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

better network model

Browse Source
This commit is contained in:
Francesco Gatti
2017-08-01 23:03:02 +02:00
parent 300b0af5dd
commit e8355cee67
22 changed files with 166 additions and 179 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/build_models.sh
Executable → Regular
View File
tests/mnist/test_mnist.cpp
+9 -10
View File
@@ -11,17 +11,16 @@ const char *output_bin = "../tests/mnist/output.bin";
int main() {
// Network layout
tkDNN::Network net;
tkDNN::dataDim_t dim(1, 1, 28, 28, 1);
tkDNN::Layer *l;
l = new tkDNN::Conv2d (&net, dim, 20, 5, 5, 1, 1, 0, 0, c0_bin);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Conv2d (&net, l->output_dim, 50, 5, 5, 1, 1, 0, 0, c1_bin);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Dense (&net, l->output_dim, 500, d2_bin);
l = new tkDNN::Activation (&net, l->output_dim, CUDNN_ACTIVATION_RELU);
l = new tkDNN::Dense (&net, l->output_dim, 10, d3_bin);
l = new tkDNN::Softmax (&net, l->output_dim);
tkDNN::Network net(dim);
tkDNN::Conv2d l0(&net, 20, 5, 5, 1, 1, 0, 0, c0_bin);
tkDNN::Pooling l1(&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Conv2d l2(&net, 50, 5, 5, 1, 1, 0, 0, c1_bin);
tkDNN::Pooling l3(&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Dense l4(&net, 500, d2_bin);
tkDNN::Activation l5(&net, CUDNN_ACTIVATION_RELU);
tkDNN::Dense l6(&net, 10, d3_bin);
tkDNN::Softmax l7(&net);
// Load input
value_type *data;
tests/mnist/test_mnistRT.cpp
+13 -14
View File
@@ -27,17 +27,16 @@ int main() {
std::cout<<"\n==== CUDNN ====\n";
// Network layout
tkDNN::Network net;
tkDNN::dataDim_t dim(1, 1, 28, 28, 1);
tkDNN::Layer *l;
l = new tkDNN::Conv2d (&net, dim, 20, 5, 5, 1, 1, 0, 0, c0_bin);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Conv2d (&net, l->output_dim, 50, 5, 5, 1, 1, 0, 0, c1_bin);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Dense (&net, l->output_dim, 500, d2_bin);
l = new tkDNN::Activation (&net, l->output_dim, CUDNN_ACTIVATION_RELU);
l = new tkDNN::Dense (&net, l->output_dim, 10, d3_bin);
l = new tkDNN::Softmax (&net, l->output_dim);
tkDNN::Network net(dim);
tkDNN::Conv2d l0(&net, 20, 5, 5, 1, 1, 0, 0, c0_bin);
tkDNN::Pooling l1(&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Conv2d l2(&net, 50, 5, 5, 1, 1, 0, 0, c1_bin);
tkDNN::Pooling l3(&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Dense l4(&net, 500, d2_bin);
tkDNN::Activation l5(&net, CUDNN_ACTIVATION_RELU);
tkDNN::Dense l6(&net, 10, d3_bin);
tkDNN::Softmax l7(&net);
// Load input
value_type *data;
@@ -72,7 +71,7 @@ int main() {
auto input = network->addInput("data", dt, DimsCHW{ 1, 28, 28});
assert(input != nullptr);
tkDNN::Conv2d *c0 = (tkDNN::Conv2d*) (net.layers[0]);
tkDNN::Conv2d *c0 = &l0;
Weights w { dt, c0->data_h, c0->inputs*c0->outputs*c0->kernelH*c0->kernelW};
Weights b { dt, c0->bias_h, c0->outputs};
// Add a convolution layer with 20 outputs and a 5x5 filter.
@@ -85,7 +84,7 @@ int main() {
assert(pool1 != nullptr);
pool1->setStride(DimsHW{2, 2});
tkDNN::Conv2d *c1 = (tkDNN::Conv2d*) (net.layers[2]);
tkDNN::Conv2d *c1 = &l2;
Weights w1 { dt, c1->data_h, c1->inputs*c1->outputs*c1->kernelH*c1->kernelW};
Weights b1 { dt, c1->bias_h, c1->outputs};
// Add a second convolution layer with 50 outputs and a 5x5 filter.
