transport.samples.kernels.streaming_sqeuclid.compute_bias_g
def compute_bias_g(f: torch.Tensor, alpha: torch.Tensor, gamma: torch.Tensor, eps: float) -> torch.TensorCompute pre-scaled bias for g-update: v = (f̂ + γ)/ε.
Args:
f: Current f potential [n]
alpha: Source squared norms [n] = cost_scale * ||x||²
gamma: Scaled log source weights [n] = eps * log(a)
eps: Regularization parameter
Returns:
v: Pre-scaled bias [n] = (f - alpha + gamma) / eps
Source: torchmatch/transport/samples/kernels/streaming_sqeuclid.py:130