echelon3 0.5.2

Config-driven PyTorch training framework: assemble nets, data, losses and trainers from YAML

echelon3

Config-driven PyTorch training framework. Every component of a training run — network, dataset, augmentations, losses, metrics, optimizer, scheduler, trainer, export — is described in a YAML config as a module / type / config triple and instantiated dynamically:

net:
  module: echelon3.nets.classifier   # import path (or a path to a .py file)
  type: ClassifierNet                # class or factory function in that module
  config: { ... }                    # constructor kwargs

There is no component registry: anything importable can be plugged in — classes from echelon3, from torch/timm/torchmetrics/albumentations, or from your own project code living next to your configs.

Documentation: https://veryviolet.github.io/echelon3/

Install

pip install echelon3

Train

echelon3-train --config-dir ./configs --config-name my_experiment

CLIs: echelon3-train, echelon3-finetune (warm-start / freeze / head-only), echelon3-evaluate, echelon3-run (inference), echelon3-export (ONNX).

Multi-GPU — built in, no torchrun

Name the GPUs and echelon3 spawns one DDP worker per GPU itself:

echelon3-train --config-dir ./configs --config-name my_experiment gpus=[0,1,2,3]

gpus is a root config key — leave it out and echelon3 uses every visible GPU on the node. dataloaders.train.config.batch_size is the global batch size; it is split across ranks automatically. torchrun (and SLURM srun) still work unchanged for multi-node / elastic jobs.

DataParallel was removed in 0.5.0 — multiple GPUs always run as DDP.

Mixed precision

Training, evaluation and inference use bf16 automatic mixed precision by default on capable GPUs (fp32 on CPU / unsupported GPUs) — a large speedup on modern hardware. Force full fp32 with precision: fp32 under trainer.config (or precision: fp32 at the config root for evaluate / run).

Quick start

examples/ has self-contained smoke runs — a classifier, a CenterNet-style detector and semantic segmentation — each with a synthetic-data generator and a minimal config that trains, validates and checkpoints on CPU or GPU.