Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-11-05 17:34:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

Split inception v3 source file

Browse Source
This commit is contained in:
Matúš Námešný
2024-05-30 18:41:33 +02:00
parent 8168f02f58
commit 229929a464
5 changed files with 1027 additions and 941 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
examples/inception_v3/CMakeLists.txt
View File

@@ -1,10 +1,13 @@
cmake_minimum_required(VERSION 3.17)
cmake_minimum_required(VERSION 3.20)
project(Inceptionv3
LANGUAGES CXX CUDA
)
add_executable(inception_v3 inception_v3.cpp)
add_executable(inception_v3
inception_v3.cpp
inception_modules.cpp
)
find_library(CUDANet_LIBRARY NAMES CUDANet HINTS ${CMAKE_CURRENT_SOURCE_DIR}/../../build)
find_package(OpenCV REQUIRED COMPONENTS core imgcodecs imgproc)
@@ -19,7 +22,7 @@ set (CUDANet_INCLUDE_DIRS
)
target_include_directories(
inception_v3 PRIVATE ${OpenCV_INCLUDE_DIRS} ${CUDANet_INCLUDE_DIRS} ${CUDAToolkit_INCLUDE_DIRS})
inception_v3 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR} ${OpenCV_INCLUDE_DIRS} ${CUDANet_INCLUDE_DIRS} ${CUDAToolkit_INCLUDE_DIRS})
target_link_libraries(inception_v3 PRIVATE ${CUDANet_LIBRARY} ${OpenCV_LIBS} CUDA::cudart)

764
examples/inception_v3/inception_modules.cpp Normal file
View File

@@ -0,0 +1,764 @@
#include "inception_v3.hpp"
#include <cudanet.cuh>
// Inception Basic Convolution 2D
BasicConv2d::BasicConv2d(
const shape2d inputShape,
const int inputChannels,
const int outputChannels,
const shape2d kernelSize,
const shape2d stride,
const shape2d padding,
const std::string &prefix
)
: outputChannels(outputChannels) {
// Create the convolution layer
conv = new CUDANet::Layers::Conv2d(
inputShape, inputChannels, kernelSize, stride, outputChannels, padding,
CUDANet::Layers::ActivationType::NONE
);
shape2d batchNormSize = conv->getOutputDims();
batchNorm = new CUDANet::Layers::BatchNorm2d(
batchNormSize, outputChannels, 1e-3f,
CUDANet::Layers::ActivationType::RELU
);
inputSize = inputShape.first * inputShape.second * inputChannels;
outputSize = batchNorm->getOutputDims().first *
batchNorm->getOutputDims().second * outputChannels;
addLayer(prefix + ".conv", conv);
addLayer(prefix + ".bn", batchNorm);
}
BasicConv2d::~BasicConv2d() {
delete conv;
delete batchNorm;
}
float *BasicConv2d::forward(const float *d_input) {
float *d_output = conv->forward(d_input);
return batchNorm->forward(d_output);
}
shape2d BasicConv2d::getOutputDims() {
return batchNorm->getOutputDims();
}
int BasicConv2d::getOutputChannels() {
return outputChannels;
}
// Inception Block A
InceptionA::InceptionA(
const shape2d inputShape,
const int inputChannels,
const int poolFeatures,
const std::string &prefix
)
: inputShape(inputShape),
inputChannels(inputChannels),
poolFeatures(poolFeatures) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 1x1
branch1x1 = new BasicConv2d(
inputShape, inputChannels, 64, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch1x1"
);
addLayer("", branch1x1);
// Branch 5x5
branch5x5_1 = new BasicConv2d(
inputShape, inputChannels, 48, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch5x5_1"
);
addLayer("", branch5x5_1);
branch5x5_2 = new BasicConv2d(
branch5x5_1->getOutputDims(), 48, 64, {5, 5}, {1, 1}, {2, 2},
prefix + ".branch5x5_2"
);
addLayer("", branch5x5_2);
// Branch 3x3
branch3x3dbl_1 = new BasicConv2d(
inputShape, inputChannels, 64, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3dbl_1"
);
addLayer("", branch3x3dbl_1);
branch3x3dbl_2 = new BasicConv2d(
branch3x3dbl_1->getOutputDims(), 64, 96, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_2"
);
addLayer("", branch3x3dbl_2);
branch3x3dbl_3 = new BasicConv2d(
branch3x3dbl_2->getOutputDims(), 96, 96, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_3"
);
addLayer("", branch3x3dbl_3);
// Branch Pool
branchPool_1 = new CUDANet::Layers::AvgPooling2d(
inputShape, inputChannels, {3, 3}, {1, 1}, {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool_1);
branchPool_2 = new BasicConv2d(
branchPool_1->getOutputDims(), inputChannels, poolFeatures, {1, 1},
{1, 1}, {0, 0}, prefix + ".branch_pool"
);
addLayer("", branchPool_2);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch1x1->getOutputSize(), branch5x5_2->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branch3x3dbl_3->getOutputSize()
);
concat_3 = new CUDANet::Layers::Concat(
concat_2->getOutputSize(), branchPool_2->getOutputSize()
);
outputSize = concat_3->getOutputSize();
}
InceptionA::~InceptionA() {
delete branch1x1;
delete branch5x5_1;
delete branch5x5_2;
delete branch3x3dbl_1;
delete branch3x3dbl_2;
delete branch3x3dbl_3;
delete branchPool_1;
delete branchPool_2;
delete concat_1;
delete concat_2;
delete concat_3;
}
float *InceptionA::forward(const float *d_input) {
float *d_branch1x1_out = branch1x1->forward(d_input);
float *d_branch5x5_out = branch5x5_1->forward(d_input);
d_branch5x5_out = branch5x5_2->forward(d_branch5x5_out);
float *d_branch3x3dbl_out = branch3x3dbl_1->forward(d_input);
d_branch3x3dbl_out = branch3x3dbl_2->forward(d_branch3x3dbl_out);
d_branch3x3dbl_out = branch3x3dbl_3->forward(d_branch3x3dbl_out);
float *d_branchPool_out = branchPool_1->forward(d_input);
d_branchPool_out = branchPool_2->forward(d_branchPool_out);
float *d_output = concat_1->forward(d_branch1x1_out, d_branch5x5_out);
d_output = concat_2->forward(d_output, d_branch3x3dbl_out);
d_output = concat_3->forward(d_output, d_branchPool_out);
return d_output;
}
shape2d InceptionA::getOutputDims() {
return branch1x1->getOutputDims();
}
int InceptionA::getOutputChannels() {
return branch1x1->getOutputChannels() + branch5x5_2->getOutputChannels() +
branch3x3dbl_3->getOutputChannels() +
branchPool_2->getOutputChannels();
}
// Inception Block B
InceptionB::InceptionB(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 3x3
branch3x3 = new BasicConv2d(
inputShape, inputChannels, 384, {3, 3}, {2, 2}, {0, 0},
prefix + ".branch3x3"
);
addLayer("", branch3x3);
// Branch 3x3dbl
branch3x3dbl_1 = new BasicConv2d(
inputShape, inputChannels, 64, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3dbl_1"
);
addLayer("", branch3x3dbl_1);
branch3x3dbl_2 = new BasicConv2d(
branch3x3dbl_1->getOutputDims(), 64, 96, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_2"
