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

Shelfnet works on cuDNN. To test everything else

Browse Source 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
This commit is contained in:
Micaela Verucchi
2020-06-22 18:11:19 +02:00
parent 9f1e30eaa9
commit 6fd261f628
11 changed files with 290 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/tkDNN/Layer.h
+17 -1
View File
@@ -21,6 +21,7 @@ enum layerType_t {
LAYER_ACTIVATION_MISH,
LAYER_FLATTEN,
LAYER_RESHAPE,
LAYER_RESIZE,
LAYER_MULADD,
LAYER_POOLING,
LAYER_SOFTMAX,
@@ -70,6 +71,7 @@ public:
case LAYER_ACTIVATION_MISH: return "ActivationMish";
case LAYER_FLATTEN: return "Flatten";
case LAYER_RESHAPE: return "Reshape";
case LAYER_RESIZE: return "Resize";
case LAYER_MULADD: return "MulAdd";
case LAYER_POOLING: return "Pooling";
case LAYER_SOFTMAX: return "Softmax";
@@ -427,6 +429,19 @@ public:
};
/**
Resize layer
*/
class Resize : public Layer {
public:
Resize(Network *net, int scale_c, int scale_h, int scale_w, bool fixed=false);
virtual ~Resize();
virtual layerType_t getLayerType() { return LAYER_RESIZE; };
virtual dnnType* infer(dataDim_t &dim, dnnType* srcData);
};
/**
MulAdd layer
@@ -545,7 +560,7 @@ public:
class Shortcut : public Layer {
public:
Shortcut(Network *net, Layer *backLayer);
Shortcut(Network *net, Layer *backLayer, bool mul=false);
virtual ~Shortcut();
virtual layerType_t getLayerType() { return LAYER_SHORTCUT; };
@@ -553,6 +568,7 @@ public:
public:
Layer *backLayer;
bool mul = false;
};
/**
include/tkDNN/kernels.h
+1 -1
View File
@@ -24,7 +24,7 @@ void softmaxForward(float *input, int n, int batch, int batch_offset,
int groups, int group_offset, int stride, float temp, float *output, cudaStream_t stream = cudaStream_t(0));
void shortcutForward(dnnType *srcData, dnnType *dstData, int n1, int c1, int h1, int w1, int s1,
int n2, int c2, int h2, int w2, int s2,
int n2, int c2, int h2, int w2, int s2, bool mul,
cudaStream_t stream = cudaStream_t(0));
void upsampleForward(dnnType *srcData, dnnType *dstData,
include/tkDNN/pluginsRT/ShortcutRT.h
+7 -5
View File
@@ -4,10 +4,11 @@
class ShortcutRT : public IPlugin {
public:
ShortcutRT(tk::dnn::dataDim_t bdim) {
ShortcutRT(tk::dnn::dataDim_t bdim, bool mul) {
this->bc = bdim.c;
this->bh = bdim.h;
this->bw = bdim.w;
this->mul = mul;
}
~ShortcutRT(){
@@ -47,15 +48,14 @@ public:
dnnType *dstData = reinterpret_cast<dnnType*>(outputs[0]);
checkCuda( cudaMemcpyAsync(dstData, srcData, batchSize*c*h*w*sizeof(dnnType), cudaMemcpyDeviceToDevice, stream));
for(int b=0; b < batchSize; ++b)
shortcutForward(srcDataBack + b*bc*bh*bw, dstData + b*c*h*w, 1, c, h, w, 1, 1, bc, bh, bw, 1, stream);
shortcutForward(srcDataBack, dstData, batchSize, c, h, w, 1, batchSize, bc, bh, bw, 1, mul, stream);
return 0;
}
virtual size_t getSerializationSize() override {
return 6*sizeof(int);
return 6*sizeof(int) + sizeof(bool);
}
virtual void serialize(void* buffer) override {
@@ -63,12 +63,14 @@ public:
tk::dnn::writeBUF(buf, bc);
tk::dnn::writeBUF(buf, bh);
tk::dnn::writeBUF(buf, bw);
tk::dnn::writeBUF(buf, mul);
tk::dnn::writeBUF(buf, c);
tk::dnn::writeBUF(buf, h);
tk::dnn::writeBUF(buf, w);
}
int c, h, w;
int bc, bh, bw;
bool mul;
};
src/LayerWgs.cpp
+2 -1
View File
@@ -26,10 +26,11 @@ LayerWgs::LayerWgs(Network *net, int inputs, int outputs,
}
readBinaryFile(weights_path.c_str(), outputs, &bias_h, &bias_d, seek);
seek += outputs;
this->batchnorm = batchnorm;
if(batchnorm) {
seek += outputs;
readBinaryFile(weights_path.c_str(), outputs, &scales_h, &scales_d, seek);
