torchloop

Lightweight PyTorch utility library for training, evaluation, and TFLite export — without the framework lock-in.

---

The Problem

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

torchloop abstracts exactly that — nothing more.

---

Install

# Base installation
pip install torchloop

# With TFLite export support
pip install torchloop[export]

# With edge deployment support
pip install torchloop[edge]

# Development setup
pip install torchloop[dev]

---

Usage

Training

from torchloop import EarlyStopping, ModelCheckpoint, Trainer

trainer = Trainer(
    model,
    optimizer=torch.optim.Adam(model.parameters()),
    criterion=torch.nn.CrossEntropyLoss(),
    device="cuda",
    use_amp=True,
    accumulate_steps=4,
    patience=5,
)

trainer.add_callback(EarlyStopping(patience=5))
trainer.add_callback(ModelCheckpoint(filepath="best.pt"))

history = trainer.fit(train_loader, val_loader, epochs=30)
trainer.save("best.pt")

Evaluation

from torchloop import Evaluator

ev = Evaluator(model, device="cuda")
results = ev.report(val_loader, class_names=["No Damage", "Minor", "Major", "Destroyed"])
# prints sklearn classification report

fig = ev.confusion_matrix(val_loader)
fig.savefig("cm.png")

per_class = ev.f1_per_class(val_loader)
# {'No Damage': 0.91, 'Minor': 0.78, ...}

Export

from torchloop.exporter import Exporter

exp = Exporter(model, input_shape=(1, 3, 224, 224))
exp.to_onnx("model.onnx")
exp.to_tflite("model.tflite", quantize=True)

Edge Deployment

from torchloop.edge import deploy_to_edge, estimate_model

stats = estimate_model(model, (1, 3, 224, 224), target_device="esp32")
print(f"RAM: {stats['estimated_ram_mb']} MB")
print(f"Latency: {stats['estimated_latency_ms']} ms")

deploy_to_edge(
    model,
    target="esp32",
    input_shape=(1, 3, 224, 224),
    output_path="model.tflite",
    quantize=True,
    quantize_type="int8",
)

---

Design Principles

• No lock-in: Works with any nn.Module. No subclassing required.
• Minimal surface area: Three modules. That's it.
• You own the model: torchloop wraps your loop, doesn't replace your architecture.

---

Roadmap

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

---

License

MIT