Methods

shiftkit.methods provides domain adaptation training loops. All trainers record identical per-epoch history dicts so their results can be directly compared.

ShiftKit methods fall into two broad families:

---

Feature-based methods

Feature-based methods work by transforming the model's learned representations so that source and target features become indistinguishable. The encoder is trained jointly on a supervised task loss (cross-entropy on source labels) and a domain alignment loss that penalises differences between the source and target latent distributions. Because alignment happens in latent space, these methods require a model with a separate encode() step and work regardless of the input modality.

Trainer	DA mechanism	Key parameter	Page
SourceOnlyTrainer	No adaptation (baseline)	—	→
MMDTrainer	Latent distribution matching via MMD	mmd_weight λ	→
LMMDTrainer	Per-class subdomain alignment via local MMD	lmmd_weight λ	→
CORALTrainer	Covariance alignment (second-order statistics)	coral_weight λ	→
DANNTrainer	Adversarial discriminator + GRL	domain_weight λ	→
SIDDATrainer	Sinkhorn optimal transport + learnable η weights	warmup_epochs	→

---

Instance-based methods

Instance-based methods do not modify the feature space. Instead, they estimate how much more or less likely each source sample is under the target distribution and reweight the training loss accordingly. This approach is theoretically grounded under the covariate shift assumption — that the label conditionals are the same across domains (p_src(y|x) = p_tgt(y|x)) while only the input marginals differ. These methods require only a standard forward() interface and compute importance weights once before training begins.

Trainer	DA mechanism	Key parameter	Page
KLIEPTrainer	Importance weighting via density ratio estimation (gradient ascent)	n_centers, weight_clip	→
KMMTrainer	Importance weighting via kernel mean matching (QP)	B, weight_clip	→

All trainers share the same interface:

trainer = AnyTrainer(model, source_loader, target_loader, ...)
history = trainer.fit(epochs=10)
result  = trainer.evaluate(test_loader, domain="target-test")

---

Method comparison

	Source Only	MMD	LMMD	CORAL	DANN	SIDDA	KLIEP
DA family	—	Feature-based	Feature-based	Feature-based	Feature-based	Feature-based	Instance-based
Alignment target	None	Full distribution	Per-class subdomains	Covariance matrix	Domain labels	Optimal transport plan	Sample weights (density ratio)
What is matched	—	All moments (via kernel)	Class-conditional moments	2nd-order statistics	Domain membership	Entire marginal distribution	p_tgt(x) / p_src(x)
Kernel required	No	Yes — RBF, bandwidth σ	Yes — RBF, bandwidth σ	No	No	No (Sinkhorn entropic OT)	Yes — RBF in input space
Needs source labels	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Needs target labels	No	No	Pseudo-labels (soft)	No	No	No	No
Adversarial training	No	No	No	No	Yes (GRL)	No	No
Learnable loss weights	No	No	No	No	No	Yes (η₁, η₂)	No
Alignment cost	—	Every batch	Every batch	Every batch	Every batch	Every batch	Once at init
Model interface	forward()	encode() + classify()	encode() + classify()	encode() + classify()	encode() + classify()	full SIDDA interface	forward() only
Computation per batch	O(n·d)	O(n²) kernel matrices	O(n²) per class	O(n·d²) covariance	O(n·d) + discriminator	O(n²) Sinkhorn iterations	O(n·m) weight lookup
Key hyperparameter	—	mmd_weight λ	lmmd_weight λ	coral_weight λ	domain_weight λ	warmup_epochs, blur schedule	n_centers, weight_clip
Warmup supported	—	Yes	Yes	Yes	Yes	Mandatory	No
Covariate shift assumption	No	No	No	No	No	No	Yes
Best suited for	Reference baseline	General distribution shift	Class-level shift with label imbalance	Shift in feature scale / correlation	Strong covariate shift with large batches	Unknown shift type; automatically reweights objectives	Covariate shift on tabular / low-dim data

Choosing a method: Start with the Source-Only baseline to measure the domain gap. For most tasks, MMD or CORAL is a fast, strong first attempt. Use LMMD when class distributions differ across domains. Use DANN when the shift is severe and batch sizes are large enough to train the discriminator. Use SIDDA when you want automatic loss balancing without manual λ tuning. Use KLIEP when the covariate shift assumption holds and you prefer instance reweighting over feature alignment — especially effective on tabular data.

---

Shared history format

Every fit() call returns a list[dict] with one entry per epoch:

Key	Type	Description
epoch	int	Epoch index (1-based)
ce_loss	float	Cross-entropy loss
mmd_loss	float	MMD² loss (0.0 if not applicable)
domain_loss	float	Adversarial domain loss (0.0 if not applicable)
da_loss	float	Sinkhorn DA loss (0.0 if not applicable)
eta1	float	Learned CE weight η₁ (SIDDA only)
eta2	float	Learned DA weight η₂ (SIDDA only)
sigma	float	Sinkhorn blur used (SIDDA only)
mean_potential	float	Mean source Kantorovich potential (SIDDA with use_potentials=True only)
total_loss	float	Total combined loss
src_acc	float	Source domain accuracy
tgt_acc	float	Target domain accuracy (tracked, not directly optimised)

---