Skip to content

GitLab

"torchvision/git@developer.sourcefind.cn:OpenDAS/vision.git" did not exist on "d6c7900d06c3388bf814cecbe90f91a9afecbefb"
Unverified Commit 12f1b3a3 authored by Adrià Arrufat's avatar Adrià Arrufat Committed by GitHub
Browse files

Refactor visitors into their own header (#2533) 

* Refactor visitors into their own header

* Update visitor links in term index

* Move documentation from layers_abstract
parent 736b4931
Hide whitespace changes
Inline Side-by-side
Showing with 613 additions and 623 deletions
dlib/cuda/cpu_dlib.h
View file @ 12f1b3a3
......@@ -8,7 +8,7 @@
#include "tensor.h"
#include "../geometry/rectangle.h"
#include "../dnn/misc.h"
#include "../dnn/utilities.h"
namespace dlib
{
......
dlib/cuda/cuda_dlib.h
View file @ 12f1b3a3
......@@ -6,7 +6,7 @@
#include "tensor.h"
#include "../geometry/rectangle.h"
#include "../dnn/misc.h"
#include "../dnn/utilities.h"
namespace dlib
{
......
dlib/dnn/layers.h
View file @ 12f1b3a3
......@@ -1838,23 +1838,6 @@ namespace dlib
};
}
template <typename net_type>
void set_all_bn_running_stats_window_sizes (
net_type& net,
unsigned long new_window_size
)
{
visit_layers(net, impl::visitor_bn_running_stats_window_size(new_window_size));
}
template <typename net_type>
void disable_duplicative_biases (
net_type& net
)
{
visit_layers(net, impl::visitor_disable_input_bias());
}
// ----------------------------------------------------------------------------------------
enum fc_bias_mode
......@@ -4428,83 +4411,6 @@ namespace dlib
template <typename SUBNET>
using reorg = add_layer<reorg_<2, 2>, SUBNET>;
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_fuse_layers
{
public:
template <typename T>
void fuse_convolution(T&) const
{
// disable other layer types
}
// handle the standard case (convolutional layer followed by affine;
template <long nf, long nr, long nc, int sy, int sx, int py, int px, typename U, typename E>
void fuse_convolution(add_layer<affine_, add_layer<con_<nf, nr, nc, sy, sx, py, px>, U>, E>& l)
{
if (l.layer_details().is_disabled())
return;
// get the convolution below the affine layer
auto& conv = l.subnet().layer_details();
// get the parameters from the affine layer as alias_tensor_instance
alias_tensor_instance gamma = l.layer_details().get_gamma();
alias_tensor_instance beta = l.layer_details().get_beta();
if (conv.bias_is_disabled())
{
conv.enable_bias();
}
tensor& params = conv.get_layer_params();
// update the biases
auto biases = alias_tensor(1, conv.num_filters());
biases(params, params.size() - conv.num_filters()) += mat(beta);
// guess the number of input channels
const long k_in = (params.size() - conv.num_filters()) / conv.num_filters() / conv.nr() / conv.nc();
// rescale the filters
DLIB_CASSERT(conv.num_filters() == gamma.k());
alias_tensor filter(1, k_in, conv.nr(), conv.nc());
const float* g = gamma.host();
for (long n = 0; n < conv.num_filters(); ++n)
{
filter(params, n * filter.size()) *= g[n];
}
// disable the affine layer
l.layer_details().disable();
}
template <typename input_layer_type>
void operator()(size_t , input_layer_type& ) const
{
// ignore other layers
}
template <typename T, typename U, typename E>
void operator()(size_t , add_layer<T, U, E>& l)
{
fuse_convolution(l);
}
};
}
template <typename net_type>
void fuse_layers (
net_type& net
)
{
DLIB_CASSERT(count_parameters(net) > 0, "The network has to be allocated before fusing the layers.");
visit_layers(net, impl::visitor_fuse_layers());
}
// ----------------------------------------------------------------------------------------
}
......
dlib/dnn/layers_abstract.h
View file @ 12f1b3a3
......@@ -1821,23 +1821,6 @@ namespace dlib
template <typename SUBNET>
using bn_fc = add_layer<bn_<FC_MODE>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <typename net_type>
void set_all_bn_running_stats_window_sizes (
const net_type& net,
unsigned long new_window_size
);
/*!
requires
- new_window_size > 0
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
ensures
- Sets the get_running_stats_window_size() field of all bn_ layers in net to
new_window_size.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
......@@ -3353,23 +3336,6 @@ namespace dlib
template <typename SUBNET>
using reorg = add_layer<reorg_<2, 2>, SUBNET>;
// ----------------------------------------------------------------------------------------
template <typename net_type>
void fuse_layers (
net_type& net
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- net has been properly allocated, that is: count_parameters(net) > 0.
ensures
- Disables all the affine_ layers that have a convolution as an input.
