SourceOnlyTrainer

A no-adaptation baseline that trains exclusively on labelled source data using cross-entropy loss. No domain alignment is applied.

Use this alongside an adaptation method (MMD, DANN, …) to quantify how much benefit domain adaptation provides on your task.

---

Usage

from shiftkit.methods import SourceOnlyTrainer

trainer = SourceOnlyTrainer(
    model=model,
    source_loader=train_src,
    target_loader=train_tgt,   # used for tgt_acc tracking only
    lr=1e-3,
)
history = trainer.fit(epochs=10)

result = trainer.evaluate(test_tgt, domain="target-test")
print(f"Target accuracy: {result['accuracy']*100:.1f}%")

---

Constructor

Parameter	Type	Default	Description
model	nn.Module	—	Any model with a standard forward() method
source_loader	DataLoader	—	Labelled source DataLoader
target_loader	DataLoader	—	Target DataLoader (labels used for tgt_acc tracking only)
lr	float	1e-3	Adam learning rate
device	str \| None	None	'cuda', 'mps', or 'cpu'; auto-detected if None

---

fit(epochs=10)

Train for epochs epochs using cross-entropy on source data only.

Returns: list[dict] — one dict per epoch:

Key	Description
epoch	Epoch number (1-indexed)
ce_loss	Mean cross-entropy loss
mmd_loss	Always 0.0 (for history-format compatibility)
total_loss	Same as ce_loss
src_acc	Source domain training accuracy
tgt_acc	Target domain accuracy (no gradient signal — purely diagnostic)

---

evaluate(loader, domain="source")

Compute classification accuracy on any labelled DataLoader.

Returns: dict with keys domain (str), accuracy (float), n_samples (int).