::stoi(value);
else if(name.find("channels") != std::string::npos)
fields.channels = std::stoi(value);
else if(name.find("batch_normalize") != std::string::npos)
fields.batch_normalize = std::stoi(value);
else if(name.find("filters") != std::string::npos)
fields.filters = std::stoi(value);
else if(name.find("activation") != std::string::npos)
fields.activation = value;
else if(name.find("size") != std::string::npos){
fields.size_x = std::stoi(value);
fields.size_y = std::stoi(value);
}
else if(name.find("size_x") != std::string::npos)
fields.size_x = std::stoi(value);
else if(name.find("size_y") != std::string::npos)
fields.size_y = std::stoi(value);
else if(name.find("stride") != std::string::npos){
fields.stride_x = std::stoi(value);
fields.stride_y = std::stoi(value);
}
else if(name.find("stride_x") != std::string::npos)
fields.stride_x = std::stoi(value);
else if(name.find("stride_y") != std::string::npos)
fields.stride_y = std::stoi(value);
else if(name.find("pad") != std::string::npos)
fields.pad = std::stoi(value);
else if(name.find("classes") != std::string::npos)
fields.classes = std::stoi(value);
else if(name.find("num") != std::string::npos)
fields.num = std::stoi(value);
else if(name.find("coords") != std::string::npos)
fields.coords = std::stoi(value);
else if(name.find("groups") != std::string::npos)
fields.groups = std::stoi(value);
else if(name.find("scale_x_y") != std::string::npos)
fields.scale_xy = std::stof(value);
else if(name.find("from") != std::string::npos)
fields.layers.push_back(std::stof(value));
else if(name.find("mask") != std::string::npos){
auto vec = fromStringToIntVec(value, ',');
fields.n_mask = vec.size();
}
else if(name.find("layers") != std::string::npos)
fields.layers = fromStringToIntVec(value, ',');
else
std::cout<<"Not supported field: "<<line<<std::endl;
return true;
}
tk::dnn::Network *darknetAddNet(darknetFields_t &fields) {
//std::cout<<"Add Net: "<<fields.type<<"\n";
dataDim_t dim(1, fields.channels, fields.height, fields.width);
return new tk::dnn::Network(dim);
}
void darknetAddLayer(tk::dnn::Network *net, darknetFields_t &f, std::string wgs_path, std::vector<tk::dnn::Layer*> &netLayers, const std::vector<std::string>& names) {
if(net == nullptr)
FatalError("Cant add a layer without a Net\n");
// padding compute
if(f.pad == 1) {
f.padding_x = f.padding_y = f.size_x /2;
}
//std::cout<<"Add layer: "<<f.type<<"\n";
if(f.type == "convolutional") {
std::string wgs = wgs_path + "/c" + std::to_string(netLayers.size()) + ".bin";
//printf("%d (%d,%d) (%d,%d) (%d,%d) %s %d %d\n", f.filters, f.size_x, f.size_y, f.stride_x, f.stride_y, f.padding_x, f.padding_y, wgs.c_str(), f.batch_normalize, f.groups);
tk::dnn::Conv2d *l= new tk::dnn::Conv2d(net, f.filters, f.size_x, f.size_y, f.stride_x,
f.stride_y, f.padding_x, f.padding_y, wgs, f.batch_normalize, false, f.groups);
netLayers.push_back(l);
} else if(f.type == "maxpool") {
if(f.stride_x == 1 && f.stride_y == 1)
netLayers.push_back(new tk::dnn::Pooling(net, f.size_x, f.size_y, f.stride_x, f.stride_y,
f.padding_x, f.padding_y, tk::dnn::POOLING_MAX_FIXEDSIZE));
else
netLayers.push_back(new tk::dnn::Pooling(net, f.size_x, f.size_y, f.stride_x, f.stride_y,
f.padding_x, f.padding_y, tk::dnn::POOLING_MAX));
} else if(f.type == "avgpool") {
netLayers.push_back(new tk::dnn::Pooling(net, f.size_x, f.size_y, f.stride_x, f.stride_y,
f.padding_x, f.padding_y, tk::dnn::POOLING_AVERAGE));
} else if(f.type == "shortcut") {
if(f.layers.size() != 1) FatalError("no layers to shortcut\n");
int layerIdx = f.layers[0];
if(layerIdx < 0)
layerIdx = netLayers.size() + layerIdx;
if(layerIdx < 0 || layerIdx >= netLayers.size()) FatalError("impossible to shortcut\n");
//std::cout<<"shortcut to "<<layerIdx<<" "<<netLayers[layerIdx]->getLayerName()<<"\n";
netLayers.push_back(new tk::dnn::Shortcut(net, netLayers[layerIdx]));
} else if(f.type == "upsample") {
netLayers.push_back(new tk::dnn::Upsample(net, f.stride_x));
} else if(f.type == "route") {
if(f.layers.size() == 0) FatalError("no layers to Route\n");
std::vector<tk::dnn::Layer*> layers;
for(int i=0; i<f.layers.size(); i++) {
int layerIdx = f.layers[i];
if(layerIdx < 0)
layerIdx = netLayers.size() + layerIdx;