- Updates the convolution weights beneath the affine_ layers to produce the same
output as with the affine_ layers enabled.
!*/
// ----------------------------------------------------------------------------------------
}
......
dlib/dnn/loss.h
View file @ 12f1b3a3
......@@ -6,7 +6,6 @@
#include "loss_abstract.h"
#include "core.h"
#include "utilities.h"
#include "misc.h"
#include "../matrix.h"
#include "../cuda/tensor_tools.h"
#include "../geometry.h"
......
dlib/dnn/misc.h deleted 100644 → 0
View file @ 736b4931
// Copyright (C) 2020 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_DNn_MISC_h
#define DLIB_DNn_MISC_h
#include "../cuda/tensor.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
template <typename label_type>
struct weighted_label
{
weighted_label()
{}
weighted_label(label_type label, float weight = 1.f)
: label(label), weight(weight)
{}
label_type label{};
float weight = 1.f;
};
// ----------------------------------------------------------------------------------------
inline double log1pexp(double x)
{
using std::exp;
using namespace std; // Do this instead of using std::log1p because some compilers
// error out otherwise (E.g. gcc 4.9 in cygwin)
if (x <= -37)
return exp(x);
else if (-37 < x && x <= 18)
return log1p(exp(x));
else if (18 < x && x <= 33.3)
return x + exp(-x);
else
return x;
}
// ----------------------------------------------------------------------------------------
template <typename T>
T safe_log(T input, T epsilon = 1e-10)
{
// Prevent trying to calculate the logarithm of a very small number (let alone zero)
return std::log(std::max(input, epsilon));
}
// ----------------------------------------------------------------------------------------
static size_t tensor_index(
const tensor& t,
const long sample,
const long k,
const long r,
const long c
)
{
return ((sample * t.k() + k) * t.nr() + r) * t.nc() + c;
}
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_DNn_MISC_h
dlib/dnn/utilities.h
View file @ 12f1b3a3
......@@ -3,7 +3,7 @@
#ifndef DLIB_DNn_UTILITIES_H_
#define DLIB_DNn_UTILITIES_H_
#include "core.h"
#include "../cuda/tensor.h"
#include "utilities_abstract.h"
#include "../geometry.h"
#include <fstream>
......@@ -31,306 +31,61 @@ namespace dlib
// ----------------------------------------------------------------------------------------
namespace impl
template <typename label_type>
struct weighted_label
{
class visitor_net_to_xml
{
public:
visitor_net_to_xml(std::ostream& out_) : out(out_) {}
template<typename input_layer_type>
void operator()(size_t idx, const input_layer_type& l)
{
out << "<layer idx='"<<idx<<"' type='input'>\n";
to_xml(l,out);
out << "</layer>\n";
}
template <typename T, typename U>
void operator()(size_t idx, const add_loss_layer<T,U>& l)
{
out << "<layer idx='"<<idx<<"' type='loss'>\n";
to_xml(l.loss_details(),out);
out << "</layer>\n";
}
weighted_label()
{}
template <typename T, typename U, typename E>
void operator()(size_t idx, const add_layer<T,U,E>& l)
{
out << "<layer idx='"<<idx<<"' type='comp'>\n";
to_xml(l.layer_details(),out);
out << "</layer>\n";
}
weighted_label(label_type label, float weight = 1.f)
: label(label), weight(weight)
{}
template <unsigned long ID, typename U, typename E>
void operator()(size_t idx, const add_tag_layer<ID,U,E>& /*l*/)
{
out << "<layer idx='"<<idx<<"' type='tag' id='"<<ID<<"'/>\n";
}
template <template<typename> class T, typename U>
void operator()(size_t idx, const add_skip_layer<T,U>& /*l*/)
{
out << "<layer idx='"<<idx<<"' type='skip' id='"<<(tag_id<T>::id)<<"'/>\n";
}
private:
std::ostream& out;
};
}
template <typename net_type>
void net_to_xml (
const net_type& net,
std::ostream& out
)
{
auto old_precision = out.precision(9);
out << "<net>\n";
visit_layers(net, impl::visitor_net_to_xml(out));
out << "</net>\n";
// restore the original stream precision.