);
addLayer("", branch3x3dbl_2);
branch3x3dbl_3 = new BasicConv2d(
branch3x3dbl_2->getOutputDims(), 96, 96, {3, 3}, {2, 2}, {1, 1},
prefix + ".branch3x3dbl_3"
);
addLayer("", branch3x3dbl_3);
branchPool = new CUDANet::Layers::MaxPooling2d(
inputShape, inputChannels, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool);
concat_1 = new CUDANet::Layers::Concat(
branch3x3->getOutputSize(), branch3x3dbl_3->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branchPool->getOutputSize()
);
outputSize = concat_2->getOutputSize();
}
InceptionB::~InceptionB() {
delete branch3x3;
delete branch3x3dbl_1;
delete branch3x3dbl_2;
delete branch3x3dbl_3;
delete branchPool;
delete concat_1;
delete concat_2;
}
float *InceptionB::forward(const float *d_input) {
float *d_branch3x3_out = branch3x3->forward(d_input);
float *d_branch3x3dbl_out = branch3x3dbl_1->forward(d_input);
d_branch3x3dbl_out = branch3x3dbl_2->forward(d_branch3x3dbl_out);
d_branch3x3dbl_out = branch3x3dbl_3->forward(d_branch3x3dbl_out);
float *d_branchPool_out = branchPool->forward(d_input);
float *d_output = concat_1->forward(d_branch3x3_out, d_branch3x3dbl_out);
d_output = concat_2->forward(d_output, d_branchPool_out);
return d_output;
}
shape2d InceptionB::getOutputDims() {
return branch3x3->getOutputDims();
}
int InceptionB::getOutputChannels() {
return branch3x3->getOutputChannels() +
branch3x3dbl_3->getOutputChannels() + inputChannels;
}
// Inception Block C
InceptionC::InceptionC(
const shape2d inputShape,
const int inputChannels,
const int nChannels_7x7,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 1x1
branch1x1 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch1x1"
);
addLayer("", branch1x1);
// Branch 7x7
branch7x7_1 = new BasicConv2d(
inputShape, inputChannels, nChannels_7x7, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch7x7_1"
);
addLayer("", branch7x7_1);
branch7x7_2 = new BasicConv2d(
branch7x7_1->getOutputDims(), nChannels_7x7, nChannels_7x7, {1, 7},
{1, 1}, {0, 3}, prefix + ".branch7x7_2"
);
addLayer("", branch7x7_2);
branch7x7_3 = new BasicConv2d(
branch7x7_2->getOutputDims(), nChannels_7x7, 192, {7, 1}, {1, 1},
{3, 0}, prefix + ".branch7x7_3"
);
addLayer("", branch7x7_3);
// Branch 7x7dbl
branch7x7dbl_1 = new BasicConv2d(
inputShape, inputChannels, nChannels_7x7, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch7x7dbl_1"
);
addLayer("", branch7x7dbl_1);
branch7x7dbl_2 = new BasicConv2d(
branch7x7dbl_1->getOutputDims(), nChannels_7x7, nChannels_7x7, {7, 1},
{1, 1}, {3, 0}, prefix + ".branch7x7dbl_2"
);
addLayer("", branch7x7dbl_2);
branch7x7dbl_3 = new BasicConv2d(
branch7x7dbl_2->getOutputDims(), nChannels_7x7, nChannels_7x7, {1, 7},
{1, 1}, {0, 3}, prefix + ".branch7x7dbl_3"
);
addLayer("", branch7x7dbl_3);
branch7x7dbl_4 = new BasicConv2d(
branch7x7dbl_3->getOutputDims(), nChannels_7x7, nChannels_7x7, {7, 1},
{1, 1}, {3, 0}, prefix + ".branch7x7dbl_4"
);
addLayer("", branch7x7dbl_4);
branch7x7dbl_5 = new BasicConv2d(
branch7x7dbl_4->getOutputDims(), nChannels_7x7, 192, {1, 7}, {1, 1},
{0, 3}, prefix + ".branch7x7dbl_5"
);
addLayer("", branch7x7dbl_5);
// Branch Pool
branchPool_1 = new CUDANet::Layers::AvgPooling2d(
inputShape, inputChannels, {3, 3}, {1, 1}, {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool_1);
branchPool_2 = new BasicConv2d(
branchPool_1->getOutputDims(), inputChannels, 192, {1, 1}, {1, 1},
{0, 0}, prefix + ".branch_pool"
);
addLayer("", branchPool_2);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch1x1->getOutputSize(), branch7x7_3->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branch7x7dbl_5->getOutputSize()
);
concat_3 = new CUDANet::Layers::Concat(
concat_2->getOutputSize(), branchPool_2->getOutputSize()
);
outputSize = concat_3->getOutputSize();
}
InceptionC::~InceptionC() {
delete branch1x1;
delete branch7x7_1;
delete branch7x7_2;
delete branch7x7_3;
delete branch7x7dbl_1;
delete branch7x7dbl_2;
delete branch7x7dbl_3;
delete branch7x7dbl_4;
delete branch7x7dbl_5;
delete branchPool_1;
delete branchPool_2;
delete concat_1;
delete concat_2;
delete concat_3;
}
float *InceptionC::forward(const float *d_input) {
float *branch1x1_output = branch1x1->forward(d_input);
float *branch7x7_output = branch7x7_1->forward(d_input);
branch7x7_output = branch7x7_2->forward(branch7x7_output);
branch7x7_output = branch7x7_3->forward(branch7x7_output);
float *branch7x7dbl_output = branch7x7dbl_1->forward(d_input);
branch7x7dbl_output = branch7x7dbl_2->forward(branch7x7dbl_output);
branch7x7dbl_output = branch7x7dbl_3->forward(branch7x7dbl_output);
branch7x7dbl_output = branch7x7dbl_4->forward(branch7x7dbl_output);
branch7x7dbl_output = branch7x7dbl_5->forward(branch7x7dbl_output);
float *branchPool_output = branchPool_1->forward(d_input);
branchPool_output = branchPool_2->forward(branchPool_output);
float *d_output = concat_1->forward(branch1x1_output, branch7x7_output);
d_output = concat_2->forward(d_output, branch7x7dbl_output);
d_output = concat_3->forward(d_output, branchPool_output);
return d_output;
}
shape2d InceptionC::getOutputDims() {
return branch1x1->getOutputDims();
}
int InceptionC::getOutputChannels() {
return branch1x1->getOutputChannels() + branch7x7_3->getOutputChannels() +
branch7x7dbl_5->getOutputChannels() +
branchPool_2->getOutputChannels();
}
// Inception Block D
InceptionD::InceptionD(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 3x3
branch3x3_1 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3_1"
);
addLayer("", branch3x3_1);
branch3x3_2 = new BasicConv2d(
inputShape, 192, 320, {3, 3}, {2, 2}, {0, 0}, prefix + ".branch3x3_2"
);
addLayer("", branch3x3_2);
// Branch 7x7x3
branch7x7x3_1 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch7x7x3_1"
);
addLayer("", branch7x7x3_1);
branch7x7x3_2 = new BasicConv2d(
inputShape, 192, 192, {1, 7}, {1, 1}, {0, 3}, prefix + ".branch7x7x3_2"
);
addLayer("", branch7x7x3_2);
branch7x7x3_3 = new BasicConv2d(
inputShape, 192, 192, {7, 1}, {1, 1}, {3, 0}, prefix + ".branch7x7x3_3"
);
addLayer("", branch7x7x3_3);
branch7x7x3_4 = new BasicConv2d(
inputShape, 192, 192, {3, 3}, {2, 2}, {0, 0}, prefix + ".branch7x7x3_4"
);
addLayer("", branch7x7x3_4);
// Branch Pool
branchPool = new CUDANet::Layers::MaxPooling2d(
