Optimal Network SNR Estimation

The optimal matched-filter SNR is the theoretical maximum SNR achievable for a known signal in stationary Gaussian noise. Sage uses it to rescale injected waveforms so that the training set covers a desired SNR distribution.

OptimalSNREstimator computes both the per-detector optimal SNR and the combined network SNR from a frequency-domain signal batch.

from sage.data.waveform.snr import OptimalSNREstimator

snr = OptimalSNREstimator()

# signal_batch: (B, n_detectors, n_freq) — complex FD output from IMRPhenomPv2
rho_net, rho_det = snr(signal_batch)

Output shapes

Variable	Shape	Description
`rho_net`	`(B,)`	Network (combined) optimal SNR: \(\rho_\text{net} = \sqrt{\sum_d \rho_d^2}\)
`rho_det`	`(B, n_detectors)`	Per-detector optimal SNR

print(rho_net.shape)    # torch.Size([512])
print(rho_det.shape)    # torch.Size([512, 2])

# Per-detector SNR for the first detector across the batch
print(rho_det[:, 0])

For a batch of 512 injections at 440 Mpc with \(m_1 = m_2 = 30\,M_\odot\), the individual detector SNRs typically range from ~6 (near-unfavorable sky location) to ~42 (optimal orientation).

SNR-based injection rescaling

In practice you do not call OptimalSNREstimator directly during training. Instead, wrap it in OptimalSNRRescaler, which draws a target SNR from a prior distribution and rescales each waveform amplitude accordingly before injection:

from sage.data.waveform.distributions.snr_rescaling import OptimalSNRRescaler
from sage.data.waveform import HalfNorm

# Half-normal target SNR prior: mean ~5, most mass between 5 and 15
target_snr_sampler = HalfNorm(scale=4.0, loc=5.0, seed=150914)
snrscaler = OptimalSNRRescaler(target_snr_sampler)

Pass snrscaler as the augment argument when constructing the waveform generator (see IMRPhenomPv2 on GPU). The rescaler computes the current optimal SNR of each waveform in the batch, then multiplies the amplitude by target_snr / current_snr.