OrgLion-ML
/
dlib
mirror of https://github.com/davisking/dlib.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Made the input_tensor_to_output_tensor() and output_tensor_to_input_tensor() 

coordinate mappings work on networks that contain skip layers.
This commit is contained in:
Davis King 2017-08-20 19:42:12 -04:00
parent 620178db05
commit dc071ceae1
2 changed files with 109 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

104
dlib/dnn/utilities.h
View File

@ -135,11 +135,57 @@ namespace dlib
point& p;
template<typename layer_type>
void operator()(size_t idx, const layer_type& net)
template<typename input_layer_type>
void operator()(const input_layer_type& net)
{
}
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 <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
@ -149,11 +195,57 @@ namespace dlib
point& p;
template<typename layer_type>
void operator()(size_t idx, const layer_type& net)
template<typename input_layer_type>
void operator()(const input_layer_type& net)
{
}
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 <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));
}
};
}
@ -164,7 +256,7 @@ namespace dlib
)
{
impl::visitor_net_map_input_to_output temp(p);
visit_layers_backwards_range<0,net_type::num_layers-1>(net, temp);
temp(net);
return p;
}
@ -175,7 +267,7 @@ namespace dlib
)
{
impl::visitor_net_map_output_to_input temp(p);
visit_layers_range<0,net_type::num_layers-1>(net, temp);
temp(net);
return p;
}

11
dlib/dnn/utilities_abstract.h
View File

@ -88,6 +88,17 @@ namespace dlib
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.
!*/
// ----------------------------------------------------------------------------------------