Model vs application responsibilities

Inspect the Android code to understand which parts of the deployment are handled by ExecuTorch, and which parts remain application responsibilities.

Application architecture

The application uses one Android Activity for the screen and separates model execution into small Java classes. The Java source is under:

    

        
        
module-3-executorch-on-android-semantic-segmentation/android/ObjectCutout/app/src/main/java/
  org/arm/education/executorch/objectcutout/

Java file	Responsibility
`MainActivity`	Builds the UI, opens the Android image picker, submits model work to the background executor, displays results on the UI thread, and closes the model when the Activity is destroyed
`ExecuTorchSegmenter`	Owns the ExecuTorch `Module` and coordinates preprocessing, `Module.forward()`, postprocessing, and timing
`AssetUtils`	Reads JSON assets and copies the packaged `.pte` file into app-private storage
`ModelSpec`	Loads `model_spec.json` and checks tensor shapes, NCHW layout, data types, resize settings, normalization values, and class counts
`ImagePreprocessor`	Converts an Android `Bitmap` into the letterboxed, normalized NCHW float array expected by the model
`MaskPostprocessor`	Converts output logits into a class mask, removes letterbox padding, creates the transparent cutout, and applies the optional highlight color
`CutoutResult`	Carries the output bitmaps, alpha mask, selected label, foreground size, and timing values back to the UI
`CheckerboardDrawable`	Draws the checkerboard behind the output image so transparent pixels are visible

Before reading the code excerpts, confirm that you understand where each deployment responsibility lives:

Responsibility	File you expect to own it
Copy `.pte` from assets into app-private storage
Load the ExecuTorch module
Decode a selected image URI
Convert a bitmap into NCHW float data
Run `Module.forward()`
Convert logits into a class mask
Keep inference off the UI thread
Release native runtime resources

The main data flow is:

Image Alt Text:Data flow from the packaged model and selected image through the Android application, ExecuTorch inference, and cutout processing.

The model pipeline remains outside MainActivity, separating inference code from Android UI and lifecycle code.

The following pages will walk through the code, and how it achieves these steps:

load the packaged .pte program
decode the selected image into a Bitmap
letterbox and normalize the image into an NCHW (batch, channels, height, width) float array
create an ExecuTorch Tensor and call Module.forward()
convert the logits into an alpha mask and cutout
return the result to the UI thread and release the module with the Activity

Model vs application responsibilities

Application architecture

Nice work