pulsatrix::LearnableScalar Class Reference

A bare learnable scalar (e.g. a GFlowNet loss's log Z), outside the Module/LRP hierarchy entirely. More...

#include <learnable_scalar.hpp>

Public Member Functions
	LearnableScalar (float initial_value=0.0f)
	Constructs a scalar with the given initial value and zero gradient.
float	value () const
	Current value.
float	grad () const
	Accumulated gradient since the last zero_grad().
void	accumulate_grad (float grad)
	Adds to the accumulated gradient – mirrors Tensor::accumulate()'s across-multiple-contributions convention.
void	zero_grad ()
	Resets the accumulated gradient to zero. Does not change value().
void	step (float learning_rate)
	Plain SGD update: value_ -= learning_rate * grad_.

Detailed Description

A bare learnable scalar (e.g. a GFlowNet loss's log Z), outside the Module/LRP hierarchy entirely.

Note

Deliberately not a Module: AdamOptimizer/SGDOptimizer only operate over Module::parameters(), and forcing a training-only scalar into the Module contract would also force a propagate_relevance definition for something with no forward pass to explain – the same reasoning that keeps MSELoss outside the Module hierarchy (see mission_shared_gflownet_machinery.md's Recon). The GFlowNet losses this class supports compute d(loss)/d(log Z) analytically themselves; no autograd machinery is needed here at all.

Not a Tensor either – a Tensor's whole reason to exist (device placement, batched storage, DeviceBackend dispatch) is irrelevant to one scalar that never participates in a tensor op.

Constructor & Destructor Documentation

LearnableScalar()

pulsatrix::LearnableScalar::LearnableScalar ( float  initial_value = 0.0f )

explicit

Constructs a scalar with the given initial value and zero gradient.

Member Function Documentation

accumulate_grad()

void pulsatrix::LearnableScalar::accumulate_grad

(

float

grad

)

Adds to the accumulated gradient – mirrors Tensor::accumulate()'s across-multiple-contributions convention.

grad()

float pulsatrix::LearnableScalar::grad ( ) const

inline

Accumulated gradient since the last zero_grad().

step()

void pulsatrix::LearnableScalar::step

(

float

learning_rate

)

Plain SGD update: value_ -= learning_rate * grad_.

Parameters

learning_rate

Step size. Unvalidated – same convention as SGDOptimizer's own constructor, which does not validate its learning rate either.

Note

Does not reset grad() – mirrors AdamOptimizer::step()'s own convention (zero_grad() is a separate, explicit call).

value()

float pulsatrix::LearnableScalar::value ( ) const

inline

Current value.

zero_grad()

void pulsatrix::LearnableScalar::zero_grad

(

)

Resets the accumulated gradient to zero. Does not change value().

---

The documentation for this class was generated from the following file:

• include/pulsatrix/learnable_scalar.hpp