if(layerIdx < 0 || layerIdx >= netLayers.size()) FatalError("impossible to route\n");
//std::cout<<"Route to "<<layerIdx<<" "<<netLayers[layerIdx]->getLayerName()<<"\n";
layers.push_back(netLayers[layerIdx]);
}
netLayers.push_back(new tk::dnn::Route(net, layers.data(), layers.size()));
} else if(f.type == "reorg") {
netLayers.push_back(new tk::dnn::Reorg(net, f.stride_x));
} else if(f.type == "region") {
netLayers.push_back(new tk::dnn::Region(net, f.classes, f.coords, f.num));
} else if(f.type == "yolo") {
std::string wgs = wgs_path + "/g" + std::to_string(netLayers.size()) + ".bin";
//printf("%d %d %s %d %f\n", f.classes, f.num/f.n_mask, wgs.c_str(), f.n_mask, f.scale_xy);
tk::dnn::Yolo *l = new tk::dnn::Yolo(net, f.classes, f.num/f.n_mask, wgs, f.n_mask, f.scale_xy);
if(names.size() != f.classes)
FatalError("Mismatch between number of classes and names");
l->classesNames = names;
netLayers.push_back(l);
} else{
FatalError("layer not supported: " + f.type);
}
// add activation
if(netLayers.size() > 0 && f.activation != "linear") {
tkdnnActivationMode_t act;
if(f.activation == "relu") act = tkdnnActivationMode_t(CUDNN_ACTIVATION_RELU);
else if(f.activation == "leaky") act = tk::dnn::ACTIVATION_LEAKY;
else if(f.activation == "mish") act = tk::dnn::ACTIVATION_MISH;
else { FatalError("activation not supported: " + f.activation); }
netLayers[netLayers.size()-1] = new tk::dnn::Activation(net, act);
};
}
std::vector<std::string> darknetReadNames(const std::string& names_file){
std::ifstream if_names(names_file);
if(!if_names.is_open())
FatalError("cloud not open names file: " + names_file);
std::vector<std::string> names;
std::string line;
while(std::getline(if_names, line))
if(line != "")
names.push_back(line);
if_names.close();
return names;
}
tk::dnn::Network* darknetParser(const std::string& cfg_file, const std::string& wgs_path, const std::string& names_file) {
tk::dnn::Network *net = nullptr;
// layers without activations to retrive correct id number
std::vector<tk::dnn::Layer*> netLayers;
std::ifstream if_cfg(cfg_file);
if(!if_cfg.is_open())
FatalError("cloud not open cfg file: " + cfg_file);
std::vector<std::string> names = darknetReadNames(names_file);
darknetFields_t fields; // will be filled with layers fields
std::string line;
while(std::getline(if_cfg, line)) {
// remove comments
std::size_t found = line.find("#");
if ( found != std::string::npos ) {
line = line.substr(0, found);
}
// skip empty lines
if(line.size() == 0)
continue;
std::string type = darknetParseType(line);
if(type.size() > 0) {
// end of filled type
if(fields.type != "") {
if(fields.type == "net")
net = darknetAddNet(fields);
else
darknetAddLayer(net, fields, wgs_path, netLayers, names);
}
// new type
//std::cout<<"type: "<<type<<"\n";
fields = darknetFields_t(); // reset to default
fields.type = type;
continue;
}
if(darknetParseFields(line, fields)) {
// already parsed do nothing
} else {
FatalError("could not parse line: " + line);
}
}
// end of filled type
if(fields.type != "") {
darknetAddLayer(net, fields, wgs_path, netLayers, names);
}
if(net == nullptr) {
FatalError("net not found\n");
}
return net;
}
std::string darknetParseType(const std::string& line);
bool divideNameAndValue(const std::string& line, std::string&name, std::string& value);
std::vector<int> fromStringToIntVec(const std::string& line, const char delimiter);
bool darknetParseFields(const std::string& line, darknetFields_t& fields);
tk::dnn::Network *darknetAddNet(darknetFields_t &fields);
void darknetAddLayer(tk::dnn::Network *net, darknetFields_t &f, std::string wgs_path,
std::vector<tk::dnn::Layer*> &netLayers, const std::vector<std::string>& names);
std::vector<std::string> darknetReadNames(const std::string& names_file);
tk::dnn::Network* darknetParser(const std::string& cfg_file, const std::string& wgs_path, const std::string& names_file);
}}
include/tkDNN/Layer.h
+3 -1
View File
@@ -509,7 +509,7 @@ public:
class Route : public Layer {
public:
Route(Network *net, Layer **layers, int layers_n);
Route(Network *net, Layer **layers, int layers_n, int groups = 1, int group_id = 0);
virtual ~Route();
virtual layerType_t getLayerType() { return LAYER_ROUTE; };
@@ -519,6 +519,8 @@ public:
