"""Couplings: how noise is paired with data before interpolation.
The coupling decides which x0 trains against which x1; straighter pairings mean
straighter probability paths and fewer sampling steps downstream.
cost = E ||x1 - x0||^2 under the pairing
A CouplingResult unpacks as the pair (x0, x1); extras such as the per-pair
weights of unbalanced OT, which relaxes the marginal constraints and reweights
the pairs, ride on it as attributes and go to the loss's weight= argument. In
a training loop the coupling runs right before the interpolant:
x0, x1 = coupling(x0, x1)
x_t, u_t = path(x0, x1, t)
"""
import torch
from torchebm.couplings import (
ExactOTCoupling,
GreedyCoupling,
IndependentCoupling,
SinkhornCoupling,
UnbalancedSinkhornCoupling,
)
from torchebm.datasets import TwoMoonsDataset
torch.manual_seed(0)
x0 = torch.randn(512, 2)
x1 = TwoMoonsDataset(n_samples=512, noise=0.05, seed=0).get_data()
def cost(a, b):
return ((b - a) ** 2).sum(-1).mean()
couplings = {
"independent": IndependentCoupling(),
"greedy": GreedyCoupling(),
"sinkhorn (reg=0.05)": SinkhornCoupling(reg=0.05),
"exact OT": ExactOTCoupling(),
}
print(f"{'coupling':22s} E||x1 - x0||^2")
for name, coupling in couplings.items():
x0_paired, x1_paired = coupling(x0, x1)
print(f"{name:22s} {cost(x0_paired, x1_paired):.3f}")
result = UnbalancedSinkhornCoupling(reg=0.05, reg_marginal=1.0)(x0, x1)
weights = result.weights
print(f"unbalanced sinkhorn {cost(result.x0, result.x1):.3f}"
f" weights: min {weights.min():.2f} max {weights.max():.2f} (uniform would be 1.0)")