Value API

Scalar values and automatic differentiation graph operations. More...

Data Structures
struct	mg_graph_checkpoint
	Snapshot of a graph's allocation state. More...

Typedefs
typedef struct mg_graph	mg_graph
	Opaque owner of a linked list of mg_value nodes.
typedef struct mg_value	mg_value
	Opaque scalar node in a computation graph.

Functions
mg_graph *	mg_graph_new (void)
	Create a new graph.
void	mg_graph_free (mg_graph *g)
	Free a graph and all values it owns.
void	mg_zero_grad (mg_graph *g)
	Set all mg_value gradients in a graph to zero.
mg_graph_checkpoint	mg_graph_save (const mg_graph *g)
	Create a snapshot of the graph's current allocation state.
void	mg_graph_restore (mg_graph *g, mg_graph_checkpoint checkpoint)
	Restore the graph to a previous checkpoint.
float	mg_data (const mg_value *v)
	Get the scalar data of a value.
float	mg_grad (const mg_value *v)
	Get the gradient of a value.
void	mg_set_data (mg_value *v, float data)
	Set the scalar data of a value.
void	mg_set_grad (mg_value *v, float grad)
	Set the gradient of a value.
mg_value *	mg_scalar (mg_graph *g, float data)
	Create a new scalar value.
mg_value *	mg_add (mg_graph *g, mg_value *a, mg_value *b)
	Add two values.
mg_value *	mg_sub (mg_graph *g, mg_value *a, mg_value *b)
	Subtract one value from another.
mg_value *	mg_mul (mg_graph *g, mg_value *a, mg_value *b)
	Multiply two values.
mg_value *	mg_div (mg_graph *g, mg_value *a, mg_value *b)
	Divide one value by another.
mg_value *	mg_pow (mg_graph *g, mg_value *a, mg_value *b)
	Raise one value to the power of another.
mg_value *	mg_neg (mg_graph *g, mg_value *a)
	Negate a value.
mg_value *	mg_square (mg_graph *g, mg_value *a)
	Square a value.
mg_value *	mg_relu (mg_graph *g, mg_value *a)
	Apply the ReLU activation function to a value.
mg_value *	mg_tanh (mg_graph *g, mg_value *a)
	Apply the tanh activation function to a value.
mg_value *	mg_exp (mg_graph *g, mg_value *a)
	Apply the exponential function to a value.
bool	mg_backward (mg_graph *g, mg_value *out)
	Compute gradients for all values that contribute to an output value.

Detailed Description

Scalar values and automatic differentiation graph operations.

Typedef Documentation

mg_graph

typedef struct mg_graph mg_graph

Opaque owner of a linked list of mg_value nodes.

A graph frees all values it owns when passed to mg_graph_free.

mg_value

typedef struct mg_value mg_value

Opaque scalar node in a computation graph.

A value stores scalar data, its gradient, and the operation and parent values that produced it.

Function Documentation

mg_add()

mg_value * mg_add	(	mg_graph *	g,
		mg_value *	a,
		mg_value *	b
	)

Add two values.

Parameters

g	Graph that owns the result.
a	Left operand.
b	Right operand.

Returns

The sum, or NULL if allocation fails.

mg_backward()

bool mg_backward	(	mg_graph *	g,
		mg_value *	out
	)

Compute gradients for all values that contribute to an output value.

On failure, the graph is left unchanged.

Parameters

g	Graph containing the values.
out	Output value to differentiate from.

Returns

true on success, false if allocation fails.

mg_data()

float mg_data

(

const mg_value *

v

)

Get the scalar data of a value.

Parameters

v	Value to inspect.

Returns

The scalar data stored in v.

mg_div()

mg_value * mg_div	(	mg_graph *	g,
		mg_value *	a,
		mg_value *	b
	)

Divide one value by another.

Parameters

g	Graph that owns the result.
a	Dividend.
b	Divisor.

