Build, test, and run MIRA.

Install MIRA from PyPI, download a trained model, and open a local camera stream. I've tested this on Windows PowerShell — on macOS use source .venv/bin/activate.

1. Install from PyPI

python -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install mira-ai

Package: mira-ai on PyPI

2. Download a model

mira download --list
mira download mira_exp019.pt

3. Check the environment

mira doctor
mira models

4. Start detection

mira live --model mira_exp019.pt

mira.yaml is the shared configuration source for classes, paths, training defaults, inference thresholds, and export settings. CLI options override these values for one run.

classes:
  names: [glass, metal, paper, plastic, trash]

training:
  default_model: yolo11n.pt
  default_epochs: 120
  default_batch_size: 32
  default_imgsz: 640

inference:
  reject_threshold: 0.55
  default_conf: 0.5
  default_iou: 0.45

Inspect the effective configuration with .\mira config.

The system separates camera capture, inference, presentation, and research tooling so each part can be tested independently.

Key modules

Launch the interactive model picker or specify a model directly. I usually just run .\mira live without arguments and pick from the list.

.\mira live
.\mira live --model mira_exp019.pt
.\mira live --model mira_exp019.pt --camera 1 --resolution 1280x720
.\mira live --model mira_exp019.pt --conf 0.25 --reject 0.55

Confidence tiers

The dashboard runs locally. It does not upload camera frames to a public server. I've been running it on 127.0.0.1:8000 during development — it doesn't need internet.

.\mira dashboard
.\mira dashboard --port 8080
.\mira dashboard --host 0.0.0.0

Open http://127.0.0.1:8000. Select a camera, load a model, and start the stream. The interface exposes class distribution, recent detections, FPS, latency, CPU, and memory history.

Dataset files stay outside Git. Registry metadata and processing code keep their origins and transformations traceable. I've kept the raw zips out of git since they're a few GB.

.\mira datasets
.\mira merge --sources taco_trashnet roboflow --output datasets/mira_tnr
.\mira validate --dataset datasets/mira_tnr

Source families

• TACO for litter in varied contexts.
• TrashNet for material classification images.
• Roboflow Trash Detection for YOLO-formatted waste objects.
• Project-specific tabletop data for controlled detector training.

Experiment YAML files record model, dataset, image size, epochs, batch size, and augmentation. Result serialization adds timing and Git metadata.

.\mira train --config experiments/exp014_yolo11n_multidataset.yaml
.\mira eval-yolo --model models/detection/mira_exp019.pt
.\mira benchmark --models models/detection/mira_exp019.pt models/detection/mira_exp019_int8_640.tflite
.\mira generate kaggle --config experiments/exp014_yolo11n_multidataset.yaml

What gets measured

Export the PyTorch reference model, then benchmark the converted artifact before choosing it for deployment.

.\mira export --model models/detection/mira_exp019.pt --formats tflite_int8
.\mira diagnostics

Run .\mira models to list all discovered artifacts. Model metadata sidecars describe task, architecture, classes, and expected input shape.

The local FastAPI service exposes control and observation endpoints.

• Crumpled white paper can be confused with plastic at high confidence.
• End-on metal cans are underrepresented in training data.
• Heavy overlap and occlusion reduce bounding-box quality.
• The broad trash category remains difficult outside controlled tabletop validation.
• Raspberry Pi and robotic sorting benchmarks are pending.
• The dashboard has unit coverage but no tracked end-to-end camera and browser verification artifact.