Skip to content

D
dlib
Commit ecb7095e authored by OranjeeGeneral's avatar OranjeeGeneral
Browse files
Options

refactored interface to reduce complexity so conv and convt layers forward… 

refactored interface to reduce complexity so conv and convt layers forward passes have to call setup explicit now and there is only one ()-operator
parent 3a91295e
Hide whitespace changes
Inline Side-by-side
Showing with 27 additions and 140 deletions
dlib/dnn/cpu_dlib.cpp
View file @ ecb7095e
......@@ -1739,65 +1739,32 @@ namespace dlib
}
}
void tensor_conv::forward_without_setup(
resizable_tensor& output,
const tensor& data,
const tensor& filters
)
{
DLIB_CASSERT(is_same_object(output,data) == false);
DLIB_CASSERT(is_same_object(output,filters) == false);
DLIB_CASSERT(filters.k() == data.k());
output.set_size(data.num_samples(),
filters.num_samples(),
1+(data.nr()+2*last_padding_y-filters.nr())/last_stride_y,
1+(data.nc()+2*last_padding_x-filters.nc())/last_stride_x);
matrix<float> temp;
for (long n = 0; n < data.num_samples(); ++n)
{
img2col(temp, data, n, filters.nr(), filters.nc(), last_stride_y, last_stride_x, last_padding_y, last_padding_x);
output.set_sample(n, mat(filters)*trans(temp));
}
}
void tensor_conv::operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters,
int stride_y,
int stride_x,
int padding_y,
int padding_x
const tensor& filters
)
{
DLIB_CASSERT(is_same_object(output,data) == false);
DLIB_CASSERT(is_same_object(output,filters) == false);
DLIB_CASSERT(filters.k() == data.k());
DLIB_CASSERT(stride_y > 0 && stride_x > 0);
DLIB_CASSERT(0 <= padding_y && padding_y < filters.nr());
DLIB_CASSERT(0 <= padding_x && padding_x < filters.nc());
DLIB_CASSERT(filters.nr() <= data.nr() + 2*padding_y,
DLIB_CASSERT(filters.nr() <= data.nr() + 2*last_padding_y,
"Filter windows must be small enough to fit into the padded image.");
DLIB_CASSERT(filters.nc() <= data.nc() + 2*padding_x,
DLIB_CASSERT(filters.nc() <= data.nc() + 2*last_padding_x,
"Filter windows must be small enough to fit into the padded image.");
output.set_size(data.num_samples(),
filters.num_samples(),
1+(data.nr()+2*padding_y-filters.nr())/stride_y,
1+(data.nc()+2*padding_x-filters.nc())/stride_x);
1+(data.nr()+2*last_padding_y-filters.nr())/last_stride_y,
1+(data.nc()+2*last_padding_x-filters.nc())/last_stride_x);
matrix<float> temp;
for (long n = 0; n < data.num_samples(); ++n)
{
img2col(temp, data, n, filters.nr(), filters.nc(), stride_y, stride_x, padding_y, padding_x);
img2col(temp, data, n, filters.nr(), filters.nc(), last_stride_y, last_stride_x, last_padding_y, last_padding_x);
output.set_sample(n, mat(filters)*trans(temp));
}
last_stride_y = stride_y;
last_stride_x = stride_x;
last_padding_y = padding_y;
last_padding_x = padding_x;
}
// ------------------------------------------------------------------------------------
......
dlib/dnn/cpu_dlib.h
View file @ ecb7095e
......@@ -378,29 +378,21 @@ namespace dlib
)
{
(void)data; /* silence compiler */
(void)filters; /* silence compiler */
DLIB_CASSERT(stride_y > 0 && stride_x > 0);
DLIB_CASSERT(0 <= padding_y && padding_y < filters.nr());
DLIB_CASSERT(0 <= padding_x && padding_x < filters.nc());
last_stride_y = stride_y;
last_stride_x = stride_x;
last_padding_y = padding_y;
last_padding_x = padding_x;
}
void forward_without_setup(
void operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters
);
void operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters,
int stride_y,
int stride_x,
int padding_y,
int padding_x
);
void get_gradient_for_data (
const tensor& gradient_input,
const tensor& filters,
......
dlib/dnn/cudnn_dlibapi.cpp
View file @ ecb7095e
......@@ -950,63 +950,10 @@ namespace dlib
clear();
}
void tensor_conv::forward_without_setup(
resizable_tensor& output,
const tensor& data,
const tensor& filters
)
{
DLIB_CASSERT(is_same_object(output,data) == false);
DLIB_CASSERT(is_same_object(output,filters) == false);
DLIB_CASSERT(filters.k() == data.k());
DLIB_CASSERT(filters.nc() <= data.nc() + 2*padding_x,