inputShape, 192, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch3x3_2->getOutputSize(), branch7x7x3_4->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branchPool->getOutputSize()
);
outputSize = concat_2->getOutputSize();
}
InceptionD::~InceptionD() {
delete branch3x3_1;
delete branch3x3_2;
delete branch7x7x3_1;
delete branch7x7x3_2;
delete branch7x7x3_3;
delete branch7x7x3_4;
delete branchPool;
delete concat_1;
delete concat_2;
}
float *InceptionD::forward(const float *d_input) {
float *branch3x3_output = branch3x3_1->forward(d_input);
branch3x3_output = branch3x3_2->forward(branch3x3_output);
float *branch7x7_output = branch7x7x3_1->forward(d_input);
branch7x7_output = branch7x7x3_2->forward(branch7x7_output);
branch7x7_output = branch7x7x3_3->forward(branch7x7_output);
branch7x7_output = branch7x7x3_4->forward(branch7x7_output);
float *branchPool_output = branchPool->forward(d_input);
float *d_output = concat_1->forward(branch3x3_output, branch7x7_output);
d_output = concat_2->forward(d_output, branchPool_output);
return d_output;
}
shape2d InceptionD::getOutputDims() {
return branch3x3_2->getOutputDims();
}
int InceptionD::getOutputChannels() {
return branch3x3_2->getOutputChannels() +
branch7x7x3_4->getOutputChannels() + inputChannels;
}
// Inception Block E
InceptionE::InceptionE(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 1x1
branch1x1 = new BasicConv2d(
inputShape, inputChannels, 320, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch1x1"
);
addLayer("", branch1x1);
// Branch 3x3
branch3x3_1 = new BasicConv2d(
inputShape, inputChannels, 384, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3_1"
);
addLayer("", branch3x3_1);
branch3x3_2a = new BasicConv2d(
inputShape, 384, 384, {1, 3}, {1, 1}, {0, 1}, prefix + ".branch3x3_2a"
);
addLayer("", branch3x3_2a);
branch3x3_2b = new BasicConv2d(
inputShape, 384, 384, {3, 1}, {1, 1}, {1, 0}, prefix + ".branch3x3_2b"
);
addLayer("", branch3x3_2b);
branch_3x3_2_concat = new CUDANet::Layers::Concat(
branch3x3_2a->getOutputSize(), branch3x3_2b->getOutputSize()
);
// Branch 3x3dbl
branch3x3dbl_1 = new BasicConv2d(
inputShape, inputChannels, 448, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3dbl_1"
);
addLayer("", branch3x3dbl_1);
branch3x3dbl_2 = new BasicConv2d(
inputShape, 448, 384, {3, 3}, {1, 1}, {1, 1}, prefix + ".branch3x3dbl_2"
);
addLayer("", branch3x3dbl_2);
branch3x3dbl_3a = new BasicConv2d(
inputShape, 384, 384, {1, 3}, {1, 1}, {0, 1},
prefix + ".branch3x3dbl_3a"
);
addLayer("", branch3x3dbl_3a);
branch3x3dbl_3b = new BasicConv2d(
inputShape, 384, 384, {3, 1}, {1, 1}, {1, 0},
prefix + ".branch3x3dbl_3b"
);
addLayer("", branch3x3dbl_3b);
branch_3x3dbl_3_concat = new CUDANet::Layers::Concat(
branch3x3dbl_3a->getOutputSize(), branch3x3dbl_3b->getOutputSize()
);
// Branch Pool
branchPool_1 = new CUDANet::Layers::AvgPooling2d(
inputShape, inputChannels, {3, 3}, {1, 1}, {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool_1);
branchPool_2 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch_pool"
);
addLayer("", branchPool_2);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch1x1->getOutputSize(), branch_3x3_2_concat->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branch_3x3dbl_3_concat->getOutputSize()
);
concat_3 = new CUDANet::Layers::Concat(
concat_2->getOutputSize(), branchPool_2->getOutputSize()
);
outputSize = concat_3->getOutputSize();
}
InceptionE::~InceptionE() {
delete branch1x1;
delete branch3x3_1;
delete branch3x3_2a;
delete branch3x3_2b;
delete branch_3x3_2_concat;
delete branch3x3dbl_1;
delete branch3x3dbl_2;
delete branch3x3dbl_3a;
delete branch3x3dbl_3b;
delete branch_3x3dbl_3_concat;
delete branchPool_1;
delete branchPool_2;
delete concat_1;
delete concat_2;
delete concat_3;
}
float *InceptionE::forward(const float *d_input) {
float *branch1x1_output = branch1x1->forward(d_input);
float *branch3x3_output = branch3x3_1->forward(d_input);
float *branch3x3_2a_output = branch3x3_2a->forward(branch3x3_output);
float *branch3x3_2b_output = branch3x3_2b->forward(branch3x3_output);
branch3x3_output =
branch_3x3_2_concat->forward(branch3x3_2a_output, branch3x3_2b_output);
float *branch3x3dbl_output = branch3x3dbl_1->forward(d_input);
branch3x3dbl_output = branch3x3dbl_2->forward(branch3x3dbl_output);
float *branch3x3dbl_3a_output =
branch3x3dbl_3a->forward(branch3x3dbl_output);
float *branch3x3dbl_3b_output =
branch3x3dbl_3b->forward(branch3x3dbl_output);
branch3x3dbl_output = branch_3x3dbl_3_concat->forward(
branch3x3dbl_3a_output, branch3x3dbl_3b_output
);
float *branchPool_output = branchPool_1->forward(d_input);
branchPool_output = branchPool_2->forward(branchPool_output);
float *d_output = concat_1->forward(branch1x1_output, branch3x3_output);
d_output = concat_2->forward(d_output, branch3x3dbl_output);
d_output = concat_3->forward(d_output, branchPool_output);
return d_output;
}
shape2d InceptionE::getOutputDims() {
return branch3x3_2a->getOutputDims();
}
int InceptionE::getOutputChannels() {
return branch1x1->getOutputChannels() + branch3x3_2a->getOutputChannels() +
branch3x3_2b->getOutputChannels() +
branch3x3dbl_3a->getOutputChannels() +
branch3x3dbl_3b->getOutputChannels() +
branchPool_2->getOutputChannels();
}
// InceptionV3 Model
InceptionV3::InceptionV3(
const shape2d inputShape,
const int inputChannels,
const int outputSize
)
: CUDANet::Model(inputShape, inputChannels, outputSize) {
conv2d_1a_3x3 = new BasicConv2d(
inputShape, inputChannels, 32, {3, 3}, {2, 2}, {0, 0}, "Conv2d_1a_3x3"
);
addLayer("", conv2d_1a_3x3);
conv2d_2a_3x3 = new BasicConv2d(
conv2d_1a_3x3->getOutputDims(), 32, 32, {3, 3}, {1, 1}, {0, 0},
"Conv2d_2a_3x3"
);
addLayer("", conv2d_2a_3x3);
conv2d_2b_3x3 = new BasicConv2d(
conv2d_2a_3x3->getOutputDims(), 32, 64, {3, 3}, {1, 1}, {1, 1},
"Conv2d_2b_3x3"
);
addLayer("", conv2d_2b_3x3);
maxpool1 = new CUDANet::Layers::MaxPooling2d(
conv2d_2b_3x3->getOutputDims(), 64, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer("Maxpool1", maxpool1);
conv2d_3b_1x1 = new BasicConv2d(
maxpool1->getOutputDims(), 64, 80, {1, 1}, {1, 1}, {0, 0},
"Conv2d_3b_1x1"
);
addLayer("", conv2d_3b_1x1);
conv2d_4a_3x3 = new BasicConv2d(
conv2d_3b_1x1->getOutputDims(), 80, 192, {3, 3}, {1, 1}, {0, 0},
"Conv2d_4a_3x3"
);
addLayer("", conv2d_4a_3x3);