out.precision(old_precision);
}
template <typename net_type>
void net_to_xml (
const net_type& net,
const std::string& filename
)
{
std::ofstream fout(filename);
net_to_xml(net, fout);
}
label_type label{};
float weight = 1.f;
};
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_net_map_input_to_output
{
public:
visitor_net_map_input_to_output(dpoint& p_) : p(p_) {}
dpoint& p;
template<typename input_layer_type>
void operator()(const input_layer_type& )
{
}
template <typename T, typename U>
void operator()(const add_loss_layer<T,U>& net)
{
(*this)(net.subnet());
}
template <typename T, typename U, typename E>
void operator()(const add_layer<T,U,E>& net)
{
(*this)(net.subnet());
p = net.layer_details().map_input_to_output(p);
}
template <bool B, typename T, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_layer<T,U,E>,B>& net)
{
(*this)(net.subnet());
p = net.layer_details().map_input_to_output(p);
}
template <size_t N, template <typename> class R, typename U>
void operator()(const repeat<N, R, U>& net)
{
(*this)(net.subnet());
for (size_t i = 0; i < N; ++i)
{
(*this)(net.get_repeated_layer(N-1-i).subnet());
}
}
template <unsigned long ID, typename U, typename E>
void operator()(const add_tag_layer<ID,U,E>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <bool is_first, unsigned long ID, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_tag_layer<ID,U,E>,is_first>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <template<typename> class TAG_TYPE, typename U>
void operator()(const add_skip_layer<TAG_TYPE,U>& net)
{
(*this)(layer<TAG_TYPE>(net));
}
template <bool is_first, template<typename> class TAG_TYPE, typename SUBNET>
void operator()(const dimpl::subnet_wrapper<add_skip_layer<TAG_TYPE,SUBNET>,is_first>& net)
{
// skip layers are an identity transform, so do nothing
(*this)(layer<TAG_TYPE>(net));
}
};
class visitor_net_map_output_to_input
{
public:
visitor_net_map_output_to_input(dpoint& p_) : p(p_) {}
dpoint& p;
template<typename input_layer_type>
void operator()(const input_layer_type& )
{
}
template <typename T, typename U>
void operator()(const add_loss_layer<T,U>& net)
{
(*this)(net.subnet());
}
template <typename T, typename U, typename E>
void operator()(const add_layer<T,U,E>& net)
{
p = net.layer_details().map_output_to_input(p);
(*this)(net.subnet());
}
template <bool B, typename T, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_layer<T,U,E>,B>& net)
{
p = net.layer_details().map_output_to_input(p);
(*this)(net.subnet());
}
template <size_t N, template <typename> class R, typename U>
void operator()(const repeat<N, R, U>& net)
{
for (size_t i = 0; i < N; ++i)
{
(*this)(net.get_repeated_layer(i).subnet());
}
(*this)(net.subnet());
}
template <unsigned long ID, typename U, typename E>
void operator()(const add_tag_layer<ID,U,E>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <bool is_first, unsigned long ID, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_tag_layer<ID,U,E>,is_first>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <template<typename> class TAG_TYPE, typename U>
void operator()(const add_skip_layer<TAG_TYPE,U>& net)
{
(*this)(layer<TAG_TYPE>(net));
}
template <bool is_first, template<typename> class TAG_TYPE, typename SUBNET>
void operator()(const dimpl::subnet_wrapper<add_skip_layer<TAG_TYPE,SUBNET>,is_first>& net)
{
// skip layers are an identity transform, so do nothing
(*this)(layer<TAG_TYPE>(net));
}
};
}
template <typename net_type>
inline dpoint input_tensor_to_output_tensor(
const net_type& net,
dpoint p
)
{
impl::visitor_net_map_input_to_output temp(p);
temp(net);
return p;
}
template <typename net_type>
inline dpoint output_tensor_to_input_tensor(
const net_type& net,
dpoint p
)
inline double log1pexp(double x)
{
impl::visitor_net_map_output_to_input temp(p);
temp(net);
return p;
using std::exp;
using namespace std; // Do this instead of using std::log1p because some compilers
// error out otherwise (E.g. gcc 4.9 in cygwin)
if (x <= -37)
return exp(x);
else if (-37 < x && x <= 18)
return log1p(exp(x));
else if (18 < x && x <= 33.3)
return x + exp(-x);
else
return x;
}
// ----------------------------------------------------------------------------------------
template <typename net_type>
size_t count_parameters(
const net_type& net
)
template <typename T>