@@ -98,7 +97,7 @@ int main() {
assert(pool2 != nullptr);
pool2->setStride(DimsHW{2, 2});
tkDNN::Dense *d2 = (tkDNN::Dense*) (net.layers[4]);
tkDNN::Dense *d2 = &l4;
Weights w2 { dt, d2->data_h, d2->inputs*d2->outputs};
Weights b2 { dt, d2->bias_h, d2->outputs};
// Add a fully connected layer with 500 outputs.
@@ -109,7 +108,7 @@ int main() {
auto relu1 = network->addActivation(*ip1->getOutput(0), ActivationType::kRELU);
assert(relu1 != nullptr);
tkDNN::Conv2d *d3 = (tkDNN::Conv2d*) (net.layers[6]);
tkDNN::Dense *d3 = &l6;
Weights w3 { dt, d3->data_h, d3->inputs*d3->outputs};
Weights b3 { dt, d3->bias_h, d3->outputs};
// Add a second fully connected layer with 20 outputs.
tests/test/test_model.py → tests/simple/test_model.py
View File
tests/test/test_simple.cpp → tests/simple/test_simple.cpp
+8 -12
View File
@@ -10,16 +10,15 @@ const char *output_bin = "../tests/test/output.bin";
int main() {
// Network layout
tkDNN::Network net;
tkDNN::dataDim_t dim(1, 1, 10, 10, 1);
tkDNN::Layer *l;
l = new tkDNN::Conv2d (&net, dim, 2, 4, 4, 2, 2, 0, 0, c0_bin);
l = new tkDNN::Activation (&net, l->output_dim, CUDNN_ACTIVATION_RELU);
l = new tkDNN::Conv2d (&net, l->output_dim, 4, 2, 2, 1, 1, 0, 0, c1_bin);
l = new tkDNN::Activation (&net, l->output_dim, CUDNN_ACTIVATION_RELU);
l = new tkDNN::Flatten (&net, l->output_dim);
l = new tkDNN::Dense (&net, l->output_dim, 4, d2_bin);
l = new tkDNN::Activation (&net, l->output_dim, CUDNN_ACTIVATION_RELU);
tkDNN::Network net(dim);
tkDNN::Conv2d l0(&net, 2, 4, 4, 2, 2, 0, 0, c0_bin);
tkDNN::Activation l1(&net, CUDNN_ACTIVATION_RELU);
tkDNN::Conv2d l2(&net, 4, 2, 2, 1, 1, 0, 0, c1_bin);
tkDNN::Activation l3(&net, CUDNN_ACTIVATION_RELU);
tkDNN::Flatten l4(&net);
tkDNN::Dense l5(&net, 4, d2_bin);
tkDNN::Activation l6(&net, CUDNN_ACTIVATION_RELU);
// Load input
value_type *data;
@@ -30,11 +29,8 @@ int main() {
dim.print(); //print initial dimension
TIMER_START
// Inference
data = net.infer(dim, data); dim.print();
TIMER_STOP
// Print result
tests/yolo/yolo.cpp
+58 -58
View File
@@ -30,74 +30,74 @@ const char *output_bin = "../tests/yolo/layers/output.bin";
int main() {
// Network layout
tkDNN::Network net;
tkDNN::dataDim_t dim(1, 3, 608, 608, 1);
tkDNN::Layer *l;
l = new tkDNN::Conv2d (&net, dim, 32, 3, 3, 1, 1, 1, 1, c0_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Network net(dim);
l = new tkDNN::Conv2d (&net, l->output_dim, 64, 3, 3, 1, 1, 1, 1, c2_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Conv2d c0 (&net, 32, 3, 3, 1, 1, 1, 1, c0_bin, true);
tkDNN::Activation a0 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Pooling p1 (&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Conv2d (&net, l->output_dim, 128, 3, 3, 1, 1, 1, 1, c4_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 64, 1, 1, 1, 1, 0, 0, c5_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 128, 3, 3, 1, 1, 1, 1, c6_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Conv2d c2 (&net, 64, 3, 3, 1, 1, 1, 1, c2_bin, true);
tkDNN::Activation a2 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Pooling p3 (&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Conv2d (&net, l->output_dim, 256, 3, 3, 1, 1, 1, 1, c8_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 128, 1, 1, 1, 1, 0, 0, c9_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 256, 3, 3, 1, 1, 1, 1, c10_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Conv2d c4 (&net, 128, 3, 3, 1, 1, 1, 1, c4_bin, true);
tkDNN::Activation a4 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c5 (&net, 64, 1, 1, 1, 1, 0, 0, c5_bin, true);
tkDNN::Activation a5 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c6 (&net, 128, 3, 3, 1, 1, 1, 1, c6_bin, true);