maxpool2 = new CUDANet::Layers::MaxPooling2d(
conv2d_4a_3x3->getOutputDims(), 192, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer("Maxpool2", maxpool2);
Mixed_5b = new InceptionA(maxpool2->getOutputDims(), 192, 32, "Mixed_5b");
addLayer("", Mixed_5b);
Mixed_5c = new InceptionA(Mixed_5b->getOutputDims(), 256, 64, "Mixed_5c");
addLayer("", Mixed_5c);
Mixed_5d = new InceptionA(Mixed_5c->getOutputDims(), 288, 64, "Mixed_5d");
addLayer("", Mixed_5d);
Mixed_6a = new InceptionB(Mixed_5d->getOutputDims(), 288, "Mixed_6a");
addLayer("", Mixed_6a);
Mixed_6b = new InceptionC(Mixed_6a->getOutputDims(), 768, 128, "Mixed_6b");
addLayer("", Mixed_6b);
Mixed_6c = new InceptionC(Mixed_6b->getOutputDims(), 768, 160, "Mixed_6c");
addLayer("", Mixed_6c);
Mixed_6d = new InceptionC(Mixed_6c->getOutputDims(), 768, 160, "Mixed_6d");
addLayer("", Mixed_6d);
Mixed_6e = new InceptionC(Mixed_6d->getOutputDims(), 768, 192, "Mixed_6e");
addLayer("", Mixed_6e);
Mixed_7a = new InceptionD(Mixed_6e->getOutputDims(), 768, "Mixed_7a");
addLayer("", Mixed_7a);
Mixed_7b = new InceptionE(Mixed_7a->getOutputDims(), 1280, "Mixed_7b");
addLayer("", Mixed_7b);
Mixed_7c = new InceptionE(Mixed_7b->getOutputDims(), 2048, "Mixed_7c");
addLayer("", Mixed_7c);
avgpool = new CUDANet::Layers::AdaptiveAvgPooling2d(
Mixed_7c->getOutputDims(), Mixed_7c->getOutputChannels(), {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer("AveragePool", avgpool);
fc = new CUDANet::Layers::Dense(
avgpool->getOutputSize(), 1000, CUDANet::Layers::ActivationType::SOFTMAX
);
addLayer("fc", fc);
}
float *InceptionV3::predict(const float *input) {
float *d_x = inputLayer->forward(input);
d_x = conv2d_1a_3x3->forward(d_x);
d_x = conv2d_2a_3x3->forward(d_x);
d_x = conv2d_2b_3x3->forward(d_x);
d_x = maxpool1->forward(d_x);
d_x = conv2d_3b_1x1->forward(d_x);
d_x = conv2d_4a_3x3->forward(d_x);
d_x = maxpool2->forward(d_x);
d_x = Mixed_5b->forward(d_x);
d_x = Mixed_5c->forward(d_x);
d_x = Mixed_5d->forward(d_x);
d_x = Mixed_6a->forward(d_x);
d_x = Mixed_6b->forward(d_x);
d_x = Mixed_6c->forward(d_x);
d_x = Mixed_6d->forward(d_x);
d_x = Mixed_6e->forward(d_x);
d_x = Mixed_7a->forward(d_x);
d_x = Mixed_7b->forward(d_x);
d_x = Mixed_7c->forward(d_x);
d_x = avgpool->forward(d_x);
d_x = fc->forward(d_x);
float *output = outputLayer->forward(d_x);
return output;
}
InceptionV3::~InceptionV3() {
delete conv2d_1a_3x3;
delete conv2d_2a_3x3;
delete conv2d_2b_3x3;
delete maxpool1;
delete conv2d_3b_1x1;
delete conv2d_4a_3x3;
delete maxpool2;
delete Mixed_5b;
delete Mixed_5c;
delete Mixed_5d;
delete Mixed_6a;
delete Mixed_6b;
delete Mixed_6c;
delete Mixed_6d;
delete Mixed_6e;
delete Mixed_7a;
delete Mixed_7b;
delete Mixed_7c;
delete avgpool;
delete fc;
}

939
examples/inception_v3/inception_v3.cpp
View File

@@ -1,935 +1,8 @@
#include "inception_v3.hpp"
#include <cudanet.cuh>
#include <opencv2/opencv.hpp>
#include <iostream>
class BasicConv2d : public CUDANet::Module {
public:
BasicConv2d(
const shape2d inputShape,
const int inputChannels,
const int outputChannels,
const shape2d kernelSize,
const shape2d stride,
const shape2d padding,
const std::string &prefix
)
: outputChannels(outputChannels) {
// Create the convolution layer
conv = new CUDANet::Layers::Conv2d(
inputShape, inputChannels, kernelSize, stride, outputChannels,
padding, CUDANet::Layers::ActivationType::NONE
);
shape2d batchNormSize = conv->getOutputDims();
batchNorm = new CUDANet::Layers::BatchNorm2d(
batchNormSize, outputChannels, 1e-3f,
CUDANet::Layers::ActivationType::RELU
);
inputSize = inputShape.first * inputShape.second * inputChannels;
outputSize = batchNorm->getOutputDims().first *
batchNorm->getOutputDims().second * outputChannels;
addLayer(prefix + ".conv", conv);
addLayer(prefix + ".bn", batchNorm);
}
~BasicConv2d() {
delete conv;
delete batchNorm;
}
float *forward(const float *d_input) {
float *d_output = conv->forward(d_input);
return batchNorm->forward(d_output);
}
shape2d getOutputDims() {
return batchNorm->getOutputDims();
}
int getOutputChannels() {
return outputChannels;
}
private:
int outputChannels;
CUDANet::Layers::Conv2d *conv;
CUDANet::Layers::BatchNorm2d *batchNorm;
};
class InceptionA : public CUDANet::Module {
public:
InceptionA(
const shape2d inputShape,
const int inputChannels,
const int poolFeatures,
const std::string &prefix
)
: inputShape(inputShape),
inputChannels(inputChannels),
poolFeatures(poolFeatures) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 1x1
branch1x1 = new BasicConv2d(
inputShape, inputChannels, 64, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch1x1"
);
addLayer("", branch1x1);
// Branch 5x5
branch5x5_1 = new BasicConv2d(
inputShape, inputChannels, 48, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch5x5_1"
);
addLayer("", branch5x5_1);
branch5x5_2 = new BasicConv2d(
branch5x5_1->getOutputDims(), 48, 64, {5, 5}, {1, 1}, {2, 2},
prefix + ".branch5x5_2"
);
addLayer("", branch5x5_2);
// Branch 3x3
branch3x3dbl_1 = new BasicConv2d(
inputShape, inputChannels, 64, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3dbl_1"
);
addLayer("", branch3x3dbl_1);
branch3x3dbl_2 = new BasicConv2d(
branch3x3dbl_1->getOutputDims(), 64, 96, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_2"
);
addLayer("", branch3x3dbl_2);
branch3x3dbl_3 = new BasicConv2d(
branch3x3dbl_2->getOutputDims(), 96, 96, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_3"
);
addLayer("", branch3x3dbl_3);
// Branch Pool
branchPool_1 = new CUDANet::Layers::AvgPooling2d(
inputShape, inputChannels, {3, 3}, {1, 1}, {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool_1);
branchPool_2 = new BasicConv2d(
branchPool_1->getOutputDims(), inputChannels, poolFeatures, {1, 1},
{1, 1}, {0, 0}, prefix + ".branch_pool"
);
addLayer("", branchPool_2);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch1x1->getOutputSize(), branch5x5_2->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branch3x3dbl_3->getOutputSize()
);
concat_3 = new CUDANet::Layers::Concat(
concat_2->getOutputSize(), branchPool_2->getOutputSize()
);
outputSize = concat_3->getOutputSize();
}
~InceptionA() {
delete branch1x1;
delete branch5x5_1;
delete branch5x5_2;
delete branch3x3dbl_1;
delete branch3x3dbl_2;
delete branch3x3dbl_3;
delete branchPool_1;