T safe_log(T input, T epsilon = 1e-10)
{
size_t num_parameters = 0;
visit_layer_parameters(net, [&](const tensor& t) { num_parameters += t.size(); });
return num_parameters;
// Prevent trying to calculate the logarithm of a very small number (let alone zero)
return std::log(std::max(input, epsilon));
}
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_learning_rate_multiplier
{
public:
visitor_learning_rate_multiplier(double new_learning_rate_multiplier_) :
new_learning_rate_multiplier(new_learning_rate_multiplier_) {}
template <typename layer>
void operator()(layer& l) const
{
set_learning_rate_multiplier(l, new_learning_rate_multiplier);
}
private:
double new_learning_rate_multiplier;
};
}
template <typename net_type>
void set_all_learning_rate_multipliers(
net_type& net,
double learning_rate_multiplier
static size_t tensor_index(
const tensor& t,
const long sample,
const long k,
const long r,
const long c
)
{
DLIB_CASSERT(learning_rate_multiplier >= 0);
impl::visitor_learning_rate_multiplier temp(learning_rate_multiplier);
visit_computational_layers(net, temp);
}
template <size_t begin, size_t end, typename net_type>
void set_learning_rate_multipliers_range(
net_type& net,
double learning_rate_multiplier
)
{
static_assert(begin <= end, "Invalid range");
static_assert(end <= net_type::num_layers, "Invalid range");
DLIB_CASSERT(learning_rate_multiplier >= 0);
impl::visitor_learning_rate_multiplier temp(learning_rate_multiplier);
visit_computational_layers_range<begin, end>(net, temp);
return ((sample * t.k() + k) * t.nr() + r) * t.nc() + c;
}
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_DNn_UTILITIES_H_
......
dlib/dnn/utilities_abstract.h
View file @ 12f1b3a3
......@@ -18,8 +18,6 @@ namespace dlib
ensures
- returns log(1+exp(x))
(except computes it using a numerically accurate method)
NOTE: For technical reasons, it is defined in misc.h.
!*/
// ----------------------------------------------------------------------------------------
......@@ -41,136 +39,6 @@ namespace dlib
layer that uses params as its parameters.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
void net_to_xml (
const net_type& net,
std::ostream& out
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- All layers in the net must provide to_xml() functions.
ensures
- Prints the given neural network object as an XML document to the given output
stream.
!*/
template <typename net_type>
void net_to_xml (
const net_type& net,
const std::string& filename
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- All layers in the net must provide to_xml() functions.
ensures
- This function is just like the above net_to_xml(), except it writes to a file
rather than an ostream.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
dpoint input_tensor_to_output_tensor(
const net_type& net,
dpoint p
);
/*!
requires
- net_type is an object of type add_layer, add_skip_layer, or add_tag_layer.
- All layers in the net must provide map_input_to_output() functions.
ensures
- Given a dpoint (i.e. a row,column coordinate) in the input tensor given to
net, this function returns the corresponding dpoint in the output tensor
net.get_output(). This kind of mapping is useful when working with fully
convolutional networks as you will often want to know what parts of the
output feature maps correspond to what parts of the input.
- If the network contains skip layers then any layers skipped over by the skip
layer are ignored for the purpose of computing this coordinate mapping. That
is, if you walk the network from the output layer to the input layer, where
each time you encounter a skip layer you jump to the layer indicated by the
skip layer, you will visit exactly the layers in the network involved in the
input_tensor_to_output_tensor() calculation. This behavior is useful since it
allows you to compute some auxiliary DNN as a separate branch of computation
that is separate from the main network's job of running some kind of fully
convolutional network over an image. For instance, you might want to have a
branch in your network that computes some global image level
summarization/feature.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
dpoint output_tensor_to_input_tensor(
const net_type& net,
dpoint p
);
/*!
requires
- net_type is an object of type add_layer, add_skip_layer, or add_tag_layer.
