pulsatrix 1.1.0
Loading...
Searching...
No Matches
pulsatrix::TrajectoryHop Struct Reference

One reverse step's split, aggregated over elements. More...

#include <trajectory_lrp.hpp>

Public Member Functions

double network_leak () const
 What the network's relevance pass didn't conserve: biases, norms, its stabilizers.
 

Public Attributes

int64_t step = 0
 The index into ContinuousTrajectory::timesteps, and that timestep.
 
int64_t timestep = 0
 
double seeded = 0
 Relevance arriving on x_prev.
 
double skip = 0
 Its split: skip (to x_t), network (into eps_theta), noise (a sink), and what the split's epsilon absorbed (a sink). seeded = skip + network + noise + absorbed.
 
double network = 0
 
double noise = 0
 
double absorbed = 0
 
double signed_skip = 0
 The signed fractions, logged next to the split: sum_i R_i h_k,i / s_i for each term, where s_i = sum_k h_k,i (elements with s_i = 0 are left out). They sum to seeded over those elements and grow without bound when the terms cancel.
 
double signed_network = 0
 
double signed_noise = 0
 
double network_to_x = 0
 Where the network share landed: x_t, the conditioning, the timestep embedding.
 
double network_to_condition = 0
 
double network_to_time = 0
 

Detailed Description

One reverse step's split, aggregated over elements.

Member Function Documentation

◆ network_leak()

double pulsatrix::TrajectoryHop::network_leak ( ) const
inline

What the network's relevance pass didn't conserve: biases, norms, its stabilizers.

Member Data Documentation

◆ absorbed

double pulsatrix::TrajectoryHop::absorbed = 0

◆ network

double pulsatrix::TrajectoryHop::network = 0

◆ network_to_condition

double pulsatrix::TrajectoryHop::network_to_condition = 0

◆ network_to_time

double pulsatrix::TrajectoryHop::network_to_time = 0

◆ network_to_x

double pulsatrix::TrajectoryHop::network_to_x = 0

Where the network share landed: x_t, the conditioning, the timestep embedding.

◆ noise

double pulsatrix::TrajectoryHop::noise = 0

◆ seeded

double pulsatrix::TrajectoryHop::seeded = 0

Relevance arriving on x_prev.

◆ signed_network

double pulsatrix::TrajectoryHop::signed_network = 0

◆ signed_noise

double pulsatrix::TrajectoryHop::signed_noise = 0

◆ signed_skip

double pulsatrix::TrajectoryHop::signed_skip = 0

The signed fractions, logged next to the split: sum_i R_i h_k,i / s_i for each term, where s_i = sum_k h_k,i (elements with s_i = 0 are left out). They sum to seeded over those elements and grow without bound when the terms cancel.

◆ skip

double pulsatrix::TrajectoryHop::skip = 0

Its split: skip (to x_t), network (into eps_theta), noise (a sink), and what the split's epsilon absorbed (a sink). seeded = skip + network + noise + absorbed.

◆ step

int64_t pulsatrix::TrajectoryHop::step = 0

The index into ContinuousTrajectory::timesteps, and that timestep.

◆ timestep

int64_t pulsatrix::TrajectoryHop::timestep = 0

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