delete branchPool_2;
delete concat_1;
delete concat_2;
delete concat_3;
}
float *forward(const float *d_input) {
float *d_branch1x1_out = branch1x1->forward(d_input);
float *d_branch5x5_out = branch5x5_1->forward(d_input);
d_branch5x5_out = branch5x5_2->forward(d_branch5x5_out);
float *d_branch3x3dbl_out = branch3x3dbl_1->forward(d_input);
d_branch3x3dbl_out = branch3x3dbl_2->forward(d_branch3x3dbl_out);
d_branch3x3dbl_out = branch3x3dbl_3->forward(d_branch3x3dbl_out);
float *d_branchPool_out = branchPool_1->forward(d_input);
d_branchPool_out = branchPool_2->forward(d_branchPool_out);
float *d_output = concat_1->forward(d_branch1x1_out, d_branch5x5_out);
d_output = concat_2->forward(d_output, d_branch3x3dbl_out);
d_output = concat_3->forward(d_output, d_branchPool_out);
return d_output;
}
shape2d getOutputDims() {
return branch1x1->getOutputDims();
}
int getOutputChannels() {
return branch1x1->getOutputChannels() +
branch5x5_2->getOutputChannels() +
branch3x3dbl_3->getOutputChannels() +
branchPool_2->getOutputChannels();
}
private:
shape2d inputShape;
int inputChannels;
int poolFeatures;
BasicConv2d *branch1x1;
BasicConv2d *branch5x5_1;
BasicConv2d *branch5x5_2;
BasicConv2d *branch3x3dbl_1;
BasicConv2d *branch3x3dbl_2;
BasicConv2d *branch3x3dbl_3;
CUDANet::Layers::AvgPooling2d *branchPool_1;
BasicConv2d *branchPool_2;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
CUDANet::Layers::Concat *concat_3;
};
class InceptionB : public CUDANet::Module {
public:
InceptionB(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 3x3
branch3x3 = new BasicConv2d(
inputShape, inputChannels, 384, {3, 3}, {2, 2}, {0, 0},
prefix + ".branch3x3"
);
addLayer("", branch3x3);
// Branch 3x3dbl
branch3x3dbl_1 = new BasicConv2d(
inputShape, inputChannels, 64, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3dbl_1"
);
addLayer("", branch3x3dbl_1);
branch3x3dbl_2 = new BasicConv2d(
branch3x3dbl_1->getOutputDims(), 64, 96, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_2"
);
addLayer("", branch3x3dbl_2);
branch3x3dbl_3 = new BasicConv2d(
branch3x3dbl_2->getOutputDims(), 96, 96, {3, 3}, {2, 2}, {1, 1},
prefix + ".branch3x3dbl_3"
);
addLayer("", branch3x3dbl_3);
branchPool = new CUDANet::Layers::MaxPooling2d(
inputShape, inputChannels, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool);
concat_1 = new CUDANet::Layers::Concat(
branch3x3->getOutputSize(), branch3x3dbl_3->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branchPool->getOutputSize()
);
outputSize = concat_2->getOutputSize();
}
~InceptionB() {
delete branch3x3;
delete branch3x3dbl_1;
delete branch3x3dbl_2;
delete branch3x3dbl_3;
delete branchPool;
delete concat_1;
delete concat_2;
}
float *forward(const float *d_input) {
float *d_branch3x3_out = branch3x3->forward(d_input);
float *d_branch3x3dbl_out = branch3x3dbl_1->forward(d_input);
d_branch3x3dbl_out = branch3x3dbl_2->forward(d_branch3x3dbl_out);
d_branch3x3dbl_out = branch3x3dbl_3->forward(d_branch3x3dbl_out);
float *d_branchPool_out = branchPool->forward(d_input);
float *d_output =
concat_1->forward(d_branch3x3_out, d_branch3x3dbl_out);
d_output = concat_2->forward(d_output, d_branchPool_out);
return d_output;
}
shape2d getOutputDims() {
return branch3x3->getOutputDims();
}
int getOutputChannels() {
return branch3x3->getOutputChannels() +
branch3x3dbl_3->getOutputChannels() + inputChannels;
}
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch3x3;
BasicConv2d *branch3x3dbl_1;
BasicConv2d *branch3x3dbl_2;
BasicConv2d *branch3x3dbl_3;
CUDANet::Layers::MaxPooling2d *branchPool;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
};
class InceptionC : public CUDANet::Module {
public:
InceptionC(
const shape2d inputShape,
const int inputChannels,
const int nChannels_7x7,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 1x1
branch1x1 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch1x1"
);
addLayer("", branch1x1);
// Branch 7x7
branch7x7_1 = new BasicConv2d(
inputShape, inputChannels, nChannels_7x7, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch7x7_1"
);
addLayer("", branch7x7_1);
branch7x7_2 = new BasicConv2d(
branch7x7_1->getOutputDims(), nChannels_7x7, nChannels_7x7, {1, 7},
{1, 1}, {0, 3}, prefix + ".branch7x7_2"
);
addLayer("", branch7x7_2);
branch7x7_3 = new BasicConv2d(
branch7x7_2->getOutputDims(), nChannels_7x7, 192, {7, 1}, {1, 1},
{3, 0}, prefix + ".branch7x7_3"
);
addLayer("", branch7x7_3);
// Branch 7x7dbl
branch7x7dbl_1 = new BasicConv2d(
inputShape, inputChannels, nChannels_7x7, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch7x7dbl_1"
);
addLayer("", branch7x7dbl_1);
branch7x7dbl_2 = new BasicConv2d(
branch7x7dbl_1->getOutputDims(), nChannels_7x7, nChannels_7x7,
{7, 1}, {1, 1}, {3, 0}, prefix + ".branch7x7dbl_2"
);
addLayer("", branch7x7dbl_2);
branch7x7dbl_3 = new BasicConv2d(
branch7x7dbl_2->getOutputDims(), nChannels_7x7, nChannels_7x7,
{1, 7}, {1, 1}, {0, 3}, prefix + ".branch7x7dbl_3"
);
addLayer("", branch7x7dbl_3);
branch7x7dbl_4 = new BasicConv2d(
branch7x7dbl_3->getOutputDims(), nChannels_7x7, nChannels_7x7,
{7, 1}, {1, 1}, {3, 0}, prefix + ".branch7x7dbl_4"
);
addLayer("", branch7x7dbl_4);
branch7x7dbl_5 = new BasicConv2d(
branch7x7dbl_4->getOutputDims(), nChannels_7x7, 192, {1, 7}, {1, 1},
{0, 3}, prefix + ".branch7x7dbl_5"
);
addLayer("", branch7x7dbl_5);
// Branch Pool
branchPool_1 = new CUDANet::Layers::AvgPooling2d(
inputShape, inputChannels, {3, 3}, {1, 1}, {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool_1);
branchPool_2 = new BasicConv2d(
branchPool_1->getOutputDims(), inputChannels, 192, {1, 1}, {1, 1},
{0, 0}, prefix + ".branch_pool"
);
addLayer("", branchPool_2);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch1x1->getOutputSize(), branch7x7_3->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branch7x7dbl_5->getOutputSize()
);
concat_3 = new CUDANet::Layers::Concat(
concat_2->getOutputSize(), branchPool_2->getOutputSize()
);
outputSize = concat_3->getOutputSize();
}
~InceptionC() {
delete branch1x1;
delete branch7x7_1;
delete branch7x7_2;
delete branch7x7_3;
delete branch7x7dbl_1;
delete branch7x7dbl_2;
delete branch7x7dbl_3;
delete branch7x7dbl_4;
delete branch7x7dbl_5;
delete branchPool_1;
delete branchPool_2;
delete concat_1;
delete concat_2;
