TorchTL – A minimal training loop abstraction for PyTorch (github.com)

TorchTL is a compact training-loop abstraction for PyTorch that provides a drop-in, dependency-free way to run common training workflows. Released as a pip package (pip install torchtl) under the Apache 2.0 license, it wraps standard PyTorch models, optimizers and losses (no subclassing required) and exposes a simple Trainer API (Trainer.fit, train_epoch, validate). The library focuses on readability and predictability while automating routine tasks: device management (CPU/CUDA), mixed-precision, gradient accumulation and clipping, checkpointing with resume, early stopping, LR scheduling, progress reporting, exponential moving average (EMA), and support for multiple batch formats. For practitioners and researchers the significance is practical: TorchTL reduces boilerplate and enforces best practices without locking you into an opinionated framework. It’s extensible via a callback system (progress, checkpoint, early-stopping, LR-scheduler callbacks are included and you can add custom Callback subclasses), and includes utilities like count_params, freeze/unfreeze layers, set_seed, get/set_lr, and EMA helpers. Typical usage shows working with DataLoader, Adam, MSELoss, mixed_precision=True, grad_acc_steps, and saving/loading checkpoints. By keeping only PyTorch as a dependency and exposing familiar primitives, TorchTL aims to speed prototyping, improve reproducibility, and simplify production handoffs while staying transparent and debuggable.