tkDNN::Activation a6 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Pooling p7 (&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Conv2d (&net, l->output_dim, 512, 3, 3, 1, 1, 1, 1, c12_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 256, 1, 1, 1, 1, 0, 0, c13_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 512, 3, 3, 1, 1, 1, 1, c14_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 256, 1, 1, 1, 1, 0, 0, c15_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 512, 3, 3, 1, 1, 1, 1, c16_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY); //29
l = new tkDNN::Pooling (&net, l->output_dim, 2, 2, 2, 2, tkDNN::POOLING_MAX);
tkDNN::Conv2d c8 (&net, 256, 3, 3, 1, 1, 1, 1, c8_bin, true);
tkDNN::Activation a8 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c9 (&net, 128, 1, 1, 1, 1, 0, 0, c9_bin, true);
tkDNN::Activation a9 (&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c10(&net, 256, 3, 3, 1, 1, 1, 1, c10_bin, true);
tkDNN::Activation a10(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Pooling p11(&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
l = new tkDNN::Conv2d (&net, l->output_dim, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 512, 1, 1, 1, 1, 0, 0, c19_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 512, 1, 1, 1, 1, 0, 0, c21_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY); //44
tkDNN::Conv2d c12(&net, 512, 3, 3, 1, 1, 1, 1, c12_bin, true);
tkDNN::Activation a12(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c13(&net, 256, 1, 1, 1, 1, 0, 0, c13_bin, true);
tkDNN::Activation a13(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c14(&net, 512, 3, 3, 1, 1, 1, 1, c14_bin, true);
tkDNN::Activation a14(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c15(&net, 256, 1, 1, 1, 1, 0, 0, c15_bin, true);
tkDNN::Activation a15(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c16(&net, 512, 3, 3, 1, 1, 1, 1, c16_bin, true);
tkDNN::Activation a16(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Pooling p17(&net, 2, 2, 2, 2, tkDNN::POOLING_MAX);
int rlayers[1] = {29};
l = new tkDNN::Route (&net, rlayers, 1);
l = new tkDNN::Conv2d (&net, l->output_dim, 64, 1, 1, 1, 1, 0, 0, c26_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Reorg (&net, l->output_dim, 2); //48
tkDNN::Conv2d c18(&net, 1024, 3, 3, 1, 1, 1, 1, c18_bin, true);
tkDNN::Activation a18(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c19(&net, 512, 1, 1, 1, 1, 0, 0, c19_bin, true);
tkDNN::Activation a19(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c20(&net, 1024, 3, 3, 1, 1, 1, 1, c20_bin, true);
tkDNN::Activation a20(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c21(&net, 512, 1, 1, 1, 1, 0, 0, c21_bin, true);
tkDNN::Activation a21(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c22(&net, 1024, 3, 3, 1, 1, 1, 1, c22_bin, true);
tkDNN::Activation a22(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c23(&net, 1024, 3, 3, 1, 1, 1, 1, c23_bin, true);
tkDNN::Activation a23(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c24(&net, 1024, 3, 3, 1, 1, 1, 1, c24_bin, true);
tkDNN::Activation a24(&net, tkDNN::ACTIVATION_LEAKY);
int rlayers2[2] = {48,44};
l = new tkDNN::Route (&net, rlayers2, 2);
tkDNN::Layer *m25_layers[1] = { &a16 };
tkDNN::Route m25(&net, m25_layers, 1);
tkDNN::Conv2d c26(&net, 64, 1, 1, 1, 1, 0, 0, c26_bin, true);
tkDNN::Activation a26(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Reorg r27(&net, 2);
l = new tkDNN::Conv2d (&net, l->output_dim, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true);
l = new tkDNN::Activation (&net, l->output_dim, tkDNN::ACTIVATION_LEAKY);
l = new tkDNN::Conv2d (&net, l->output_dim, 425, 1, 1, 1, 1, 0, 0, c30_bin, false);
tkDNN::Layer *m28_layers[2] = { &r27, &a24 };
tkDNN::Route m28(&net, m28_layers, 2);
l = new tkDNN::Region (&net, l->output_dim, 80, 4, 5, 0.6f);
tkDNN::Conv2d c29(&net, 1024, 3, 3, 1, 1, 1, 1, c29_bin, true);
tkDNN::Activation a29(&net, tkDNN::ACTIVATION_LEAKY);
tkDNN::Conv2d c30(&net, 425, 1, 1, 1, 1, 0, 0, c30_bin, false);
tkDNN::Region g31(&net, 80, 4, 5, 0.6f);
// Load input
value_type *data;