Returns

The quotient, or NULL if allocation fails.

mg_exp()

mg_value * mg_exp	(	mg_graph *	g,
		mg_value *	a
	)

Apply the exponential function to a value.

Parameters

g	Graph that owns the result.
a	Input value.

Returns

The exponential value, or NULL if allocation fails.

mg_grad()

float mg_grad

(

const mg_value *

v

)

Get the gradient of a value.

Parameters

v	Value to inspect.

Returns

The gradient stored in v.

mg_graph_free()

void mg_graph_free

(

mg_graph *

g

)

Free a graph and all values it owns.

Parameters

g	Graph to free. Passing NULL is allowed.

mg_graph_new()

mg_graph * mg_graph_new

(

void

)

Create a new graph.

Returns

A new graph, or NULL if allocation fails.

mg_graph_restore()

void mg_graph_restore	(	mg_graph *	g,
		mg_graph_checkpoint	checkpoint
	)

Restore the graph to a previous checkpoint.

Parameters

g	Graph to restore.
checkpoint	Checkpoint returned by mg_graph_save.

mg_graph_save()

mg_graph_checkpoint mg_graph_save

(

const mg_graph *

g

)

Create a snapshot of the graph's current allocation state.

Parameters

g	Graph to snapshot.

Returns

A checkpoint that can be passed to mg_graph_restore.

mg_mul()

mg_value * mg_mul	(	mg_graph *	g,
		mg_value *	a,
		mg_value *	b
	)

Multiply two values.

Parameters

g	Graph that owns the result.
a	Left operand.
b	Right operand.

Returns

The product, or NULL if allocation fails.

mg_neg()

mg_value * mg_neg	(	mg_graph *	g,
		mg_value *	a
	)

Negate a value.

Parameters

g	Graph that owns the result.
a	Value to negate.

Returns

The negated value, or NULL if allocation fails.

mg_pow()

mg_value * mg_pow	(	mg_graph *	g,
		mg_value *	a,
		mg_value *	b
	)

Raise one value to the power of another.

Parameters

g	Graph that owns the result.
a	Base.
b	Exponent.

Returns

The power result, or NULL if allocation fails.

mg_relu()

mg_value * mg_relu	(	mg_graph *	g,
		mg_value *	a
	)

Apply the ReLU activation function to a value.

Parameters

g	Graph that owns the result.
a	Input value.

Returns

The activated value, or NULL if allocation fails.

mg_scalar()

mg_value * mg_scalar	(	mg_graph *	g,
		float	data
	)

Create a new scalar value.

Parameters

g	Graph that owns the new value.
data	Initial scalar data.

Returns

The new value, or NULL if allocation fails.

mg_set_data()

void mg_set_data	(	mg_value *	v,
		float	data
	)

Set the scalar data of a value.

Parameters

v	Value to update.
data	New scalar data.

mg_set_grad()

void mg_set_grad	(	mg_value *	v,
		float	grad
	)

Set the gradient of a value.

Parameters

v	Value to update.
grad	New gradient.

mg_square()

mg_value * mg_square	(	mg_graph *	g,
		mg_value *	a
	)

Square a value.

Parameters

g	Graph that owns the result.
a	Value to square.

Returns

The squared value, or NULL if allocation fails.

mg_sub()

mg_value * mg_sub	(	mg_graph *	g,
		mg_value *	a,
		mg_value *	b
	)

Subtract one value from another.

Parameters

g	Graph that owns the result.
a	Minuend.
b	Subtrahend.

Returns

The difference, or NULL if allocation fails.

mg_tanh()

mg_value * mg_tanh	(	mg_graph *	g,
		mg_value *	a
	)

Apply the tanh activation function to a value.

Parameters

g	Graph that owns the result.
a	Input value.

Returns

The activated value, or NULL if allocation fails.

mg_zero_grad()

void mg_zero_grad

(

mg_graph *

g

)

Set all mg_value gradients in a graph to zero.

Parameters

g	Graph whose values should be reset.