seek += outputs;
readBinaryFile(weights_path.c_str(), outputs, &mean_h, &mean_d, seek);
src/NetworkRT.cpp
+2 -2
View File
@@ -512,7 +512,7 @@ ILayer* NetworkRT::convert_layer(ITensor *input, Shortcut *l) {
else
{
// plugin version
IPlugin *plugin = new ShortcutRT(l->backLayer->output_dim);
IPlugin *plugin = new ShortcutRT(l->backLayer->output_dim, l->mul);
ITensor **inputs = new ITensor*[2];
inputs[0] = input;
inputs[1] = back_tens;
@@ -682,7 +682,7 @@ IPlugin* PluginFactory::createPlugin(const char* layerName, const void* serialDa
bdim.w = readBUF<int>(buf);
bdim.l = 1;
ShortcutRT *r = new ShortcutRT(bdim);
ShortcutRT *r = new ShortcutRT(bdim, readBUF<bool>(buf));
r->c = readBUF<int>(buf);
r->h = readBUF<int>(buf);
r->w = readBUF<int>(buf);
src/Reshape.cpp
+5
View File
@@ -15,6 +15,11 @@ Reshape::Reshape(Network *net, dataDim_t new_dim) : Layer(net) {
output_dim.w = new_dim.w;
output_dim.l = new_dim.l;
output_dim = new_dim;
if(input_dim.tot() != output_dim.tot())
FatalError("Reshape dimension mismatch");
}
Reshape::~Reshape() {
src/Resize.cpp
+38
View File
@@ -0,0 +1,38 @@
#include <iostream>
#include "Layer.h"
#include "kernels.h"
namespace tk { namespace dnn {
Resize::Resize(Network *net, int scale_c, int scale_h, int scale_w, bool fixed) : Layer(net) {
if(fixed){
output_dim.c = scale_c;
output_dim.h = scale_h;
output_dim.w = scale_w;
}
else{
output_dim.c *= scale_c;
output_dim.h *= scale_h;
output_dim.w *= scale_w;
}
checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) );
}
Resize::~Resize() {
checkCuda( cudaFree(dstData) );
}
dnnType* Resize::infer(dataDim_t &dim, dnnType* srcData) {
resizeForward(srcData, dstData, dim.n, dim.c, dim.h, dim.w,
output_dim.c, output_dim.h, output_dim.w);
dim = output_dim;
return dstData;
}
}}
src/Shortcut.cpp
+9 -6
View File
@@ -5,15 +5,18 @@
namespace tk { namespace dnn {
Shortcut::Shortcut(Network *net, Layer *backLayer) : Layer(net) {
Shortcut::Shortcut(Network *net, Layer *backLayer, bool mul) : Layer(net) {
this->backLayer = backLayer;
this->mul = mul;
checkCuda( cudaMalloc(&dstData, output_dim.tot()*sizeof(dnnType)) );
if( /*backLayer->output_dim.c != input_dim.c ||*/
backLayer->output_dim.w != input_dim.w ||
backLayer->output_dim.h != input_dim.h )
FatalError("Shortcut dim missmatch");
//FIXME
// if( /*backLayer->output_dim.c != input_dim.c ||*/
// backLayer->output_dim.w != input_dim.w ||
// backLayer->output_dim.h != input_dim.h )
// FatalError("Shortcut dim missmatch");
}
Shortcut::~Shortcut() {
@@ -26,7 +29,7 @@ dnnType* Shortcut::infer(dataDim_t &dim, dnnType* srcData) {
dataDim_t bdim = this->backLayer->output_dim;
checkCuda(cudaMemcpy(dstData, srcData, dim.tot()*sizeof(dnnType), cudaMemcpyDeviceToDevice));
shortcutForward(this->backLayer->dstData, dstData, dim.n, dim.c, dim.h, dim.w, 1, bdim.n, bdim.c, bdim.h, bdim.w, 1);
shortcutForward(this->backLayer->dstData, dstData, dim.n, dim.c, dim.h, dim.w, 1, bdim.n, bdim.c, bdim.h, bdim.w, 1, mul);
//update data dimensions
dim = output_dim;
src/kernels/resize.cu
+14 -27
View File
@@ -1,46 +1,33 @@
#include "kernels.h"
#include <stdio.h>
#define MIN(a,b) (((a)<(b))?(a):(b))
#define MAX(a,b) (((a)>(b))?(a):(b))
__global__ void resize_kernel( int i_N,float *x, int i_w, int i_h, int i_c,
__global__ void resize_kernel( int size,float *x, int i_w, int i_h, int i_c,
int o_w, int o_h, int o_c, int batch, float *out)
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if(i >= i_N) return;
int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if(id >= size) return;