- All layers in the net must provide map_output_to_input() functions.
ensures
- This function provides the reverse mapping of input_tensor_to_output_tensor().
That is, given a dpoint in net.get_output(), what is the corresponding dpoint
in the input tensor?
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
inline size_t count_parameters(
const net_type& net
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
ensures
- Returns the number of allocated parameters in the network. E.g. if the network has not
been trained then, since nothing has been allocated yet, it will return 0.
!*/
// ----------------------------------------------------------------------------------------
template<typename net_type>
void set_all_learning_rate_multipliers(
net_type& net,
double learning_rate_multiplier
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- learning_rate_multiplier >= 0
ensures
- Sets all learning_rate_multipliers and bias_learning_rate_multipliers in net
to learning_rate_multiplier.
!*/
// ----------------------------------------------------------------------------------------
template <size_t begin, size_t end, typename net_type>
void set_learning_rate_multipliers_range(
net_type& net,
double learning_rate_multiplier
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- learning_rate_multiplier >= 0
- begin <= end <= net_type::num_layers
ensures
- Loops over the layers in the range [begin,end) in net and calls
set_learning_rate_multiplier on them with the value of
learning_rate_multiplier.
!*/
// ----------------------------------------------------------------------------------------
}
......
dlib/dnn/visitors.h
View file @ 12f1b3a3
......@@ -9,6 +9,408 @@
namespace dlib
{
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_net_map_input_to_output
{
public:
visitor_net_map_input_to_output(dpoint& p_) : p(p_) {}
dpoint& p;
template<typename input_layer_type>
void operator()(const input_layer_type& )
{
}
template <typename T, typename U>
void operator()(const add_loss_layer<T,U>& net)
{
(*this)(net.subnet());
}
template <typename T, typename U, typename E>
void operator()(const add_layer<T,U,E>& net)
{
(*this)(net.subnet());
p = net.layer_details().map_input_to_output(p);
}
template <bool B, typename T, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_layer<T,U,E>,B>& net)
{
(*this)(net.subnet());
p = net.layer_details().map_input_to_output(p);
}
template <size_t N, template <typename> class R, typename U>
void operator()(const repeat<N, R, U>& net)
{
(*this)(net.subnet());
for (size_t i = 0; i < N; ++i)
{
(*this)(net.get_repeated_layer(N-1-i).subnet());
}
}
template <unsigned long ID, typename U, typename E>
void operator()(const add_tag_layer<ID,U,E>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <bool is_first, unsigned long ID, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_tag_layer<ID,U,E>,is_first>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <template<typename> class TAG_TYPE, typename U>
void operator()(const add_skip_layer<TAG_TYPE,U>& net)
{
(*this)(layer<TAG_TYPE>(net));
}
template <bool is_first, template<typename> class TAG_TYPE, typename SUBNET>
void operator()(const dimpl::subnet_wrapper<add_skip_layer<TAG_TYPE,SUBNET>,is_first>& net)
{
// skip layers are an identity transform, so do nothing
(*this)(layer<TAG_TYPE>(net));
}
};
class visitor_net_map_output_to_input
{
public:
visitor_net_map_output_to_input(dpoint& p_) : p(p_) {}
dpoint& p;
template<typename input_layer_type>
void operator()(const input_layer_type& )
{
}
template <typename T, typename U>
void operator()(const add_loss_layer<T,U>& net)
{
(*this)(net.subnet());
}
template <typename T, typename U, typename E>
void operator()(const add_layer<T,U,E>& net)
{
p = net.layer_details().map_output_to_input(p);
(*this)(net.subnet());
}
template <bool B, typename T, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_layer<T,U,E>,B>& net)
{
p = net.layer_details().map_output_to_input(p);
(*this)(net.subnet());
}
template <size_t N, template <typename> class R, typename U>
void operator()(const repeat<N, R, U>& net)
{
for (size_t i = 0; i < N; ++i)
{
(*this)(net.get_repeated_layer(i).subnet());
}
(*this)(net.subnet());
}
template <unsigned long ID, typename U, typename E>
