4.1.8. Parameter

class primitiv::Parameter

Class to manage a trainable tensor parameter.

Inherits from primitiv::mixins::Nonmovable< Parameter >

Public Functions

Parameter()

Creates an invalid parameter object.

Parameter(const Shape &shape, const std::vector<float> &value, Device *device)

Creates a new Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• value: List of initial values. Order of elements should be the column-major (Fortran) order.
• device: The device object to manage internal memory.

Parameter(const Shape &shape, const std::vector<float> &value, Device &device)

Creates a new Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• value: List of initial values. Order of elements should be the column-major (Fortran) order.
• device: The device object to manage internal memory.

Parameter(const Shape &shape, const std::vector<float> &value)

Creates a new Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• value: List of initial values. Order of elements should be the column-major (Fortran) order.

Parameter(const Shape &shape, const Initializer &initializer, Device *device)

Creates a new Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• initializer: An Initializer object.
• device: The device object to manage internal memory.

Parameter(const Shape &shape, const Initializer &initializer, Device &device)

Creates a new Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• initializer: An Initializer object.
• device: The device object to manage internal memory.

Parameter(const Shape &shape, const Initializer &initializer)

Creates a new Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• initializer: An Initializer object.

void init(const Shape &shape, const std::vector<float> &value, Device *device)

Initializes the Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• value: List of initial values. Order of elements should be the column-major (Fortran) order.
• device: The device object to manage internal memory.

void init(const Shape &shape, const std::vector<float> &value, Device &device)

Initializes the Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• value: List of initial values. Order of elements should be the column-major (Fortran) order.
• device: The device object to manage internal memory.

void init(const Shape &shape, const std::vector<float> &value)

Initializes the Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• value: List of initial values. Order of elements should be the column-major (Fortran) order.

void init(const Shape &shape, const Initializer &initializer, Device *device)

Initializes the Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• initializer: An Initializer object.
• device: The device object to manage internal memory.

void init(const Shape &shape, const Initializer &initializer, Device &device)

Initializes the Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• initializer: An Initializer object.
• device: The device object to manage internal memory.

void init(const Shape &shape, const Initializer &initializer)

Initializes the Parameter object.

Parameters

• shape: The shape of the parameter. The batch size should be 1.
• initializer: An Initializer object.

void load(const std::string &path, bool with_stats, Device *device)

Loads parameters from specified file.

Parameters

• path: File path to load parameters.
• with_stats: Whether or not to load all additional statistics as well as parameter values if the file has them.
• device: The device object to manage internal memory.

void load(const std::string &path, bool with_stats, Device &device)

Loads parameters from specified file.

Parameters

• path: File path to load parameters.
• with_stats: Whether or not to load all additional statistics as well as parameter values if the file has them.
• device: The device object to manage internal memory.

void load(const std::string &path, bool with_stats)

Loads parameters from specified file.

Parameters

• path: File path to load parameters.
• with_stats: Whether or not to load all additional statistics as well as parameter values if the file has them.

void load(const std::string &path)

Loads parameters from specified file.

Parameters

• path: File path to load parameters.

void save(const std::string &path, bool with_stats) const

Saves current parameters into specified file.

Parameters

• path: File path to save parameters.
• with_stats: Whether or not to save all additional statistics as well as parameter values if the parameter object has them.

void save(const std::string &path) const

Saves current parameters into specified file.

Parameters

• path: File path to save parameters.

bool valid() const

Returns whether the parameter is valid or not.

Return

true or false w.r.t. the parameter is valid or not.

void reset_gradient()

Set all gradients to 0.

void add_stats(const std::string &name, const Shape &shape)

Adds a new optional statistics tensor.

Remark

All elements in the new statistics tensor is initialized by 0.

Parameters

• name: Name of the statistics.
• shape: Shape of the tensor.

bool has_stats(const std::string &name) const

Checks whether the statistics with name name exists or not.

Return

true if the entry exists, false otherwise.

Parameters

• name: Name of the statistics.

Shape shape() const

Returns the shape of the parameter.

Return

Shape object.

Device &device() const

Returns the Device object to manage the internal memory.

Return

Pointer of the Device object.

const Tensor &value() const

Returns the values of the parameter.

Return

A tensor representing the parameter tensor.

Tensor &value()

Returns the values of the parameter.

Return

A tensor representing the parameter tensor.

const Tensor &gradient() const

Returns the current gradient of the parameter.

Return

A tensor representing the gradient of the value.

Tensor &gradient()

Returns the current gradient of the parameter.

Return

A tensor representing the gradient of the value.

const Tensor &stats(const std::string &name) const

Returns the current opotional statistics tensor specified by given name.

Return

A tensor.

Parameters

• name: Name of the statistics.

Tensor &stats(const std::string &name)

Returns the current opotional statistics tensor specified by given name.

Return

A tensor.

Parameters

• name: Name of the statistics.