int out_index = i;
int out_w = i%o_w;
i = i/o_w;
int out_h = i%o_h;
i = i/o_h;
int out_c = i%o_c;
i = i/o_c;
int i = id % o_w;
id /= o_w;
int j = id % o_h;
id /= o_h;
int k = id % o_c;
id /= o_c;
int b = id % batch;
//copying last column/last row as padding
int in_index = ((i*i_c + MIN(out_c,i_c-1))*i_h + MIN(out_h,i_h-1))*i_w + MIN(out_w, i_w-1);
out[out_index] = x[in_index];
int out_index = i + o_w*(j + o_h*(k + o_c*b));
int add_index = i/(o_w/i_w) + i_w*(j/(o_h/i_h) + i_h*(k + i_c*b));
out[out_index] = x[add_index];
}
void resizeForward( dnnType* srcData, dnnType* dstData, int n, int i_c, int i_h, int i_w,
int o_c, int o_h, int o_w, cudaStream_t stream )
{
int i_size = n*i_c*i_h*i_w;
int o_size = n*o_c*o_h*o_w;
int blocks = (o_size+255)/256;
int threads = 256;
if(i_c == o_c && i_h == o_h && i_w == o_w )
{
checkCuda(cudaMemcpy(dstData, srcData, i_size*sizeof(dnnType), cudaMemcpyDeviceToDevice));
}
else
{
checkCuda(cudaMemset(dstData, 0, o_size*sizeof(dnnType)));
resize_kernel<<<blocks, threads, 0, stream>>>(o_size, srcData, i_w, i_h, i_c, o_w, o_h, o_c, n, dstData);
// printDeviceVector(i_size, srcData);
// printDeviceVector(o_size, dstData);
}
resize_kernel<<<blocks, threads, 0, stream>>>(o_size, srcData, i_w, i_h, i_c, o_w, o_h, o_c, n, dstData);
}
src/kernels/shortcut.cu
+48 -15
View File
@@ -21,27 +21,60 @@ __global__ void shortcut_kernel(int size, int minw, int minh, int minc, int stri
//out[out_index] += add[add_index];
}
__global__ void shortcut_mul_kernel(int size, int minw, int minh, int minc, int sample, int batch,
int w1, int h1, int c1, dnnType *mul,
int w2, int h2, int c2, float s1, float s2, dnnType *out)
{
int id = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if (id >= size) return;
int i = id % minw;
id /= minw;
int j = id % minh;
id /= minh;
int k = id % minc;
id /= minc;
int b = id % batch;
int out_index = i*sample + w1*(j*sample + h1*(k + c1*b));
out[out_index] = out[out_index] * mul[k + c2*b];
}
void shortcutForward(dnnType* srcData, dnnType* dstData, int n1, int c1, int h1, int w1, int s1,
int n2, int c2, int h2, int w2, int s2,
cudaStream_t stream)
bool mul, cudaStream_t stream)
{
assert(n1 == n2);
int batch = n1;
int minw = (w1 < w2) ? w1 : w2;
int minh = (h1 < h2) ? h1 : h2;
int minc = (c1 < c2) ? c1 : c2;
if(!mul){
int minw = (w1 < w2) ? w1 : w2;
int minh = (h1 < h2) ? h1 : h2;
int minc = (c1 < c2) ? c1 : c2;
int stride = w1/w2;
int sample = w2/w1;
assert(stride == h1/h2);
assert(sample == h2/h1);
if(stride < 1) stride = 1;
if(sample < 1) sample = 1;
int stride = w1/w2;
int sample = w2/w1;
assert(stride == h1/h2);
assert(sample == h2/h1);
if(stride < 1) stride = 1;
if(sample < 1) sample = 1;
int size = batch * minw * minh * minc;
int blocks = (size+255)/256;
int threads = 256;
shortcut_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, stride, sample, batch,
w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
}
else{
int minw = w1;
int minh = h1;
int minc = c1;
int sample = 1;
int size = batch * minw * minh * minc;
int blocks = (size+255)/256;
int threads = 256;
shortcut_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, stride, sample, batch,
w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
int size = batch * minw * minh * minc;
int blocks = (size+255)/256;
int threads = 256;
shortcut_mul_kernel<<<blocks, threads, 0, stream>>>(size, minw, minh, minc, sample, batch,
w1, h1, c1, srcData, w2, h2, c2, s1, s2, dstData);
}
}
tests/shelfnet/shelfnet.cpp
+147 -28
View File
@@ -1,5 +1,9 @@
