Interpolants and Couplings¶
Transport-based training regresses a model onto a path connecting noise to data. Two independent choices define that path: the interpolant (the shape of the path in time) and the coupling (which noise point is paired with which data point). TorchEBM keeps them orthogonal, so any loss that consumes pairs composes with any of either.
Interpolants: the path¶
An interpolant defines
with \(x_0 \sim \mathcal{N}(0, I)\) and \(x_1 \sim p_{\text{data}}\); the conditional velocity \(u_t\) is the regression target of flow matching, and the same coefficients convert between velocity, score, and noise predictions.
Common schedules illustrate the design space: a linear path (\(\alpha = t\), \(\sigma = 1 - t\)) gives the straight lines of flow matching; a cosine path smooths the endpoints; a variance-preserving schedule reproduces the diffusion setting where score and noise prediction are natural. The shipped schedules, generated from the installed package:
graph TD
BaseInterpolant(["BaseInterpolant"])
BaseInterpolant --> CosineInterpolant
BaseInterpolant --> LinearInterpolant
BaseInterpolant --> VarianceExplodingInterpolant
BaseInterpolant --> VariancePreservingInterpolant The call returns the path point and the target velocity. Beyond it, the base class carries the conversion algebra (velocity_to_score, score_to_velocity, velocity_to_noise, compute_drift, compute_diffusion), which is what lets FlowSampler accept any prediction type over any path.
Couplings: the pairing¶
Under an independent pairing, conditional targets are noisy: many \((x_0, x_1)\) pairs cross, and the marginal velocity field the model must average over is curved. A coupling re-pairs each batch before interpolation. The families range from keeping the incoming pairing (independent), through cost-based pairings of increasing fidelity (greedy nearest-neighbor, entropic Sinkhorn, exact OT, and unbalanced variants that emit per-pair weights), to model-induced couplings that pair \(x_0\) with the model's own output (e.g. reflow). The shipped set, generated from the installed package:
graph TD
BaseCostCoupling(["BaseCostCoupling"])
BaseCoupling(["BaseCoupling"])
BaseModelCoupling(["BaseModelCoupling"])
BaseCostCoupling --> ExactOTCoupling
BaseCoupling --> BaseCostCoupling
BaseCostCoupling --> GreedyCoupling
BaseCoupling --> IndependentCoupling
BaseModelCoupling --> ReflowCoupling
BaseCoupling --> BaseModelCoupling
BaseCostCoupling --> SinkhornCoupling
BaseCostCoupling --> UnbalancedSinkhornCoupling Results are CouplingResult objects: they unpack as (x0, x1) and carry extras as attributes, so unpacking code keeps working when a coupling adds one. The current extra is weights, per-pair importance weights produced by unbalanced OT, which losses take through weight=. Couplings are also addressable by registry string via get_coupling.
Minibatch OT couplings reduce crossings between the paired paths, so the marginal flow is straighter1; this lowers the variance of the regression target and the number of integration steps generation needs2. Reflow is the model-induced version of the same idea: the training loop pairs noise with the model's own outputs, x1 = sampler(model, x0, n_steps=50), to iteratively rectify the flow.
Where they sit in the loop¶
Both objects are called by the training loop, right before the model:
Each training step is: couple the batch, draw \(t\), interpolate, call the model, regress; a weighted coupling's weights go to the loss's weight= argument. Everything upstream (which energy, which network) and downstream (which sampler, which integrator) is untouched by the choice.