Signals and storage

Choose this based on storage budget, task complexity (e.g., number of classes, annotation density) and how often you need overlays during training.

Light mode — train keeps only per-sample loss, eval keeps full data:

# Training step: store per-sample loss only
train_loss = sig["loss"](outputs, targets, batch_ids=ids)

# Evaluation step: store predictions + targets for overlay analysis
eval_preds = decode_and_nms(outputs.detach())
eval_loss = sig["loss"](outputs, targets, batch_ids=ids, preds=eval_preds, targets=targets)

Use this when you want lighter train-time writes but still need rich eval-time inspection in Studio.

Note

We still include ‘batch_ids’ in the signal call for both train and eval, so you can still sort and filter by sample in the UI. The studio will not store the full arrays for train, but it will still let you inspect the loss per sample and history.

Standard mode — both train and eval store full data:

# Training step
train_preds = decode_and_nms(outputs.detach())
train_loss = sig["loss"](outputs, targets, batch_ids=ids, preds=train_preds, targets=targets)

# Evaluation step
eval_preds = decode_and_nms(outputs.detach())
eval_loss = sig["loss"](outputs, targets, batch_ids=ids, preds=eval_preds, targets=targets)

preds should be processed predictions (after NMS, argmax, etc.) rather than raw model outputs, because the studio renders them directly as overlays. The optional targets override is useful when the annotation fed to the loss function differs from the one the studio should display (e.g. encoded anchors vs. decoded boxes).

Summary table

Mode	Call	Stores	Use when
Light	train: sig(out, tgt, batch_ids=ids) / eval: sig(out, tgt, batch_ids=ids, preds=preds, targets=tgt)	train: loss only / eval: loss + predictions + targets	Large or medium datasets, lower write cost during training
Standard	train + eval: sig(out, tgt, batch_ids=ids, preds=preds, targets=tgt)	train + eval: loss + predictions + targets	Smaller datasets, maximum observability on both phases