pulsatrix::DetailedBalanceLoss Class Reference

‘Δ(s,s’) = log F(s) + log P_F(s'|s) − log F(s') − log P_B(s|s'),loss = Δ(s,s')²-- the per-*transition* credit-assignment alternative toTrajectoryBalanceLoss`'s per-trajectory constraint. More...

#include <detailed_balance_loss.hpp>

Public Member Functions
	DetailedBalanceLoss ()=default
float	forward (float log_flow_s, float log_pf, float log_flow_s_next, float log_pb)
	Computes ‘Δ(s,s’)and the lossΔ(s,s')², cachingΔfor the accessors below. @param log_flow_slog F_θ(s). @param log_pflog P_F(s'|s)(orlog P_F(stop|s)for the exit transition). @param log_flow_s_nextlog F_θ(s')for a non-terminals', orlog R(x)for the exit transition (s' = x, the trajectory's terminal state). @param log_pblog P_B(s|s')for a real transition, or0.0for the exit transition. @return The scalar loss,Δ(s,s')²`.
float	delta () const
	‘Δ(s,s’)` from the most recent forward() call.
float	grad_log_flow_s () const
	d(loss)/d(log F(s)) = 2Δ.
float	grad_log_flow_s_next () const
	‘d(loss)/d(log F(s’)) = -2Δ‘. Only meaningful when s’ is non-terminal (a real F_θ network call) – callers must not apply this to the exit transition's fixed log R(x) endpoint, which has no network to backprop into.
float	grad_weight_for_log_pf () const
	The weight w = -2Δ such that this transition's policy-network gradient is w · (p[k] − 1{k==a}) – TrajectoryBalanceLoss/PolicyGradientLoss's exact softmax/log gradient identity shape.

Detailed Description

‘Δ(s,s’) = log F(s) + log P_F(s'|s) − log F(s') − log P_B(s|s'),loss = Δ(s,s')²-- the per-*transition* credit-assignment alternative toTrajectoryBalanceLoss`'s per-trajectory constraint.

Note

The terminal/exit transition is not a special case of this class: passing log_f_s_next = log R(x) (the reward, not a learned flow) and log_pb = 0.0 (log(1), the trivial backward policy from a unique virtual sink state) reduces the general formula to Δ_exit = log F(s_n) + log P_F(stop|s_n) − log R(x) exactly. See mission_detailed_balance_loss.md's Design section.

Structurally identical math to TrajectoryBalanceLoss (a + b − c − d, squared) at a different granularity – implemented as its own class rather than sharing a base, matching this codebase's own precedent (MSELoss/CrossEntropyLoss/ BCEWithLogitsLoss/KLDivergenceLoss are all independent despite similar shapes).

Not a Module subclass, for MSELoss's own reason.

Constructor & Destructor Documentation

DetailedBalanceLoss()

pulsatrix::DetailedBalanceLoss::DetailedBalanceLoss ( )

default

Member Function Documentation

delta()

float pulsatrix::DetailedBalanceLoss::delta

(

)

const

‘Δ(s,s’)` from the most recent forward() call.

forward()

float pulsatrix::DetailedBalanceLoss::forward	(	float	log_flow_s,
		float	log_pf,
		float	log_flow_s_next,
		float	log_pb
	)

Computes ‘Δ(s,s’)and the lossΔ(s,s')², cachingΔfor the accessors below. @param log_flow_slog F_θ(s). @param log_pflog P_F(s'|s)(orlog P_F(stop|s)for the exit transition). @param log_flow_s_nextlog F_θ(s')for a non-terminals', orlog R(x)for the exit transition (s' = x, the trajectory's terminal state). @param log_pblog P_B(s|s')for a real transition, or0.0for the exit transition. @return The scalar loss,Δ(s,s')²`.

grad_log_flow_s()

float pulsatrix::DetailedBalanceLoss::grad_log_flow_s

(

)

const

d(loss)/d(log F(s)) = 2Δ.

grad_log_flow_s_next()

float pulsatrix::DetailedBalanceLoss::grad_log_flow_s_next

(

)

const

‘d(loss)/d(log F(s’)) = -2Δ‘. Only meaningful when s’ is non-terminal (a real F_θ network call) – callers must not apply this to the exit transition's fixed log R(x) endpoint, which has no network to backprop into.

grad_weight_for_log_pf()

float pulsatrix::DetailedBalanceLoss::grad_weight_for_log_pf

(

)

const

The weight w = -2Δ such that this transition's policy-network gradient is w · (p[k] − 1{k==a}) – TrajectoryBalanceLoss/PolicyGradientLoss's exact softmax/log gradient identity shape.

---

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

• include/pulsatrix/detailed_balance_loss.hpp