void operator()(const add_tag_layer<ID,U,E>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <bool is_first, unsigned long ID, typename U, typename E>
void operator()(const dimpl::subnet_wrapper<add_tag_layer<ID,U,E>,is_first>& net)
{
// tag layers are an identity transform, so do nothing
(*this)(net.subnet());
}
template <template<typename> class TAG_TYPE, typename U>
void operator()(const add_skip_layer<TAG_TYPE,U>& net)
{
(*this)(layer<TAG_TYPE>(net));
}
template <bool is_first, template<typename> class TAG_TYPE, typename SUBNET>
void operator()(const dimpl::subnet_wrapper<add_skip_layer<TAG_TYPE,SUBNET>,is_first>& net)
{
// skip layers are an identity transform, so do nothing
(*this)(layer<TAG_TYPE>(net));
}
};
}
template <typename net_type>
inline dpoint input_tensor_to_output_tensor(
const net_type& net,
dpoint p
)
{
impl::visitor_net_map_input_to_output temp(p);
temp(net);
return p;
}
template <typename net_type>
inline dpoint output_tensor_to_input_tensor(
const net_type& net,
dpoint p
)
{
impl::visitor_net_map_output_to_input temp(p);
temp(net);
return p;
}
// ----------------------------------------------------------------------------------------
template <typename net_type>
size_t count_parameters(
const net_type& net
)
{
size_t num_parameters = 0;
visit_layer_parameters(net, [&](const tensor& t) { num_parameters += t.size(); });
return num_parameters;
}
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_learning_rate_multiplier
{
public:
visitor_learning_rate_multiplier(double new_learning_rate_multiplier_) :
new_learning_rate_multiplier(new_learning_rate_multiplier_) {}
template <typename layer>
void operator()(layer& l) const
{
set_learning_rate_multiplier(l, new_learning_rate_multiplier);
}
private:
double new_learning_rate_multiplier;
};
}
template <typename net_type>
void set_all_learning_rate_multipliers(
net_type& net,
double learning_rate_multiplier
)
{
DLIB_CASSERT(learning_rate_multiplier >= 0);
impl::visitor_learning_rate_multiplier temp(learning_rate_multiplier);
visit_computational_layers(net, temp);
}
template <size_t begin, size_t end, typename net_type>
void set_learning_rate_multipliers_range(
net_type& net,
double learning_rate_multiplier
)
{
static_assert(begin <= end, "Invalid range");
static_assert(end <= net_type::num_layers, "Invalid range");
DLIB_CASSERT(learning_rate_multiplier >= 0);
impl::visitor_learning_rate_multiplier temp(learning_rate_multiplier);
visit_computational_layers_range<begin, end>(net, temp);
}
// ----------------------------------------------------------------------------------------
template <typename net_type>
void set_all_bn_running_stats_window_sizes (
net_type& net,
unsigned long new_window_size
)
{
visit_layers(net, impl::visitor_bn_running_stats_window_size(new_window_size));
}
template <typename net_type>
void disable_duplicative_biases (
net_type& net
)
{
visit_layers(net, impl::visitor_disable_input_bias());
}
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_fuse_layers
{
public:
template <typename T>
void fuse_convolution(T&) const
{
// disable other layer types
}
// handle the standard case (convolutional layer followed by affine;
template <long nf, long nr, long nc, int sy, int sx, int py, int px, typename U, typename E>
void fuse_convolution(add_layer<affine_, add_layer<con_<nf, nr, nc, sy, sx, py, px>, U>, E>& l)
{
if (l.layer_details().is_disabled())
return;
// get the convolution below the affine layer
auto& conv = l.subnet().layer_details();
// get the parameters from the affine layer as alias_tensor_instance
alias_tensor_instance gamma = l.layer_details().get_gamma();
alias_tensor_instance beta = l.layer_details().get_beta();
if (conv.bias_is_disabled())
{
conv.enable_bias();
}
tensor& params = conv.get_layer_params();
// update the biases
auto biases = alias_tensor(1, conv.num_filters());
biases(params, params.size() - conv.num_filters()) += mat(beta);
// guess the number of input channels
const long k_in = (params.size() - conv.num_filters()) / conv.num_filters() / conv.nr() / conv.nc();
// rescale the filters
DLIB_CASSERT(conv.num_filters() == gamma.k());
alias_tensor filter(1, k_in, conv.nr(), conv.nc());
const float* g = gamma.host();
for (long n = 0; n < conv.num_filters(); ++n)
{
filter(params, n * filter.size()) *= g[n];
}
// disable the affine layer
l.layer_details().disable();
}
