echelon3

Describe a PyTorch training run in YAML — network, data, augmentations, losses, metrics, optimizer, scheduler, trainer, export — and run it with one command.

No training-loop boilerplate and no component registry: every piece is a module / type / config triple that echelon3 imports and instantiates, so anything importable drops straight in — classes from echelon3, from torch / timm / torchmetrics / albumentations, or from your own project code sitting next to your configs.

# my_run.yaml — a whole training run
net:  { module: my_pkg.nets, type: MyNet, config: { channels: 32 } }
data:
  train: { module: my_pkg.data, type: MyDataset, config: { split: train } }
  test:  { module: my_pkg.data, type: MyDataset, config: { split: test } }
dataloaders:
  train: { module: torch.utils.data, type: DataLoader, config: { batch_size: 64, shuffle: true } }
  test:  { module: torch.utils.data, type: DataLoader, config: { batch_size: 128 } }
loss:
  - ce: { module: torch.nn, type: CrossEntropyLoss, config: {} }
optimizer: { module: torch.optim, type: AdamW, config: { lr: 3e-4 } }
trainer:   { module: echelon3.trainers.baseline, type: Trainer, config: { epochs: 50 } }
target:    { path: ./out }

echelon3 train -cd . -cn my_run     # trains, validates, keeps the best, checkpoints, resumes

What you get

One CLI, five tasks — echelon3 train | finetune | evaluate | run | export.
Multi-GPU with no torchrun — gpus=[0,1,2,3] spawns one DDP worker per GPU; batch_size is global and split across ranks automatically.
bf16 mixed precision by default on capable GPUs (precision: fp32 to opt out), plus an optional torch.compile knob and TF32.
Batteries included in the trainer — automatic resume, keep-best-checkpoint, validation schedule, TensorBoard / mlops logging, clean Ctrl-C.
Override anything on the CLI — trainer.config.epochs=100 optimizer.config.lr=1e-4, ~scheduler drops a section, +key=… adds one; typed values, list literals, ${oc.env:VAR,default} interpolation, and defaults: config composition (OmegaConf).
Your own code, no registration — reference nets / datasets / losses by import path and run from your repo root (the cwd is on sys.path).

Install

pip install echelon3
# extras: echelon3[export] (ONNX)   echelon3[detection]   echelon3[smp]

CLI & overrides

echelon3 train    -cd ./configs -cn my_run          # train (auto-resumes from target.path)
echelon3 finetune -cd ./configs -cn my_run          # + warm-start / freeze / head-only
echelon3 evaluate -cd ./configs -cn my_run          # score the latest checkpoint
echelon3 export   -cd ./configs -cn my_run          # preprocess→net→postprocess to ONNX
echelon3 run      -cd ./configs -cn my_run          # inference over images / video

# override any config value; +add, ~delete, typed values, multi-GPU:
echelon3 train -cd ./configs -cn my_run \
    trainer.config.epochs=100 optimizer.config.lr=5e-4 \
    dataloaders.train.config.batch_size=256 \
    gpus=[0,1] +trainer.config.compile=true ~scheduler

--config-dir/-cd picks the directory, --config-name/-cn the YAML.

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. For example:

cd examples/segmentation
python gen_seg_data.py --root ./seg_data
SEG_DATA=./seg_data echelon3 train -cd . -cn segmentation_smoke device=cpu

Use it from your AI coding agent

Marketplace-installable plugins teach Codex and Claude Code the echelon3 config format and CLI, so the agent writes correct configs and runs them for you — repo: https://github.com/veryviolet/echelon3-agent-skills.

# Codex
codex plugin marketplace add veryviolet/echelon3-agent-skills
codex plugin add echelon3@veryviolet

# Claude Code (run inside the REPL)
/plugin marketplace add veryviolet/echelon3-agent-skills
/plugin install echelon3@veryviolet

Full documentation

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

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

echelon3-0.11.0.tar.gz (141.7 kB view details)

Uploaded Aug 11, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

echelon3-0.11.0-py3-none-any.whl (185.5 kB view details)

Uploaded Aug 11, 2026 Python 3

File details

Details for the file echelon3-0.11.0.tar.gz.

File metadata

Download URL: echelon3-0.11.0.tar.gz
Upload date: Aug 11, 2026
Size: 141.7 kB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for echelon3-0.11.0.tar.gz
Algorithm	Hash digest
SHA256	`c033db0aea073fe794d18103cd86cbc145d3872554260af6680be49cb8cde075`
MD5	`53167a069845c72a3e3bbd04c93d6da7`
BLAKE2b-256	`76ddc4cb8bf8d88315c5d9ef6cf98aaf1a3bcc27d0d7aab09c1620199ef48dc2`

See more details on using hashes here.

Provenance

The following attestation bundles were made for echelon3-0.11.0.tar.gz:

Publisher: publish.yml on veryviolet/echelon3

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Statement:
- Statement type: https://in-toto.io/Statement/v1
- Predicate type: https://docs.pypi.org/attestations/publish/v1
- Subject name: echelon3-0.11.0.tar.gz
- Subject digest: c033db0aea073fe794d18103cd86cbc145d3872554260af6680be49cb8cde075
- Sigstore transparency entry: 2420428084
- Sigstore integration time: Aug 11, 2026
Source repository:
- Permalink: veryviolet/echelon3@83965edc76ce769f016dc0ddc58d81517f3089b2
- Branch / Tag: refs/tags/v0.11.0
- Owner: https://github.com/veryviolet
- Access: public
Publication detail:
- Token Issuer: https://token.actions.githubusercontent.com
- Runner Environment: github-hosted
- Publication workflow: publish.yml@83965edc76ce769f016dc0ddc58d81517f3089b2
- Trigger Event: push