#include <iostream>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include "tkdnn.h"
#include "NetworkViz.h"
const char *output_bin1 = "shelfnet/debug/classification_headers-5.bin";
@@ -29,35 +33,49 @@ const char *backbone[] = {
"shelfnet/layers/backbone-layer4-1-conv2.bin"};
const char *conv_out[] = {
"shelfnet/layers/conv_out-conv-conv.bin",
"shelfnet/layers/conv_out-conv_out.bin",
"shelfnet/layers/conv_out16-conv-conv.bin",
"shelfnet/layers/conv_out16-conv_out.bin",
"shelfnet/layers/conv_out32-conv-conv.bin",
"shelfnet/layers/conv_out32-conv_out.bin",
"shelfnet/layers/conv_out-conv-conv.bin",
"shelfnet/layers/conv_out-conv_out.bin"};
"shelfnet/layers/conv_out32-conv_out.bin"
};
const char *decoder[] = {
"shelfnet/layers/decoder-bottom-conv1.bin",
"shelfnet/layers/decoder-up_conv_list-0-conv_atten.bin",
"shelfnet/layers/decoder-bottom-conv12.bin",
"shelfnet/layers/decoder-up_conv_list-0-conv-conv.bin",
"shelfnet/layers/decoder-up_conv_list-1-conv_atten.bin",
"shelfnet/layers/decoder-up_conv_list-1-conv-conv.bin",
"shelfnet/layers/decoder-up_conv_list-0-conv_atten.bin",
"shelfnet/layers/decoder-up_dense_list-0-conv.bin",
"shelfnet/layers/decoder-up_dense_list-1-conv.bin"};
"shelfnet/layers/decoder-up_conv_list-1-conv-conv.bin",
"shelfnet/layers/decoder-up_conv_list-1-conv_atten.bin",
"shelfnet/layers/decoder-up_dense_list-1-conv.bin"
};
const char *ladder[] = {
"shelfnet/layers/ladder-bottom-conv1.bin",
"shelfnet/layers/ladder-down_conv_list-0.bin",
"shelfnet/layers/ladder-down_conv_list-1.bin",
"shelfnet/layers/ladder-down_module_list-0-conv1.bin",
"shelfnet/layers/ladder-down_module_list-1-conv1.bin",
"shelfnet/layers/ladder-inconv-conv1.bin",
"shelfnet/layers/ladder-up_conv_list-0-conv_atten.bin",
"shelfnet/layers/ladder-inconv-conv12.bin",
"shelfnet/layers/ladder-down_module_list-0-conv1.bin",
"shelfnet/layers/ladder-down_module_list-0-conv12.bin",
"shelfnet/layers/ladder-down_conv_list-0.bin",
"shelfnet/layers/ladder-down_module_list-1-conv1.bin",
"shelfnet/layers/ladder-down_module_list-1-conv12.bin",
"shelfnet/layers/ladder-down_conv_list-1.bin",
"shelfnet/layers/ladder-bottom-conv1.bin",
"shelfnet/layers/ladder-bottom-conv12.bin",
"shelfnet/layers/ladder-up_conv_list-0-conv-conv.bin",
"shelfnet/layers/ladder-up_conv_list-1-conv_atten.bin",
"shelfnet/layers/ladder-up_conv_list-1-conv-conv.bin",
"shelfnet/layers/ladder-up_conv_list-0-conv_atten.bin",
"shelfnet/layers/ladder-up_dense_list-0-conv.bin",
"shelfnet/layers/ladder-up_conv_list-1-conv-conv.bin",
"shelfnet/layers/ladder-up_conv_list-1-conv_atten.bin",
"shelfnet/layers/ladder-up_dense_list-1-conv.bin"};
const char *trans[] = {
@@ -71,11 +89,11 @@ int main()
int classes = 19;
// Network layout
// Network layout
tk::dnn::dataDim_t dim(1, 3, 1024, 1024, 1);
tk::dnn::Network net(dim);
int bi = 0;
int bi = 0, di = 0, li = 0, ci = 0;
new tk::dnn::Conv2d(&net, 64, 7, 7, 2, 2, 3, 3, backbone[bi++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
tk::dnn::Layer* last = new tk::dnn::Pooling (&net, 3, 3, 2, 2, 1, 1, tk::dnn::POOLING_MAX);
@@ -105,24 +123,122 @@ int main()
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, backbone[bi++], true);
if(i != 2)
{new tk::dnn::Shortcut(&net, last);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
features.push_back(last);}
features.push_back(last);
}
// for(int i=0; i<features.size(); ++i){
// new tk::dnn::Route(&net, &features[i], 1);
// int out_channel = pow(2,6+i);
// new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, trans[i], true);
