MDI Machine Learning Classifier

This capstone application combines the MDI desktop framework, Swing, sPlot, and ONNX Runtime in an end-to-end image-classification workflow.

Requirements

• Java 17 or newer
• A locally installed io.github.heddle:mdi:1.2.2-SNAPSHOT
• An ONNX image-classification model

The default model location is models/mobilenetv2-12.onnx; the ImageNet label file is already stored at models/imagenet_labels.txt. Large model binaries are not committed to this repository. Obtain MobileNet V2 from the ONNX Model Zoo or supply another compatible single-input, single-output RGB classifier.

Build and run

mvn clean package
mvn exec:java \
  -Dexec.mainClass=edu.cnu.mdi.mlclassifier.app.ClassifierApp \
  -Dexec.args="--model=/path/to/model.onnx --labels=/path/to/labels.txt"

The application also accepts --model=/path/to/model.onnx and --labels=/path/to/labels.txt. When omitted, paths are resolved relative to the launch directory.

Open an image by dropping it onto the classifier view or by choosing Image > Open Image…. Successfully opened files are retained in Image > Recent Images.

Use Model > Open ONNX Model… to load or replace the classifier without restarting the application. The view keeps a recent-model list and offers separate controls for the labels file and input normalization. If the startup model is missing, the application remains usable and immediately offers a model chooser.

Each recent model retains its own labels path and normalization setting. The last successfully loaded profile is restored on the next launch; explicit --model or --labels command-line arguments take precedence.

Use Results > Number of Classes to request 1, 3, 5, 10, or 20 ranked classifications. The selection is remembered, and changing it reclassifies the current image without blocking the Swing event-dispatch thread.

Completed classifications are retained in a bounded, in-memory comparison history. Results > Comparison History… compares model, prediction, confidence, inference time, uncertainty, model size, and normalization. The latest run can be copied as text, and the complete history can be saved as CSV with one row per ranked class.

Predicted probability is presented as model confidence, not as guaranteed correctness or calibrated real-world accuracy. Model > Find Models Online… opens the ONNX Model Zoo when desktop browsing is supported. The platform menu shortcut opens an image; adding Shift opens a model.

Obtaining models

The historical ONNX Model Zoo is now preserved rather than actively updated, and its model files have moved to the ONNX Model Zoo organization on Hugging Face. The default model used by this project is the 14 MB MobileNet V2 opset-12 model. Download mobilenetv2-12.onnx, then open it from the Model menu or place it in models/ before starting the application. The repository already includes the corresponding ImageNet labels file.

Model downloads are deliberately not initiated by the application. Models can be large, have model-specific licenses and preprocessing requirements, and may use external data files. Keeping acquisition explicit avoids an implicit network dependency and lets the user verify the model card and license first.

This demo expects a floating-point RGB image classifier with one rank-4 input and one output. A model must be paired with the labels and normalization its model card specifies; merely having an .onnx extension does not guarantee that it is compatible with this application.

Pipeline

1. Inspect the ONNX input tensor and determine NCHW or NHWC layout.
2. Center-crop the source to the model aspect ratio.
3. Resize and apply the configured channel normalization.
4. Run inference on a dedicated worker thread.
5. Preserve probability outputs or apply stable softmax to logits.
6. Present top classes in an sPlot bar chart and expose diagnostics through MDI feedback.

The application assumes one rank-4 RGB input and one floating-point output. The model's preprocessing requirements must match the selected normalization.

Tests

The test suite includes a generated 177-byte ONNX RGB classifier under src/test/resources. It exercises real ONNX Runtime session creation, tensor layout discovery, preprocessing, synchronous and asynchronous inference, labels, structured measurements, and resource cleanup without depending on the large models in models/. The reproducible generator is retained under src/test/scripts and requires the Python onnx package only when regenerating the fixture.