One sub-trajectory pair's contribution to the SubTB(λ) loss: Δ(i,j) = log F(s_i) + Σ log P_F − log F(s_j) − Σ log P_B (summed over the edges spanned by [i,j)), weighted by ‘pair_weight_ratio = λ^{j-i} / Σ_{i’<j'} λ^{j'-i'}` (the pre-normalized share of the total weighted-average loss this specific pair contributes).
More...
#include <subtb_loss.hpp>
|
| | SubTBLoss ()=default |
| |
| float | forward (float log_flow_i, float sum_log_pf, float log_flow_j, float sum_log_pb, float pair_weight_ratio) |
| | Computes this pair's Δ(i,j) and its weighted contribution to the total SubTB loss, pair_weight_ratio · Δ(i,j)².
|
| |
| float | delta () const |
| | Δ(i,j) (unweighted) from the most recent forward() call.
|
| |
| float | grad_log_flow_i () const |
| | pair_weight_ratio · 2Δ(i,j).
|
| |
| float | grad_log_flow_j () const |
| | pair_weight_ratio · (-2Δ(i,j)). Only meaningful when s_j is non-terminal (a real F_θ network call) – callers must not apply this when j is the trajectory's terminal sink position (fixed log R(x), no network to backprop into).
|
| |
| float | grad_weight_for_log_pf_range () const |
| | The weight w = pair_weight_ratio · (-2Δ(i,j)), applied uniformly to every edge t in [i,j)'s own P_F softmax/log gradient identity (w · (p[k] − 1{k==a_t})) – one Δ(i,j) affects every decision point the sub-trajectory spans equally, via the chain rule through sum_log_pf.
|
| |
One sub-trajectory pair's contribution to the SubTB(λ) loss: Δ(i,j) = log F(s_i) + Σ log P_F − log F(s_j) − Σ log P_B (summed over the edges spanned by [i,j)), weighted by ‘pair_weight_ratio = λ^{j-i} / Σ_{i’<j'} λ^{j'-i'}` (the pre-normalized share of the total weighted-average loss this specific pair contributes).
- Note
- Structurally identical math to
TrajectoryBalanceLoss/DetailedBalanceLoss (a+b-c-d, squared), generalized to an arbitrary sub-trajectory span with an explicit weighting factor – TrajectoryBalanceLoss is the (i=0, j=n+1, ratio=1) special case; DetailedBalanceLoss is any adjacent-pair (i, i+1, ratio=1) special case. See mission_subtb_loss.md's Design section.
-
pair_weight_ratio is computed by the caller (from λ, i, j, and the total pair-weight sum W for the trajectory), not by this class – W depends on every pair in the trajectory, information this single-pair class has no reason to own.
-
Not a Module subclass, for
MSELoss's own reason.
◆ SubTBLoss()
| pulsatrix::SubTBLoss::SubTBLoss |
( |
| ) |
|
|
default |
◆ delta()
| float pulsatrix::SubTBLoss::delta |
( |
| ) |
const |
Δ(i,j) (unweighted) from the most recent forward() call.
◆ forward()
| float pulsatrix::SubTBLoss::forward |
( |
float |
log_flow_i, |
|
|
float |
sum_log_pf, |
|
|
float |
log_flow_j, |
|
|
float |
sum_log_pb, |
|
|
float |
pair_weight_ratio |
|
) |
| |
Computes this pair's Δ(i,j) and its weighted contribution to the total SubTB loss, pair_weight_ratio · Δ(i,j)².
- Parameters
-
| log_flow_i | log F_θ(s_i). |
| sum_log_pf | Σ_{t=i}^{j-1} log P_F(s_{t+1}|s_t) over the spanned edges. |
| log_flow_j | log F_θ(s_j) for a non-terminal s_j, or log R(x) if j is the trajectory's terminal sink position. |
| sum_log_pb | Σ_{t=i}^{j-1} log P_B(s_t|s_{t+1}) over the spanned edges (the exit edge contributes 0.0 to this sum, per DetailedBalanceLoss's own convention). |
| pair_weight_ratio | λ^{j-i} / W, this pair's pre-normalized share of the total weighted-average loss. Must be in [0, 1] for the total loss across all pairs of a trajectory to sum to a true weighted average – not validated here (a caller-internal invariant, not an external boundary). |
- Returns
- This pair's weighted contribution to the total loss.
◆ grad_log_flow_i()
| float pulsatrix::SubTBLoss::grad_log_flow_i |
( |
| ) |
const |
pair_weight_ratio · 2Δ(i,j).
◆ grad_log_flow_j()
| float pulsatrix::SubTBLoss::grad_log_flow_j |
( |
| ) |
const |
pair_weight_ratio · (-2Δ(i,j)). Only meaningful when s_j is non-terminal (a real F_θ network call) – callers must not apply this when j is the trajectory's terminal sink position (fixed log R(x), no network to backprop into).
◆ grad_weight_for_log_pf_range()
| float pulsatrix::SubTBLoss::grad_weight_for_log_pf_range |
( |
| ) |
const |
The weight w = pair_weight_ratio · (-2Δ(i,j)), applied uniformly to every edge t in [i,j)'s own P_F softmax/log gradient identity (w · (p[k] − 1{k==a_t})) – one Δ(i,j) affects every decision point the sub-trajectory spans equally, via the chain rule through sum_log_pf.
The documentation for this class was generated from the following file:
1.9.8