qmb-mvp (0.2.0)

2026-04-23T17:59:24-04:00

To install the package using pip, run the following command:

pip install --index-url /api/packages/qoherent/pypi/simple/ --no-deps qmb-mvp

For more information on the PyPI registry, see the documentation.

Add your description here

qmb-mvp

qmb-mvp is a command-line toolkit for RF model training, evaluation, export, hyperparameter search, and pruning. The repository now uses one YAML-driven runtime rooted in src/qmb; the old Hydra/Prefect stack has been removed.

Install

uv venv --python 3.13
uv pip install torch torchvision
uv sync --inexact

For CPU-only installs:

UV_TORCH_BACKEND=cpu uv pip install torch torchvision
UV_TORCH_BACKEND=cpu uv sync --inexact

CLI

qmb --help
qmb train --config path/to/train.yaml
qmb eval path/to/checkpoints/best.ckpt --split test
qmb export path/to/checkpoints/best.ckpt
qmb hpo --config path/to/hpo.yaml
qmb prune path/to/checkpoints/best.ckpt --config path/to/prune.yaml

Training writes timestamped runs under outputs/ by default with logs, checkpoints, resolved config snapshots, metrics, evaluation artifacts, and optional ONNX exports.

Config Shape

Training configs are plain YAML mappings with these top-level sections:

task
data
model
optimization
runtime
evaluation
export

Common features supported by the current runtime include:

IQ HDF5 datasets for legacy, structured, Curator, DeepSig 2018, and auto-detection modes
Split-specific transform pipelines
Torch and timm optimizers and schedulers
IQ model families including Tiny CNN, LeNet, VTCNN2, ResNet1D, MobileNetV3, EfficientNetV2, and timm-backed IQ classifiers
Resume, early stopping, gradient clipping, torch.compile, ONNX export, Optuna HPO, and structured pruning

See the example configs under documentation/v2/examples for working YAML layouts.

Built-In Tasks

classification is the only built-in runnable task.
regression is a stub and should be implemented as a custom extension task.

Extension Authoring

The primary supported customization workflow is a copyable extension folder whose Python files import only from the public qmb.api* surface. Start from the reference folders under documentation/v2/examples/extensions/:

minimal_task_extension/ for a custom task that reuses built-in qmb datasets, models, and losses
full_task_extension/ for a custom task plus custom dataset, model, and dataset config resolver

Those folders can live on the same machine running qmb, or in a repository that the runner clones before execution. The runtime then loads the extension through:

runtime:
  component_modules:
    - /path/to/my_qmb_extension

runtime.component_modules supports two forms:

package directories containing __init__.py
direct .py module paths

Relative filesystem inputs in qmb configs are resolved from the config file location. That lets examples/train.yaml use component_modules: [".."] to load the parent extension directory without depending on the caller's working directory.

Resolved input paths are persisted into run artifacts and checkpoints, so later eval, export, and prune commands remain portable across working directories.

qmb.api.scaffold remains available for internal tooling that needs to materialize the packaged template assets, but it is not required for the primary user workflow.

Layout

src/qmb/
|-- artifacts/
|-- api/
|-- cli/
|-- compression/
|-- core/
|-- data/
|-- datasets/
|-- engine/
|-- evaluation/
|-- export/
|-- losses/
|-- models/
|-- tasks/
`-- transforms/

custom_components/ is now a repo-local/internal example area. It is useful for developing and testing project-specific tasks inside this repository, but it is not the primary user workflow and is not shipped in package builds.

Testing

uv sync --group dev
uv run --group dev pytest

Requires Python: >=3.13

Details

PyPI

2026-04-23 17:59:24 -04:00

2

250 KiB

Assets (2)

qmb_mvp-0.2.0-py3-none-any.whl 149 KiB

qmb_mvp-0.2.0.tar.gz 101 KiB

Versions (2) View all

0.2.0

2026-04-23

0.1.6

2026-02-03