Stop rewriting PyTorch boilerplate.¶

torchloop handles the training loop, evaluation, logging, and export — so you can focus on the model.

The Problem¶

You write the same PyTorch training loop in every project. Same checkpoint logic. Same metric assembly. Same export steps. It's tedious and inconsistent.

torchloop abstracts exactly that — nothing more.

What's Inside¶

Module	What it does
Trainer	Training loop, early stopping, AMP, LR scheduler
Evaluator	Classification report, confusion matrix, per-class F1
Callbacks	W&B and MLflow logging hooks
Exporter	PyTorch → ONNX → TFLite
Edge	Deployment utilities and FLOPs estimation

Learn more about Export & Edge Deployment →

Quick Start¶

from torchloop import Trainer, Evaluator
import torch.optim.lr_scheduler as sched

scheduler = sched.ReduceLROnPlateau(optimizer, patience=5)

trainer = Trainer(
    model=model,
    optimizer=optimizer,
    criterion=criterion,
    device="cuda",
    scheduler=scheduler,
    amp=True,
    patience=10,
)

history = trainer.fit(train_loader, val_loader, epochs=50)
trainer.save("best.pt")

ev = Evaluator(model, device="cuda")
ev.report(val_loader, class_names=["No Damage", "Minor", "Major", "Destroyed"])
fig = ev.confusion_matrix(val_loader)
fig.savefig("confusion_matrix.png")

Install¶

BasicWith LoggingWith ExportWith EdgeEverything

pip install torchloop

pip install torchloop[logging]

pip install torchloop[export]

pip install torchloop[edge]

pip install torchloop[all]

Design Principles¶

Philosophy

No lock-in: Works with any nn.Module. No subclassing required.

Minimal surface area: Three core modules. That's it.

You own the model: torchloop wraps your loop, doesn't replace your architecture.

Roadmap¶

[x] v0.1.0 — Trainer, Evaluator, Exporter
[x] v0.2.0 — LR scheduler support, mixed precision (AMP)
[x] v0.2.1 — Gradient accumulation + callbacks
[x] v0.2.2 — Edge submodule
[ ] v0.3.0 — W&B / MLflow hooks + CoreML export
[ ] v0.3.1 — Model pruning utilities