delete concat_3;
}
float *forward(const float *d_input) {
float *branch1x1_output = branch1x1->forward(d_input);
float *branch7x7_output = branch7x7_1->forward(d_input);
branch7x7_output = branch7x7_2->forward(branch7x7_output);
branch7x7_output = branch7x7_3->forward(branch7x7_output);
float *branch7x7dbl_output = branch7x7dbl_1->forward(d_input);
branch7x7dbl_output = branch7x7dbl_2->forward(branch7x7dbl_output);
branch7x7dbl_output = branch7x7dbl_3->forward(branch7x7dbl_output);
branch7x7dbl_output = branch7x7dbl_4->forward(branch7x7dbl_output);
branch7x7dbl_output = branch7x7dbl_5->forward(branch7x7dbl_output);
float *branchPool_output = branchPool_1->forward(d_input);
branchPool_output = branchPool_2->forward(branchPool_output);
float *d_output = concat_1->forward(branch1x1_output, branch7x7_output);
d_output = concat_2->forward(d_output, branch7x7dbl_output);
d_output = concat_3->forward(d_output, branchPool_output);
return d_output;
}
shape2d getOutputDims() {
return branch1x1->getOutputDims();
}
int getOutputChannels() {
return branch1x1->getOutputChannels() +
branch7x7_3->getOutputChannels() +
branch7x7dbl_5->getOutputChannels() +
branchPool_2->getOutputChannels();
}
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch1x1;
BasicConv2d *branch7x7_1;
BasicConv2d *branch7x7_2;
BasicConv2d *branch7x7_3;
BasicConv2d *branch7x7dbl_1;
BasicConv2d *branch7x7dbl_2;
BasicConv2d *branch7x7dbl_3;
BasicConv2d *branch7x7dbl_4;
BasicConv2d *branch7x7dbl_5;
CUDANet::Layers::AvgPooling2d *branchPool_1;
BasicConv2d *branchPool_2;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
CUDANet::Layers::Concat *concat_3;
};
class InceptionD : public CUDANet::Module {
public:
InceptionD(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 3x3
branch3x3_1 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3_1"
);
addLayer("", branch3x3_1);
branch3x3_2 = new BasicConv2d(
inputShape, 192, 320, {3, 3}, {2, 2}, {0, 0}, prefix + ".branch3x3_2"
);
addLayer("", branch3x3_2);
// Branch 7x7x3
branch7x7x3_1 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch7x7x3_1"
);
addLayer("", branch7x7x3_1);
branch7x7x3_2 = new BasicConv2d(
inputShape, 192, 192, {1, 7}, {1, 1}, {0, 3},
prefix + ".branch7x7x3_2"
);
addLayer("", branch7x7x3_2);
branch7x7x3_3 = new BasicConv2d(
inputShape, 192, 192, {7, 1}, {1, 1}, {3, 0},
prefix + ".branch7x7x3_3"
);
addLayer("", branch7x7x3_3);
branch7x7x3_4 = new BasicConv2d(
inputShape, 192, 192, {3, 3}, {2, 2}, {0, 0},
prefix + ".branch7x7x3_4"
);
addLayer("", branch7x7x3_4);
// Branch Pool
branchPool = new CUDANet::Layers::MaxPooling2d(
inputShape, 192, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch3x3_2->getOutputSize(), branch7x7x3_4->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branchPool->getOutputSize()
);
outputSize = concat_2->getOutputSize();
}
~InceptionD() {
delete branch3x3_1;
delete branch3x3_2;
delete branch7x7x3_1;
delete branch7x7x3_2;
delete branch7x7x3_3;
delete branch7x7x3_4;
delete branchPool;
delete concat_1;
delete concat_2;
}
float *forward(const float *d_input) {
float *branch3x3_output = branch3x3_1->forward(d_input);
branch3x3_output = branch3x3_2->forward(branch3x3_output);
float *branch7x7_output = branch7x7x3_1->forward(d_input);
branch7x7_output = branch7x7x3_2->forward(branch7x7_output);
branch7x7_output = branch7x7x3_3->forward(branch7x7_output);
branch7x7_output = branch7x7x3_4->forward(branch7x7_output);
float *branchPool_output = branchPool->forward(d_input);
float *d_output = concat_1->forward(branch3x3_output, branch7x7_output);
d_output = concat_2->forward(d_output, branchPool_output);
return d_output;
}
shape2d getOutputDims() {
return branch3x3_2->getOutputDims();
}
int getOutputChannels() {
return branch3x3_2->getOutputChannels() +
branch7x7x3_4->getOutputChannels() + inputChannels;
}
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch3x3_1;
BasicConv2d *branch3x3_2;
BasicConv2d *branch7x7x3_1;
BasicConv2d *branch7x7x3_2;
BasicConv2d *branch7x7x3_3;
BasicConv2d *branch7x7x3_4;
CUDANet::Layers::MaxPooling2d *branchPool;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
};
class InceptionE : public CUDANet::Module {
public:
InceptionE(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
)
: inputShape(inputShape), inputChannels(inputChannels) {
inputSize = inputShape.first * inputShape.second * inputChannels;
// Branch 1x1
branch1x1 = new BasicConv2d(
inputShape, inputChannels, 320, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch1x1"
);
addLayer("", branch1x1);
// Branch 3x3
branch3x3_1 = new BasicConv2d(
inputShape, inputChannels, 384, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3_1"
);
addLayer("", branch3x3_1);
branch3x3_2a = new BasicConv2d(
inputShape, 384, 384, {1, 3}, {1, 1}, {0, 1},
prefix + ".branch3x3_2a"
);
addLayer("", branch3x3_2a);
branch3x3_2b = new BasicConv2d(
inputShape, 384, 384, {3, 1}, {1, 1}, {1, 0},
prefix + ".branch3x3_2b"
);
addLayer("", branch3x3_2b);
branch_3x3_2_concat = new CUDANet::Layers::Concat(
branch3x3_2a->getOutputSize(), branch3x3_2b->getOutputSize()
);
// Branch 3x3dbl
branch3x3dbl_1 = new BasicConv2d(
inputShape, inputChannels, 448, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch3x3dbl_1"
);
addLayer("", branch3x3dbl_1);
branch3x3dbl_2 = new BasicConv2d(
inputShape, 448, 384, {3, 3}, {1, 1}, {1, 1},
prefix + ".branch3x3dbl_2"
);
addLayer("", branch3x3dbl_2);
branch3x3dbl_3a = new BasicConv2d(
inputShape, 384, 384, {1, 3}, {1, 1}, {0, 1},
prefix + ".branch3x3dbl_3a"
);
addLayer("", branch3x3dbl_3a);
branch3x3dbl_3b = new BasicConv2d(
inputShape, 384, 384, {3, 1}, {1, 1}, {1, 0},
prefix + ".branch3x3dbl_3b"
);
addLayer("", branch3x3dbl_3b);
branch_3x3dbl_3_concat = new CUDANet::Layers::Concat(
branch3x3dbl_3a->getOutputSize(), branch3x3dbl_3b->getOutputSize()
);
// Branch Pool
branchPool_1 = new CUDANet::Layers::AvgPooling2d(
inputShape, inputChannels, {3, 3}, {1, 1}, {1, 1},
CUDANet::Layers::ActivationType::NONE
);
addLayer(prefix + ".branch_pool", branchPool_1);
branchPool_2 = new BasicConv2d(
inputShape, inputChannels, 192, {1, 1}, {1, 1}, {0, 0},
prefix + ".branch_pool"
);
addLayer("", branchPool_2);
// Concat
concat_1 = new CUDANet::Layers::Concat(
branch1x1->getOutputSize(), branch_3x3_2_concat->getOutputSize()