template <typename input_layer_type>
void operator()(size_t , input_layer_type& ) const
{
// ignore other layers
}
template <typename T, typename U, typename E>
void operator()(size_t , add_layer<T, U, E>& l)
{
fuse_convolution(l);
}
};
}
template <typename net_type>
void fuse_layers (
net_type& net
)
{
DLIB_CASSERT(count_parameters(net) > 0, "The network has to be allocated before fusing the layers.");
visit_layers(net, impl::visitor_fuse_layers());
}
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_net_to_xml
{
public:
visitor_net_to_xml(std::ostream& out_) : out(out_) {}
template<typename input_layer_type>
void operator()(size_t idx, const input_layer_type& l)
{
out << "<layer idx='"<<idx<<"' type='input'>\n";
to_xml(l,out);
out << "</layer>\n";
}
template <typename T, typename U>
void operator()(size_t idx, const add_loss_layer<T,U>& l)
{
out << "<layer idx='"<<idx<<"' type='loss'>\n";
to_xml(l.loss_details(),out);
out << "</layer>\n";
}
template <typename T, typename U, typename E>
void operator()(size_t idx, const add_layer<T,U,E>& l)
{
out << "<layer idx='"<<idx<<"' type='comp'>\n";
to_xml(l.layer_details(),out);
out << "</layer>\n";
}
template <unsigned long ID, typename U, typename E>
void operator()(size_t idx, const add_tag_layer<ID,U,E>& /*l*/)
{
out << "<layer idx='"<<idx<<"' type='tag' id='"<<ID<<"'/>\n";
}
template <template<typename> class T, typename U>
void operator()(size_t idx, const add_skip_layer<T,U>& /*l*/)
{
out << "<layer idx='"<<idx<<"' type='skip' id='"<<(tag_id<T>::id)<<"'/>\n";
}
private:
std::ostream& out;
};
}
template <typename net_type>
void net_to_xml (
const net_type& net,
std::ostream& out
)
{
auto old_precision = out.precision(9);
out << "<net>\n";
visit_layers(net, impl::visitor_net_to_xml(out));
out << "</net>\n";
// restore the original stream precision.
out.precision(old_precision);
}
template <typename net_type>
void net_to_xml (
const net_type& net,
const std::string& filename
)
{
std::ofstream fout(filename);
net_to_xml(net, fout);
}
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
namespace impl
{
class visitor_net_to_dot
......
dlib/dnn/visitors_abstract.h
View file @ 12f1b3a3
......@@ -10,6 +10,170 @@
namespace dlib
{
// ----------------------------------------------------------------------------------------
template <typename net_type>
void set_all_bn_running_stats_window_sizes (
const net_type& net,
unsigned long new_window_size
);
/*!
requires
- new_window_size > 0
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
ensures
- Sets the get_running_stats_window_size() field of all bn_ layers in net to
new_window_size.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
void fuse_layers (
net_type& net
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- net has been properly allocated, that is: count_parameters(net) > 0.
ensures
- Disables all the affine_ layers that have a convolution as an input.
- Updates the convolution weights beneath the affine_ layers to produce the same
output as with the affine_ layers enabled.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
void net_to_xml (
const net_type& net,
std::ostream& out
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- All layers in the net must provide to_xml() functions.
ensures
- Prints the given neural network object as an XML document to the given output
stream.
!*/
template <typename net_type>
void net_to_xml (
const net_type& net,
const std::string& filename
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- All layers in the net must provide to_xml() functions.
ensures
- This function is just like the above net_to_xml(), except it writes to a file
rather than an ostream.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
dpoint input_tensor_to_output_tensor(
const net_type& net,
dpoint p
);
/*!
requires
- net_type is an object of type add_layer, add_skip_layer, or add_tag_layer.
- All layers in the net must provide map_input_to_output() functions.
ensures
- Given a dpoint (i.e. a row,column coordinate) in the input tensor given to
net, this function returns the corresponding dpoint in the output tensor
net.get_output(). This kind of mapping is useful when working with fully
convolutional networks as you will often want to know what parts of the
output feature maps correspond to what parts of the input.
