10.4Testing Model Accuracy on Device

The same model may produce slightly different results on different hardware due to floating-point precision differences between GPU, CPU, and NPU inference.

def test_model_accuracy_matches_server(driver):
    """On-device predictions must agree with server-side predictions."""
    validation_set = load_validation_set("test_data/validation/")
    agreements = 0
    total = len(validation_set)

    for image_path, expected_label in validation_set:
        # Get on-device prediction
        driver.push_file("/sdcard/test_images/current.jpg", source_path=image_path)
        driver.find_element(AppiumBy.ACCESSIBILITY_ID, "classify-image").click()
        result = driver.find_element(AppiumBy.ACCESSIBILITY_ID, "classification-result")
        device_prediction = result.text

        # Compare with server-side prediction
        if device_prediction.lower() == expected_label.lower():
            agreements += 1

        driver.find_element(AppiumBy.ACCESSIBILITY_ID, "clear-result").click()

    accuracy = agreements / total
    assert accuracy >= 0.99, \
        f"On-device accuracy {accuracy:.2%} is below 99% threshold"

def test_model_handles_edge_case_inputs(driver):
    """Model should handle adversarial and edge case inputs gracefully."""
    edge_cases = [
        ("test_data/blank_white.jpg", "no_object_detected"),
        ("test_data/blank_black.jpg", "no_object_detected"),
        ("test_data/tiny_1x1.jpg", "invalid_input"),
        ("test_data/very_large.jpg", "cat"),  # Should still classify correctly
        ("test_data/rotated_90.jpg", "cat"),  # Rotation invariance
        ("test_data/low_quality.jpg", "cat"),  # JPEG quality = 10
    ]

    for image_path, expected_category in edge_cases:
        driver.push_file("/sdcard/test_images/edge.jpg", source_path=image_path)
        driver.find_element(AppiumBy.ACCESSIBILITY_ID, "classify-image").click()

        # Should not crash regardless of input
        result = driver.find_element(AppiumBy.ACCESSIBILITY_ID, "classification-result")
        assert result.is_displayed(), f"Model crashed on {image_path}"

        driver.find_element(AppiumBy.ACCESSIBILITY_ID, "clear-result").click()