);
concat_2 = new CUDANet::Layers::Concat(
concat_1->getOutputSize(), branch_3x3dbl_3_concat->getOutputSize()
);
concat_3 = new CUDANet::Layers::Concat(
concat_2->getOutputSize(), branchPool_2->getOutputSize()
);
outputSize = concat_3->getOutputSize();
}
~InceptionE() {
delete branch1x1;
delete branch3x3_1;
delete branch3x3_2a;
delete branch3x3_2b;
delete branch_3x3_2_concat;
delete branch3x3dbl_1;
delete branch3x3dbl_2;
delete branch3x3dbl_3a;
delete branch3x3dbl_3b;
delete branch_3x3dbl_3_concat;
delete branchPool_1;
delete branchPool_2;
delete concat_1;
delete concat_2;
delete concat_3;
}
float *forward(const float *d_input) {
float *branch1x1_output = branch1x1->forward(d_input);
float *branch3x3_output = branch3x3_1->forward(d_input);
float *branch3x3_2a_output = branch3x3_2a->forward(branch3x3_output);
float *branch3x3_2b_output = branch3x3_2b->forward(branch3x3_output);
branch3x3_output = branch_3x3_2_concat->forward(
branch3x3_2a_output, branch3x3_2b_output
);
float *branch3x3dbl_output = branch3x3dbl_1->forward(d_input);
branch3x3dbl_output = branch3x3dbl_2->forward(branch3x3dbl_output);
float *branch3x3dbl_3a_output =
branch3x3dbl_3a->forward(branch3x3dbl_output);
float *branch3x3dbl_3b_output =
branch3x3dbl_3b->forward(branch3x3dbl_output);
branch3x3dbl_output = branch_3x3dbl_3_concat->forward(
branch3x3dbl_3a_output, branch3x3dbl_3b_output
);
float *branchPool_output = branchPool_1->forward(d_input);
branchPool_output = branchPool_2->forward(branchPool_output);
float *d_output = concat_1->forward(branch1x1_output, branch3x3_output);
d_output = concat_2->forward(d_output, branch3x3dbl_output);
d_output = concat_3->forward(d_output, branchPool_output);
return d_output;
}
shape2d getOutputDims() {
return branch3x3_2a->getOutputDims();
}
int getOutputChannels() {
return branch3x3_2a->getOutputChannels() +
branch3x3_2b->getOutputChannels() +
branch3x3dbl_3a->getOutputChannels() +
branch3x3dbl_3b->getOutputChannels() +
branchPool_2->getOutputChannels();
}
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch1x1;
BasicConv2d *branch3x3_1;
BasicConv2d *branch3x3_2a;
BasicConv2d *branch3x3_2b;
CUDANet::Layers::Concat *branch_3x3_2_concat;
BasicConv2d *branch3x3dbl_1;
BasicConv2d *branch3x3dbl_2;
BasicConv2d *branch3x3dbl_3a;
BasicConv2d *branch3x3dbl_3b;
CUDANet::Layers::Concat *branch_3x3dbl_3_concat;
CUDANet::Layers::AvgPooling2d *branchPool_1;
BasicConv2d *branchPool_2;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
CUDANet::Layers::Concat *concat_3;
};
class InceptionV3 : public CUDANet::Model {
public:
InceptionV3(
const shape2d inputShape,
const int inputChannels,
const int outputSize
)
: CUDANet::Model(inputShape, inputChannels, outputSize) {
conv2d_1a_3x3 = new BasicConv2d(
inputShape, inputChannels, 32, {3, 3}, {2, 2}, {0, 0},
"Conv2d_1a_3x3"
);
addLayer("", conv2d_1a_3x3);
conv2d_2a_3x3 = new BasicConv2d(
conv2d_1a_3x3->getOutputDims(), 32, 32, {3, 3}, {1, 1}, {0, 0},
"Conv2d_2a_3x3"
);
addLayer("", conv2d_2a_3x3);
conv2d_2b_3x3 = new BasicConv2d(
conv2d_2a_3x3->getOutputDims(), 32, 64, {3, 3}, {1, 1}, {1, 1},
"Conv2d_2b_3x3"
);
addLayer("", conv2d_2b_3x3);
maxpool1 = new CUDANet::Layers::MaxPooling2d(
conv2d_2b_3x3->getOutputDims(), 64, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer("Maxpool1", maxpool1);
conv2d_3b_1x1 = new BasicConv2d(
maxpool1->getOutputDims(), 64, 80, {1, 1}, {1, 1}, {0, 0},
"Conv2d_3b_1x1"
);
addLayer("", conv2d_3b_1x1);
conv2d_4a_3x3 = new BasicConv2d(
conv2d_3b_1x1->getOutputDims(), 80, 192, {3, 3}, {1, 1}, {0, 0},
"Conv2d_4a_3x3"
);
addLayer("", conv2d_4a_3x3);
maxpool2 = new CUDANet::Layers::MaxPooling2d(
conv2d_4a_3x3->getOutputDims(), 192, {3, 3}, {2, 2}, {0, 0},
CUDANet::Layers::ActivationType::NONE
);
addLayer("Maxpool2", maxpool2);
Mixed_5b =
new InceptionA(maxpool2->getOutputDims(), 192, 32, "Mixed_5b");
addLayer("", Mixed_5b);
Mixed_5c =
new InceptionA(Mixed_5b->getOutputDims(), 256, 64, "Mixed_5c");
addLayer("", Mixed_5c);
Mixed_5d =
new InceptionA(Mixed_5c->getOutputDims(), 288, 64, "Mixed_5d");
addLayer("", Mixed_5d);
Mixed_6a = new InceptionB(Mixed_5d->getOutputDims(), 288, "Mixed_6a");
addLayer("", Mixed_6a);
Mixed_6b =
new InceptionC(Mixed_6a->getOutputDims(), 768, 128, "Mixed_6b");
addLayer("", Mixed_6b);
Mixed_6c =
new InceptionC(Mixed_6b->getOutputDims(), 768, 160, "Mixed_6c");
addLayer("", Mixed_6c);
Mixed_6d =
new InceptionC(Mixed_6c->getOutputDims(), 768, 160, "Mixed_6d");
addLayer("", Mixed_6d);
Mixed_6e =
new InceptionC(Mixed_6d->getOutputDims(), 768, 192, "Mixed_6e");
addLayer("", Mixed_6e);
Mixed_7a = new InceptionD(Mixed_6e->getOutputDims(), 768, "Mixed_7a");
addLayer("", Mixed_7a);
Mixed_7b = new InceptionE(Mixed_7a->getOutputDims(), 1280, "Mixed_7b");
addLayer("", Mixed_7b);
Mixed_7c = new InceptionE(Mixed_7b->getOutputDims(), 2048, "Mixed_7c");
addLayer("", Mixed_7c);
fc = new CUDANet::Layers::Dense(
Mixed_7c->getOutputSize(), 1000,
CUDANet::Layers::ActivationType::SOFTMAX
);
addLayer("fc", fc);
}
float *predict(const float *input) {
float *d_x = inputLayer->forward(input);
d_x = conv2d_1a_3x3->forward(d_x);
d_x = conv2d_2a_3x3->forward(d_x);
d_x = conv2d_2b_3x3->forward(d_x);
d_x = maxpool1->forward(d_x);
d_x = conv2d_3b_1x1->forward(d_x);
d_x = conv2d_4a_3x3->forward(d_x);
d_x = maxpool2->forward(d_x);
d_x = Mixed_5b->forward(d_x);
d_x = Mixed_5c->forward(d_x);
d_x = Mixed_5d->forward(d_x);
d_x = Mixed_6a->forward(d_x);
d_x = Mixed_6b->forward(d_x);
d_x = Mixed_6c->forward(d_x);
d_x = Mixed_6d->forward(d_x);
d_x = Mixed_6e->forward(d_x);
d_x = Mixed_7a->forward(d_x);
d_x = Mixed_7b->forward(d_x);
d_x = Mixed_7c->forward(d_x);
d_x = fc->forward(d_x);
float *output = outputLayer->forward(d_x);
return output;
}
~InceptionV3() {
delete conv2d_1a_3x3;
delete conv2d_2a_3x3;
delete conv2d_2b_3x3;
delete maxpool1;
delete conv2d_3b_1x1;
delete conv2d_4a_3x3;
delete maxpool2;
delete Mixed_5b;
delete Mixed_5c;
delete Mixed_5d;
delete Mixed_6a;
delete Mixed_6b;
delete Mixed_6c;
delete Mixed_6d;
delete Mixed_6e;
delete Mixed_7a;
delete Mixed_7b;
delete Mixed_7c;
delete fc;
}
private:
BasicConv2d *conv2d_1a_3x3;
BasicConv2d *conv2d_2a_3x3;
BasicConv2d *conv2d_2b_3x3;
CUDANet::Layers::MaxPooling2d *maxpool1;
BasicConv2d *conv2d_3b_1x1;
BasicConv2d *conv2d_4a_3x3;