static const int MAX_LAYERS = 32;
Layer *layers[MAX_LAYERS]; //ids of layers to be merged
int layers_n; //number of layers
int groups;
int group_id;
};
include/tkDNN/NetworkRT.h
+1 -1
View File
@@ -28,7 +28,7 @@ using namespace nvinfer1;
#include "pluginsRT/ActivationMishRT.h"
#include "pluginsRT/ReorgRT.h"
#include "pluginsRT/RegionRT.h"
//#include "pluginsRT/RouteRT.h"
#include "pluginsRT/RouteRT.h"
#include "pluginsRT/ShortcutRT.h"
#include "pluginsRT/YoloRT.h"
#include "pluginsRT/UpsampleRT.h"
include/tkDNN/NetworkViz.h
+12
View File
@@ -0,0 +1,12 @@
#pragma once
#include <iostream>
#include <opencv2/core/types.hpp>
#include "tkdnn.h"
namespace tk { namespace dnn {
cv::Mat vizFloat2colorMap(cv::Mat map);
cv::Mat vizData2Mat(dnnType *dataInput, tk::dnn::dataDim_t dim, int imgdim);
cv::Mat vizLayer2Mat(tk::dnn::Network *net, int layer, int imgdim = 1000);
}}
include/tkDNN/pluginsRT/RouteRT.h
+19 -10
View File
@@ -8,7 +8,9 @@ class RouteRT : public IPlugin {
*/
public:
RouteRT() {
RouteRT(int groups, int group_id) {
this->groups = groups;
this->group_id = group_id;
}
~RouteRT(){
@@ -22,7 +24,7 @@ public:
Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override {
int out_c = 0;
for(int i=0; i<nbInputDims; i++) out_c += inputs[i].d[0];
return DimsCHW{out_c, inputs[0].d[1], inputs[0].d[2]};
return DimsCHW{out_c/groups, inputs[0].d[1], inputs[0].d[2]};
}
void configure(const Dims* inputDims, int nbInputs, const Dims* outputDims, int nbOutputs, int maxBatchSize) override {
@@ -34,6 +36,7 @@ public:
}
h = inputDims[0].d[1];
w = inputDims[0].d[2];
c /= groups;
}
int initialize() override {
@@ -49,15 +52,18 @@ public:
}
virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override {
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
int offset = 0;
for(int i=0; i<in; i++) {
dnnType *input = (dnnType*)reinterpret_cast<const dnnType*>(inputs[i]);
int in_dim = c_in[i]*h*w;
checkCuda( cudaMemcpyAsync(dstData + offset, input, in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) );
offset += in_dim;
for(int b=0; b<batchSize; b++) {
int offset = 0;
for(int i=0; i<in; i++) {
dnnType *input = (dnnType*)reinterpret_cast<const dnnType*>(inputs[i]);
int in_dim = c_in[i]*h*w;
int part_in_dim = in_dim / this->groups;
checkCuda( cudaMemcpyAsync(dstData + b*c*w*h + offset, input + b*c*w*h*groups + this->group_id*part_in_dim, part_in_dim*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream) );
offset += part_in_dim;
}
}
return 0;
@@ -65,11 +71,13 @@ public:
virtual size_t getSerializationSize() override {
return (4+MAX_INPUTS)*sizeof(int);
return (6+MAX_INPUTS)*sizeof(int);
}
virtual void serialize(void* buffer) override {
char *buf = reinterpret_cast<char*>(buffer);
tk::dnn::writeBUF(buf, groups);
tk::dnn::writeBUF(buf, group_id);
tk::dnn::writeBUF(buf, in);
for(int i=0; i<MAX_INPUTS; i++)
tk::dnn::writeBUF(buf, c_in[i]);
@@ -83,4 +91,5 @@ public:
int in;
int c_in[MAX_INPUTS];
int c, h, w;
int groups, group_id;
};
include/tkDNN/test.h
+8 -4
View File
@@ -1,7 +1,7 @@
#include <tkdnn.h>
int testInference(std::vector<std::string> input_bins, std::vector<std::string> output_bins,
tk::dnn::Network *net, tk::dnn::NetworkRT *netRT = nullptr) {
tk::dnn::Network *net, tk::dnn::NetworkRT *netRT = nullptr) {
std::vector<tk::dnn::Layer*> outputs;
for(int i=0; i<net->num_layers; i++) {
@@ -67,7 +67,11 @@ int testInference(std::vector<std::string> input_bins, std::vector<std::string>
std::cout<<"CUDNN vs TRT ";
ret_cudnn_tensorrt |= checkResult(odim, cudnn_out[i], rt_out[i]) == 0 ? 0 : ERROR_CUDNNvsTENSORRT;
}
}
return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
}
delete [] out_h;
checkCuda( cudaFree(out) );
}
delete [] input_h;
checkCuda( cudaFree(data) );
return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
}
include/tkDNN/utils.h
+4
View File
@@ -118,4 +118,8 @@ void matrixMulAdd( cublasHandle_t handle, dnnType* srcData, dnnType* dstData,
void getMemUsage(double& vm_usage_kb, double& resident_set_kb);
void printCudaMemUsage();
void removePathAndExtension(const std::string &full_string, std::string &name);
static inline bool isCudaPointer(void *data) {
cudaPointerAttributes attr;
return cudaPointerGetAttributes(&attr, data) == 0;
}
#endif //UTILS_H