Skip to content

D
dlib
Commit 1974e68d authored by Fm's avatar Fm
Browse files
Options

Removed friend declaration of dnn_tester from core.h

parent d32bcdfa
Hide whitespace changes
Inline Side-by-side
Showing with 64 additions and 69 deletions
dlib/dnn/core.h
View file @ 1974e68d
......@@ -648,7 +648,6 @@ namespace dlib
friend class add_skip_layer;
template <size_t N, template<typename> class L, typename S>
friend class repeat;
friend class dnn_tester;
// Allow copying networks from one to another as long as their corresponding
// layers can be constructed from each other.
......@@ -1521,7 +1520,6 @@ namespace dlib
friend class add_skip_layer;
template <size_t N, template<typename> class L, typename S>
friend class repeat;
friend class dnn_tester;
// You wouldn't put a tag on a layer if you didn't want to access its forward
// outputs. So this is always true.
......
dlib/test/dnn.cpp
View file @ 1974e68d
......@@ -11,11 +11,12 @@
#include "tester.h"
namespace dlib
namespace
{
using namespace std;
using namespace test;
using namespace dlib;
using namespace std;
logger dlog("test.dnn");
......@@ -1186,7 +1187,7 @@ namespace dlib
r*stride_y+y_offset,
window_width,
window_height)));
float err = std::abs(image_plane(A,s,k)(r,c) - expected);
float err = abs(image_plane(A,s,k)(r,c) - expected);
DLIB_TEST_MSG(err < 1e-5, err << " " << expected << " " << image_plane(A,s,k)(r,c));
}
}
......@@ -1511,6 +1512,66 @@ namespace dlib
}
#endif//DLIB_USE_CUDA
template <typename SUBNET> using concat_block1 = con<5,1,1,1,1,SUBNET>;
template <typename SUBNET> using concat_block2 = con<8,3,3,1,1,SUBNET>;
template <typename SUBNET> using concat_block3 = max_pool<3,3,1,1,SUBNET>;
template <typename SUBNET> using concat_incept = inception3<concat_block1,concat_block2,concat_block3,SUBNET>;
void test_concat()
{
using namespace dlib::tt;
print_spinner();
using net_type = concat_incept<input<matrix<float>>>;
resizable_tensor data(10, 1, 111, 222);
data = matrix_cast<float>(gaussian_randm(data.num_samples(), data.k() * data.nr() * data.nc(), 1));
net_type net;
auto& out = net.forward(data);
auto& b1o = layer<itag1>(net).get_output();
auto& b2o = layer<itag2>(net).get_output();
auto& b3o = layer<itag3>(net).get_output();
resizable_tensor dest(10, 14, 111, 222);
copy_tensor(dest, 0, b1o, 0, b1o.k());
copy_tensor(dest, b1o.k(), b2o, 0, b2o.k());
copy_tensor(dest, b1o.k() + b2o.k(), b3o, 0, b3o.k());
DLIB_TEST(dest.size() == out.size());
int error = memcmp(dest.host(), out.host(), dest.size());
DLIB_TEST(error == 0);
resizable_tensor gr(10, 14, 111, 222);
gr = matrix_cast<float>(gaussian_randm(gr.num_samples(), gr.k() * gr.nr() * gr.nc(), 1));
resizable_tensor params;
net.layer_details().backward(gr, net, params);
auto& b1g = layer<itag1>(net).subnet().get_gradient_input();
auto& b2g = layer<itag2>(net).subnet().get_gradient_input();
auto& b3g = layer<itag3>(net).subnet().get_gradient_input();
resizable_tensor g1(10, 5, 111, 222);
resizable_tensor g2(10, 8, 111, 222);
resizable_tensor g3(10, 1, 111, 222);
copy_tensor(g1, 0, gr, 0, g1.k());
copy_tensor(g2, 0, gr, g1.k(), g2.k());
copy_tensor(g3, 0, gr, g1.k() + g2.k(), g3.k());
DLIB_TEST(g1.size() == b1g.size());
error = memcmp(g1.host(), b1g.host(), b1g.size());
DLIB_TEST(error == 0);
DLIB_TEST(g2.size() == b2g.size());
error = memcmp(g2.host(), b2g.host(), b2g.size());
DLIB_TEST(error == 0);
DLIB_TEST(g3.size() == b3g.size());
error = memcmp(g3.host(), b3g.host(), b3g.size());
DLIB_TEST(error == 0);
}
// ----------------------------------------------------------------------------------------
class dnn_tester : public tester
......@@ -1522,8 +1583,6 @@ namespace dlib
"Runs tests on the deep neural network tools.")
{}
void test_concat();
void perform_test (
)
{
......@@ -1579,68 +1638,6 @@ namespace dlib
test_concat();
}
} a;
template <typename SUBNET> using concat_block1 = con<5,1,1,1,1,SUBNET>;
template <typename SUBNET> using concat_block2 = con<8,3,3,1,1,SUBNET>;
template <typename SUBNET> using concat_block3 = max_pool<3,3,1,1,SUBNET>;
template <typename SUBNET> using concat_incept = inception3<concat_block1,concat_block2,concat_block3,SUBNET>;
void dnn_tester::test_concat()
{
using namespace dlib::tt;
print_spinner();
using net_type = concat_incept<input<matrix<float>>>;
resizable_tensor data(10, 1, 111, 222);
data = matrix_cast<float>(gaussian_randm(data.num_samples(), data.k() * data.nr() * data.nc(), 1));
net_type net;
auto& out = net.forward(data);
auto& b1o = layer<itag1>(net).get_output();
auto& b2o = layer<itag2>(net).get_output();
auto& b3o = layer<itag3>(net).get_output();
resizable_tensor dest(10, 14, 111, 222);
copy_tensor(dest, 0, b1o, 0, b1o.k());
copy_tensor(dest, b1o.k(), b2o, 0, b2o.k());
copy_tensor(dest, b1o.k() + b2o.k(), b3o, 0, b3o.k());
DLIB_TEST(dest.size() == out.size());
int error = memcmp(dest.host(), out.host(), dest.size());
DLIB_TEST(error == 0);
resizable_tensor gr(10, 14, 111, 222);
gr = matrix_cast<float>(gaussian_randm(gr.num_samples(), gr.k() * gr.nr() * gr.nc(), 1));
memcpy(net.get_gradient_input(), gr);
net.back_propagate_error(data);
auto& b1g = layer<itag1>(net).subnet().x_grad;
auto& b2g = layer<itag2>(net).subnet().x_grad;
auto& b3g = layer<itag3>(net).subnet().x_grad;
resizable_tensor g1(10, 5, 111, 222);
resizable_tensor g2(10, 8, 111, 222);
resizable_tensor g3(10, 1, 111, 222);
copy_tensor(g1, 0, gr, 0, g1.k());
copy_tensor(g2, 0, gr, g1.k(), g2.k());
copy_tensor(g3, 0, gr, g1.k() + g2.k(), g3.k());
DLIB_TEST(g1.size() == b1g.size());
error = memcmp(g1.host(), b1g.host(), b1g.size());
DLIB_TEST(error == 0);
DLIB_TEST(g2.size() == b2g.size());
error = memcmp(g2.host(), b2g.host(), b2g.size());
DLIB_TEST(error == 0);
DLIB_TEST(g3.size() == b3g.size());
error = memcmp(g3.host(), b3g.host(), b3g.size());
DLIB_TEST(error == 0);
}
}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment