Learnable per-fidelity sigma parameters in Tobit loss

[smcolby]

Problem

The Tobit loss uses a single fixed sigma parameter for all records regardless of fidelity. This is a mis-specification:

• DRC records have typical pEC50 measurement noise of σ ≈ 0.15 (median std.error in well-characterized screens)
• PS records encode threshold uncertainty, where σ ≈ 0.5–1.0 is more appropriate

A single σ = 0.5 simultaneously underestimates DRC precision and overestimates PS precision.

Proposed Solution

Replace the fixed sigma scalar with learnable per-fidelity log-sigma parameters:

# moal/model.py
self.log_sigma_drc = nn.Parameter(torch.tensor(math.log(0.2)))  # init ~0.2
self.log_sigma_ps  = nn.Parameter(torch.tensor(math.log(0.5)))  # init ~0.5

@property
def sigma_drc(self): return self.log_sigma_drc.exp()

@property
def sigma_ps(self): return self.log_sigma_ps.exp()

In _tobit_loss(), use sigma_drc for CensoringType.EXACT records and sigma_ps for LEFT/INTERVAL records.

Alternative (simpler): Keep sigma fixed but expose sigma_drc and sigma_ps as separate constructor parameters (no gradient through sigma).

Backward compatibility: Keep sigma as a constructor parameter that initializes both; deprecate when per-fidelity sigma is set.

Files

• moal/model.py