CUDANet::Layers::MaxPooling2d *maxpool2;
InceptionA *Mixed_5b;
InceptionA *Mixed_5c;
InceptionA *Mixed_5d;
InceptionB *Mixed_6a;
InceptionC *Mixed_6b;
InceptionC *Mixed_6c;
InceptionC *Mixed_6d;
InceptionC *Mixed_6e;
InceptionD *Mixed_7a;
InceptionE *Mixed_7b;
InceptionE *Mixed_7c;
CUDANet::Layers::Dense *fc;
};
#include <opencv2/opencv.hpp>
std::vector<float>
readAndNormalizeImage(const std::string &imagePath, int width, int height) {
@@ -973,8 +46,9 @@ int main(int argc, const char *const argv[]) {
const int inputChannels = 3;
const int outputSize = 1000;
InceptionV3 *inception_v3 = new InceptionV3(inputSize, inputChannels, outputSize);
inception_v3->printSummary();
InceptionV3 *inception_v3 =
new InceptionV3(inputSize, inputChannels, outputSize);
// inception_v3->printSummary();
std::cout << std::endl;
@@ -998,5 +72,4 @@ int main(int argc, const char *const argv[]) {
std::cout << "Prediction: " << maxIndex << " " << classLabel << std::endl;
return 0;
}

244
examples/inception_v3/inception_v3.hpp Normal file
View File

@@ -0,0 +1,244 @@
#ifndef INCEPTION_V3_HPP
#define INCEPTION_V3_HPP
#include <cudanet.cuh>
class BasicConv2d : public CUDANet::Module {
public:
BasicConv2d(
const shape2d inputShape,
const int inputChannels,
const int outputChannels,
const shape2d kernelSize,
const shape2d stride,
const shape2d padding,
const std::string &prefix
);
~BasicConv2d();
float *forward(const float *d_input);
shape2d getOutputDims();
int getOutputChannels();
private:
int outputChannels;
CUDANet::Layers::Conv2d *conv;
CUDANet::Layers::BatchNorm2d *batchNorm;
};
class InceptionA : public CUDANet::Module {
public:
InceptionA(
const shape2d inputShape,
const int inputChannels,
const int poolFeatures,
const std::string &prefix
);
~InceptionA();
float *forward(const float *d_input);
shape2d getOutputDims();
int getOutputChannels();
private:
shape2d inputShape;
int inputChannels;
int poolFeatures;
BasicConv2d *branch1x1;
BasicConv2d *branch5x5_1;
BasicConv2d *branch5x5_2;
BasicConv2d *branch3x3dbl_1;
BasicConv2d *branch3x3dbl_2;
BasicConv2d *branch3x3dbl_3;
CUDANet::Layers::AvgPooling2d *branchPool_1;
BasicConv2d *branchPool_2;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
CUDANet::Layers::Concat *concat_3;
};
class InceptionB : public CUDANet::Module {
public:
InceptionB(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
);
~InceptionB();
float *forward(const float *d_input);
shape2d getOutputDims();
int getOutputChannels();
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch3x3;
BasicConv2d *branch3x3dbl_1;
BasicConv2d *branch3x3dbl_2;
BasicConv2d *branch3x3dbl_3;
CUDANet::Layers::MaxPooling2d *branchPool;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
};
class InceptionC : public CUDANet::Module {
public:
InceptionC(
const shape2d inputShape,
const int inputChannels,
const int nChannels_7x7,
const std::string &prefix
);
~InceptionC();
float *forward(const float *d_input);
shape2d getOutputDims();
int getOutputChannels();
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch1x1;
BasicConv2d *branch7x7_1;
BasicConv2d *branch7x7_2;
BasicConv2d *branch7x7_3;
BasicConv2d *branch7x7dbl_1;
BasicConv2d *branch7x7dbl_2;
BasicConv2d *branch7x7dbl_3;
BasicConv2d *branch7x7dbl_4;
BasicConv2d *branch7x7dbl_5;
CUDANet::Layers::AvgPooling2d *branchPool_1;
BasicConv2d *branchPool_2;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
CUDANet::Layers::Concat *concat_3;
};
class InceptionD : public CUDANet::Module {
public:
InceptionD(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
);
~InceptionD();
float *forward(const float *d_input);
shape2d getOutputDims();
int getOutputChannels();
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch3x3_1;
BasicConv2d *branch3x3_2;
BasicConv2d *branch7x7x3_1;
BasicConv2d *branch7x7x3_2;
BasicConv2d *branch7x7x3_3;
BasicConv2d *branch7x7x3_4;
CUDANet::Layers::MaxPooling2d *branchPool;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
};
class InceptionE : public CUDANet::Module {
public:
InceptionE(
const shape2d inputShape,
const int inputChannels,
const std::string &prefix
);
~InceptionE();
float *forward(const float *d_input);
shape2d getOutputDims();
int getOutputChannels();
private:
shape2d inputShape;
int inputChannels;
BasicConv2d *branch1x1;
BasicConv2d *branch3x3_1;
BasicConv2d *branch3x3_2a;
BasicConv2d *branch3x3_2b;
CUDANet::Layers::Concat *branch_3x3_2_concat;
BasicConv2d *branch3x3dbl_1;
BasicConv2d *branch3x3dbl_2;
BasicConv2d *branch3x3dbl_3a;
BasicConv2d *branch3x3dbl_3b;
CUDANet::Layers::Concat *branch_3x3dbl_3_concat;
CUDANet::Layers::AvgPooling2d *branchPool_1;
BasicConv2d *branchPool_2;
CUDANet::Layers::Concat *concat_1;
CUDANet::Layers::Concat *concat_2;
CUDANet::Layers::Concat *concat_3;
};
class InceptionV3 : public CUDANet::Model {
public:
InceptionV3(
const shape2d inputShape,
const int inputChannels,
const int outputSize
);
~InceptionV3();
float *predict(const float *input);
private:
BasicConv2d *conv2d_1a_3x3;
BasicConv2d *conv2d_2a_3x3;
BasicConv2d *conv2d_2b_3x3;
CUDANet::Layers::MaxPooling2d *maxpool1;
BasicConv2d *conv2d_3b_1x1;
BasicConv2d *conv2d_4a_3x3;
CUDANet::Layers::MaxPooling2d *maxpool2;
InceptionA *Mixed_5b;
InceptionA *Mixed_5c;
InceptionA *Mixed_5d;
InceptionB *Mixed_6a;
InceptionC *Mixed_6b;
InceptionC *Mixed_6c;
InceptionC *Mixed_6d;
InceptionC *Mixed_6e;
InceptionD *Mixed_7a;
InceptionE *Mixed_7b;
InceptionE *Mixed_7c;
CUDANet::Layers::AdaptiveAvgPooling2d *avgpool;
CUDANet::Layers::Dense *fc;
};
#endif

8
examples/inception_v3/inception_v3.py
View File

@@ -2,13 +2,15 @@ import torchvision
import sys
sys.path.append('../../tools') # Ugly hack
from utils import export_model_weights, print_model_parameters
from utils import export_model_weights, print_model_parameters, predict
if __name__ == "__main__":
inception = torchvision.models.inception_v3(weights=torchvision.models.Inception_V3_Weights.DEFAULT)
inception.eval()
print_model_parameters(inception) # print layer names and number of parameters
# print_model_parameters(inception) # print layer names and number of parameters
export_model_weights(inception, 'inception_v3_weights.bin')
# export_model_weights(inception, 'inception_v3_weights.bin')
print(predict(inception, "./margot.jpg"))