"Filter windows must be small enough to fit into the padded image."
<< "\n\t filters.nc(): " << filters.nc()
<< "\n\t data.nc(): " << data.nc()
<< "\n\t padding_x: " << padding_x
);
DLIB_CASSERT(filters.nr() <= data.nr() + 2*padding_y,
"Filter windows must be small enough to fit into the padded image."
<< "\n\t filters.nr(): " << filters.nr()
<< "\n\t data.nr(): " << data.nr()
<< "\n\t padding_y: " << padding_y
);
output.set_size(out_num_samples, out_k, out_nr, out_nc);
DLIB_ASSERT(output.num_samples() == data.num_samples(),out_num_samples << " " << data.num_samples());
DLIB_ASSERT(output.k() == filters.num_samples());
DLIB_ASSERT(output.nr() == 1+(data.nr()+2*padding_y-filters.nr())/stride_y);
DLIB_ASSERT(output.nc() == 1+(data.nc()+2*padding_x-filters.nc())/stride_x);
const float alpha = 1;
const float beta = 0;
CHECK_CUDNN(cudnnConvolutionForward(
context(),
&alpha,
descriptor(data),
data.device(),
(const cudnnFilterDescriptor_t)filter_handle,
filters.device(),
(const cudnnConvolutionDescriptor_t)conv_handle,
(cudnnConvolutionFwdAlgo_t)forward_algo,
forward_workspace,
forward_workspace_size_in_bytes,
&beta,
descriptor(output),
output.device()));
}
void tensor_conv::operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters,
int stride_y,
int stride_x,
int padding_y,
int padding_x
const tensor& filters
)
{
DLIB_CASSERT(is_same_object(output,data) == false);
......@@ -1027,8 +974,6 @@ namespace dlib
);
setup(data,filters,stride_y,stride_x,padding_y,padding_x);
output.set_size(out_num_samples, out_k, out_nr, out_nc);
DLIB_ASSERT(output.num_samples() == data.num_samples(),out_num_samples << " " << data.num_samples());
......
dlib/dnn/cudnn_dlibapi.h
View file @ ecb7095e
......@@ -205,11 +205,7 @@ namespace dlib
void operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters,
int stride_y,
int stride_x,
int padding_y,
int padding_x
const tensor& filters
);
/*!
requires
......@@ -279,11 +275,6 @@ namespace dlib
int padding_x
);
void forward_without_setup(
resizable_tensor& output,
const tensor& data,
const tensor& filters);
private:
/*!
......
dlib/dnn/layers.h
View file @ ecb7095e
......@@ -147,14 +147,16 @@ namespace dlib
template <typename SUBNET>
void forward(const SUBNET& sub, resizable_tensor& output)
{
conv.setup(sub.get_output(),
filters(params,0),
_stride_y,
_stride_x,
padding_y_,
padding_x_);
conv(output,
sub.get_output(),
filters(params,0),
_stride_y,
_stride_x,
padding_y_,
padding_x_
);
filters(params,0));
tt::add(1,output,1,biases(params,filters.size()));
}
......
dlib/dnn/tensor_tools.h
View file @ ecb7095e
......@@ -879,12 +879,8 @@ namespace dlib { namespace tt
void operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters,
int stride_y,
int stride_x,
int padding_y,
int padding_x
) { impl(output,data,filters,stride_y,stride_x,padding_y,padding_x); }
const tensor& filters
) { impl(output,data,filters); }
/*!
requires
- stride_y > 0
......@@ -947,15 +943,7 @@ namespace dlib { namespace tt
this gradient to filters_gradient.
!*/
void operator() (
resizable_tensor& output,
const tensor& data,
const tensor& filters
)
{
impl.forward_without_setup(output,data,filters);
}
void setup(
const tensor& data,
const tensor& filters,
......
dlib/test/dnn.cpp
View file @ ecb7095e
......@@ -805,8 +805,10 @@ namespace
padding_y = (filters.nr()-data.nr()+1)/2;
if (!(filters.nc() <= data.nc() + 2*padding_x))
padding_x = (filters.nc()-data.nc()+1)/2;
conv1(output1, data, filters, stride_y,stride_x, padding_y, padding_x);
conv2(output2, data, filters, stride_y,stride_x, padding_y, padding_x);
conv1.setup(data,filters,stride_y,stride_x,padding_y,padding_x);
conv1(output1, data, filters);
conv2.setup(data,filters,stride_y,stride_x,padding_y,padding_x);
conv2(output2, data, filters);
dlog << LINFO << "forward error: "<< max(abs(mat(output1)-mat(output2)));
DLIB_TEST_MSG(max(abs(mat(output1)-mat(output2))) < 1e-3, max(abs(mat(output1)-mat(output2)))
<<"\n\t padding_y: "<< padding_y
......
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