Skip to content

D
dlib
Commit e679d66a authored by Davis King's avatar Davis King
Browse files
Options

filled out the trainer spec

parent 0ad2cb71
Hide whitespace changes
Inline Side-by-side
Showing with 125 additions and 4 deletions
dlib/dnn/trainer_abstract.h
View file @ e679d66a
......@@ -28,6 +28,10 @@ namespace dlib
in solvers_abstract.h
WHAT THIS OBJECT REPRESENTS
This object is a tool training a deep neural network. To use it you supply
a neural network type and a solver, then you call train() with your
training data and it will output a new network instance that has hopefully
learned something useful from your training data.
!*/
......@@ -38,46 +42,129 @@ namespace dlib
dnn_trainer(
);
/*!
ensures
- #get_net() == a default initialized net_type object.
- #get_solvers() == a set of default initialized solvers.
!*/
explicit dnn_trainer(
const net_type& net
);
/*!
ensures
- #get_net() == net
- #get_solvers() == a set of default initialized solvers.
!*/
dnn_trainer(
const net_type& net,
const solver_type& solver
);
/*!
ensures
- #get_net() == net
- #get_solvers() == a set of solvers that are all initialized with the
provided solver instance.
!*/
const net_type& get_net (
) const;
/*!
ensures
- returns the neural network object in this trainer. This is the network
that is optimized when you call train().
!*/
void set_net (
const net_type& net
);
/*!
ensures
- #get_net() == net
!*/
void set_solver (
const solver_type& solver_
const solver_type& solver
);
/*!
ensures
- assigns solver to all the solvers in this object. I.e. solver will be
assigned to each element in get_solvers().
!*/
const sstack<solver_type,net_type::num_layers>& get_solvers (
) const;
/*!
ensures
- returns the solvers used to optimize each layer of the neural network
get_net(). In particular, the first layer's solver is
get_solvers().top(), the second layer's solver is
get_solvers().pop().top(), and so on.
!*/
sstack<solver_type,net_type::num_layers>& get_solvers (
);
/*!
ensures
- returns the solvers used to optimize each layer of the neural network
get_net(). In particular, the first layer's solver is
get_solvers().top(), the second layer's solver is
get_solvers().pop().top(), and so on.
!*/
unsigned long get_mini_batch_size (
) const;
/*!
ensures
- During training, we call the network's update() routine over and over
with training data. The number of training samples we give to each call
to update is the "mini-batch size", which is defined by
get_mini_batch_size().
!*/
void set_mini_batch_size (
unsigned long batch_size
);
/*!
requires
- batch_size > 0
ensures
- #get_mini_batch_size() == batch_size
!*/
unsigned long get_num_epochs (
) const;
/*!
ensures
- Returns the number of passes over the training data we will execute when
train() is called.
!*/
void set_num_epochs (
unsigned long num
) const;
/*!
requires
- num > 0
ensures
- @get_num_epochs() == num
!*/
void be_verbose (
);
/*!
ensures
- This object will print status messages to standard out so that a
user can observe the progress of the algorithm.
!*/
void be_quiet (
);
/*!
ensures
- This object will not print anything to standard out
!*/
const net_type& train (
const std::vector<input_type>& data,
......@@ -86,7 +173,25 @@ namespace dlib
/*!
requires
- data.size() == labels.size()
- TODO: the net has a supervised loss layer.
- net_type uses a supervised loss.
i.e. net_type::label_type != no_label_type.
ensures
- Trains a supervised neural network based on the given training data.
The goal of training is to find the network parameters that minimize
get_net().compute_loss(data.begin(), data.end(), labels.begin()).
- The optimizer will run for get_num_epochs() epochs and each layer in the
network will be optimized by its corresponding solver in get_solvers().
- returns #get_net()
(i.e. the trained network can also be accessed by calling get_net() after
train() finishes executing)
- Each call to train DOES NOT reinitialize the state of get_net() or
get_solvers(). That is, the state of the solvers and network contained
inside this trainer is the starting point for the optimization each time
train() is called. For example, calling train() 1 time and having it
execute 100 epochs of training is equivalent to calling train() 10 times
and having it execute 10 epochs of training during each call. This also
means you can serialize a trainer to disk and then, at a later date,
deserialize it and resume training your network where you left off.
!*/
const net_type& train (
......@@ -94,9 +199,25 @@ namespace dlib
);
/*!
requires
- TODO: the net has an unsupervised loss layer.
- net_type uses an unsupervised loss.
i.e. net_type::label_type == no_label_type.
ensures
- trains an auto-encoder
- Trains an unsupervised neural network based on the given training data.
The goal of training is to find the network parameters that minimize
get_net().compute_loss(data.begin(), data.end()).
- The optimizer will run for get_num_epochs() epochs and each layer in the
network will be optimized by its corresponding solver in get_solvers().
- returns #get_net()
(i.e. the trained network can also be accessed by calling get_net() after
train() finishes executing)
- Each call to train DOES NOT reinitialize the state of get_net() or
get_solvers(). That is, the state of the solvers and network contained
inside this trainer is the starting point for the optimization each time
train() is called. For example, calling train() 1 time and having it
execute 100 epochs of training is equivalent to calling train() 10 times
and having it execute 10 epochs of training during each call. This also
means you can serialize a trainer to disk and then, at a later date,
deserialize it and resume training your network where you left off.
!*/
};
......
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