// new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
for(int i=0; i<features.size(); ++i){
new tk::dnn::Route(&net, &features[i], 1);
int out_channel = pow(2,6+i);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, trans[i], true);
features[i] = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
}
//DECODER
last = features[2];
std::vector<tk::dnn::Layer*> up_out;
//bottom
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, decoder[di++], true, false, 1, true);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
up_out.push_back(last);
for(int i=0; i<2; ++i){
int out_channel = pow(2,7-i);
//up-conv
std::cout<<out_channel<<std::endl;
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, decoder[di++], true);
tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
new tk::dnn::Route(&net, &last, 1);
new tk::dnn::Shortcut(&net, act, true);
//interpolate
new tk::dnn::Resize(&net, 1,2,2);
new tk::dnn::Shortcut(&net, features[1-i]);
//up-dense
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, decoder[di++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
up_out.push_back(last);
}
//LADDER
std::vector<tk::dnn::Layer*> down_out;
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Shortcut(&net, last);
new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
for(int i=0; i<2;++i){
int out_channel = pow(2,6+i);
tk::dnn::Layer* l_last = new tk::dnn::Shortcut(&net, up_out[2-i]);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Shortcut(&net, l_last);
l_last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
down_out.push_back(l_last);
new tk::dnn::Conv2d (&net, out_channel*2, 3, 3, 2, 2, 1, 1, ladder[li++], false);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
}
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Conv2d (&net, 256, 3, 3, 1, 1, 1, 1, ladder[li++], true, false, 1, true);
new tk::dnn::Shortcut(&net, last);
last = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_RELU);
up_out.clear();
up_out.push_back(last);
for(int i=0; i<2; ++i){
int out_channel = pow(2,7-i);
//up-conv
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Pooling(&net, last->output_dim.w, last->output_dim.h, last->output_dim.w, last->output_dim.h, 0, 0, tk::dnn::POOLING_AVERAGE);
new tk::dnn::Conv2d (&net, out_channel, 1, 1, 1, 1, 0, 0, ladder[li++], true);
tk::dnn::Layer* act = new tk::dnn::Activation (&net, CUDNN_ACTIVATION_SIGMOID);
new tk::dnn::Route(&net, &last, 1);
new tk::dnn::Shortcut(&net, act, true);
//interpolate
new tk::dnn::Resize(&net, 1,2,2);
new tk::dnn::Shortcut(&net, down_out[1-i]);
// //up-dense
new tk::dnn::Conv2d (&net, out_channel, 3, 3, 1, 1, 1, 1, ladder[li++], true);
last = new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
up_out.push_back(last);
}
// for(int i=2;i>=0;--i){
// new tk::dnn::Route(&net, &up_out[i], 1);
new tk::dnn::Conv2d (&net, 64, 3, 3, 1, 1, 1, 1, conv_out[ci++], true);
new tk::dnn::Activation (&net, tk::dnn::ACTIVATION_LEAKY);
new tk::dnn::Conv2d (&net, 19, 3, 3, 1, 1, 1, 1, conv_out[ci++], false);
/*up_out[i] =*/ new tk::dnn::Resize(&net, 19, net.input_dim.h, net.input_dim.w, true);
// }
new tk::dnn::Softmax(&net);
const char *output_bin = "shelfnet/debug/backbone-layer4-1-bn2.bin";
const char *output_bin = "shelfnet/debug/fofmaf.bin";
@@ -210,4 +326,7 @@ int main()
// ret_cudnn_tensorrt |= checkResult(loc->output_dim.tot(), loc->dstData, rt_out4) == 0 ? 0 : ERROR_CUDNNvsTENSORRT;
// return ret_cudnn | ret_tensorrt | ret_cudnn_tensorrt;
cv::Mat viz = vizLayer2Mat(&net, net.num_layers-1);
cv::imwrite("test.png", viz);
}