solver-pytorch

solver-pytorch provides the lightweight training infrastructure behind thsolver. It wraps the repetitive parts of PyTorch experiments such as configuration parsing, dataloader setup, checkpointing, logging, distributed training, and evaluation, while leaving the model and batch logic in user code.

The project is intentionally small. Instead of introducing a large framework, it gives you a base thsolver.solver.Solver, a simple filelist-based thsolver.dataset.Dataset, registry helpers for models and datasets, and utility modules for samplers, learning-rate schedules, and metric tracking.

Key benefits of thsolver include:

• Simple experiment loops. Subclass thsolver.solver.Solver and implement only the hooks that define your task.

• Config-driven runs. Use yacs configuration files with command-line overrides and automatic config backups in the log directory.

• Built-in training utilities. Reuse checkpointing, TensorBoard logging, learning-rate scheduling, mixed precision, and DDP launch modes.

• Easy extension points. Register model and dataset factories with small decorators and keep the rest of the pipeline generic.

Notes

• Installation
• Getting Started
• Configuration
• Extending thsolver
• Distributed Training

Package Reference

• thsolver.config
• thsolver.solver
• thsolver.dataset
• thsolver.registry
• thsolver.sampler
• thsolver.lr_scheduler
• thsolver.tracker

Indices and Tables

• Index

• Module Index