loss Namespace Reference

Functions
	base_loss_components (pred_wdl, target_wdl, wdl_weight=1.0)
	base_loss (pred_wdl, target_wdl, wdl_weight=1.0)
	killer_loss (pred_wdl, target_wdl, wdl_weight=1.0)
	survivor_loss (pred_wdl, target_wdl, wdl_weight=1.0)
	gater_loss (pred_gates, target_gates)
	moe_loss (pred_wdl, target_wdl, expert_type='BASE')

Variables
int	B = 8
	pred_wdl = F.softmax(torch.randn(B, 3), dim=1)
	target_wdl = F.softmax(torch.randn(B, 3), dim=1)
	loss = moe_loss(pred_wdl, target_wdl, expert_type)
	pred_gates = torch.sigmoid(torch.randn(B, 2))
	target_gates = torch.rand(B, 2)
	g_loss = gater_loss(pred_gates, target_gates)

Detailed Description

@file loss.py
@brief Loss functions for training the chess neural network.

Provides functions to compute custom loss metrics combining Win-Draw-Loss (WDL)
probabilities and regularization penalties.

Function Documentation

base_loss()

base_loss	(		pred_wdl,
			target_wdl,
			wdl_weight = 1.0 )

Here is the call graph for this function:

Here is the caller graph for this function:

base_loss_components()

base_loss_components	(		pred_wdl,
			target_wdl,
			wdl_weight = 1.0 )

Standard loss for base experts (Tactical, Strategic, Major End, Minor End).

WDL is the primary objective using Brier-style MSE.

Args:
    pred_wdl: [B, 3] predicted WDL probabilities (from softmax)
    target_wdl: [B, 3] target WDL probabilities
    
Returns:
    (total_loss, weighted_wdl_loss)

Here is the caller graph for this function:

gater_loss()

gater_loss	(		pred_gates,
			target_gates )

Distillation loss for SmartGater.

Args:
    pred_gates: [B, 2] predicted [survivor, killer] gates
    target_gates: [B, 2] teacher [survivor, killer] gates from C++
    
Returns:
    Scalar MSE loss

killer_loss()

killer_loss	(		pred_wdl,
			target_wdl,
			wdl_weight = 1.0 )

Zoomed loss for Killer expert (winning positions).

KILLER sees positions where STM is winning (eval > 150cp).
Since they will be combined with base experts via residual addition,
the killer expert learns to predict a CORRECTION that emphasizes win certainty.

Zoomed WDL: Maps WinProb [0.60, 1.0] -> [0.0, 1.0]
This creates stronger gradients in winning positions where base experts plateau.

The output should be interpreted as:
- High zoomed_win = "base expert underestimates win probability, add more"
- Low zoomed_win = "base expert is already confident enough"

Here is the caller graph for this function:

moe_loss()

moe_loss	(		pred_wdl,
			target_wdl,
			expert_type = 'BASE' )

Unified loss function with expert type selection.

Args:
    pred_wdl: [B, 3]
    target_wdl: [B, 3]
    expert_type: 'BASE', 'KILLER', or 'SURVIVOR'

Here is the call graph for this function:

survivor_loss()

survivor_loss	(		pred_wdl,
			target_wdl,
			wdl_weight = 1.0 )

Zoomed loss for Survivor expert (losing positions).

SURVIVOR sees positions where STM is losing (eval < -150cp).
The goal is to find defensive resources - draws or prolonging the game.

Zoomed WDL: Maps LossProb [0.60, 1.0] -> [0.0, 1.0]
Focus: Maximize draw probability, not just predict loss correctly.

The output emphasizes:
- Draw probability as the primary "hope" signal
- Lower zoomed_loss = more swindle potential

Here is the caller graph for this function:

Variable Documentation

B

int loss.B = 8

g_loss

loss.g_loss = gater_loss(pred_gates, target_gates)

loss

loss.loss = moe_loss(pred_wdl, target_wdl, expert_type)

pred_gates

loss.pred_gates = torch.sigmoid(torch.randn(B, 2))

pred_wdl

loss.pred_wdl = F.softmax(torch.randn(B, 3), dim=1)

target_gates

loss.target_gates = torch.rand(B, 2)

target_wdl

loss.target_wdl = F.softmax(torch.randn(B, 3), dim=1)