- If the network contains skip layers then any layers skipped over by the skip
layer are ignored for the purpose of computing this coordinate mapping. That
is, if you walk the network from the output layer to the input layer, where
each time you encounter a skip layer you jump to the layer indicated by the
skip layer, you will visit exactly the layers in the network involved in the
input_tensor_to_output_tensor() calculation. This behavior is useful since it
allows you to compute some auxiliary DNN as a separate branch of computation
that is separate from the main network's job of running some kind of fully
convolutional network over an image. For instance, you might want to have a
branch in your network that computes some global image level
summarization/feature.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
dpoint output_tensor_to_input_tensor(
const net_type& net,
dpoint p
);
/*!
requires
- net_type is an object of type add_layer, add_skip_layer, or add_tag_layer.
- All layers in the net must provide map_output_to_input() functions.
ensures
- This function provides the reverse mapping of input_tensor_to_output_tensor().
That is, given a dpoint in net.get_output(), what is the corresponding dpoint
in the input tensor?
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
inline size_t count_parameters(
const net_type& net
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
ensures
- Returns the number of allocated parameters in the network. E.g. if the network has not
been trained then, since nothing has been allocated yet, it will return 0.
!*/
// ----------------------------------------------------------------------------------------
template<typename net_type>
void set_all_learning_rate_multipliers(
net_type& net,
double learning_rate_multiplier
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- learning_rate_multiplier >= 0
ensures
- Sets all learning_rate_multipliers and bias_learning_rate_multipliers in net
to learning_rate_multiplier.
!*/
// ----------------------------------------------------------------------------------------
template <size_t begin, size_t end, typename net_type>
void set_learning_rate_multipliers_range(
net_type& net,
double learning_rate_multiplier
);
/*!
requires
- net_type is an object of type add_layer, add_loss_layer, add_skip_layer, or
add_tag_layer.
- learning_rate_multiplier >= 0
- begin <= end <= net_type::num_layers
ensures
- Loops over the layers in the range [begin,end) in net and calls
set_learning_rate_multiplier on them with the value of
learning_rate_multiplier.
!*/
// ----------------------------------------------------------------------------------------
template <typename net_type>
......
docs/docs/term_index.xml
View file @ 12f1b3a3
......@@ -34,14 +34,18 @@
<term file="dlib/algs.h.html" name="stack_based_memory_block" include="dlib/algs.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="net_to_xml" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="log1pexp" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="randomize_parameters" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="input_tensor_to_output_tensor" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="output_tensor_to_input_tensor" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="count_parameters" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="set_all_learning_rate_multipliers" include="dlib/dnn.h"/>
<term file="dlib/dnn/utilities_abstract.h.html" name="set_learning_rate_multipliers_range" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="set_all_bn_running_stats_window_sizes" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="disable_duplicative_biases" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="fuse_layers" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="net_to_xml" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="net_to_dot" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="input_tensor_to_output_tensor" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="output_tensor_to_input_tensor" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="count_parameters" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="set_all_learning_rate_multipliers" include="dlib/dnn.h"/>
<term file="dlib/dnn/visitors_abstract.h.html" name="set_learning_rate_multipliers_range" include="dlib/dnn.h"/>
<term file="dlib/dnn/core_abstract.h.html" name="tuple_head" include="dlib/dnn.h"/>
<term file="dlib/dnn/core_abstract.h.html" name="tuple_tail" include="dlib/dnn.h"/>
<term file="dlib/dnn/core_abstract.h.html" name="get_learning_rate_multiplier" include="dlib/dnn.h"/>
......@@ -103,9 +107,6 @@
<term file="dlib/dnn/layers_abstract.h.html" name="layer_mode" include="dlib/dnn.h"/>
<term file="dlib/dnn/layers_abstract.h.html" name="CONV_MODE" include="dlib/dnn.h"/>
<term file="dlib/dnn/layers_abstract.h.html" name="FC_MODE" include="dlib/dnn.h"/>
<term file="dlib/dnn/layers_abstract.h.html" name="set_all_bn_running_stats_window_sizes" include="dlib/dnn.h"/>
<term file="dlib/dnn/layers_abstract.h.html" name="disable_duplicative_biases" include="dlib/dnn.h"/>
<term file="dlib/dnn/layers_abstract.h.html" name="fuse_layers" include="dlib/dnn.h"/>
<term file="dlib/cuda/tensor_abstract.h.html" name="tensor" include="dlib/cuda/tensor.h"/>
<term file="dlib/cuda/tensor_abstract.h.html" name="resizable_tensor" include="dlib/cuda/tensor.h"/>
<term file="dlib/cuda/tensor_abstract.h.html" name="alias_tensor_instance" include="dlib/cuda/tensor.h"/>
......
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