frogermcs

@RunWith(AndroidJUnit4.class)
@LargeTest
public class MLModelTest {

    @Rule
    public ActivityTestRule<ModelTestActivity> mainActivityActivityRule = new ActivityTestRule<>(ModelTestActivity.class);

    @Test
    public void testClassificationUI() {
        ModelTestActivity activity = mainActivityActivityRule.getActivity();

frogermcs

public class ModelTestActivity extends AppCompatActivity {

    private ImageView ivPreview;
    private TextView tvClassification;
    private ModelClassificator modelClassificator;

    @Override
    protected void onCreate(@Nullable Bundle savedInstanceState) {
        super.onCreate(savedInstanceState);
        setContentView(com.frogermcs.imageclassificationtester.test.R.layout.activity_model_test);

frogermcs

from PIL import Image

VAL_BATCH_DIR = "validation_batch"
!mkdir {VAL_BATCH_DIR}

# Export batch to *.jpg files with specific naming convention.
# Make sure they are exported in the full quality, otherwise the inference
# process will return different results. 

for n in range(32):

frogermcs

public class ModelClassificator {
    private static final int MAX_CLASSIFICATION_RESULTS = 3;
    private static final float CLASSIFICATION_THRESHOLD = 0.2f;

    private final Interpreter interpreter;
    private final List<String> labels;
    private final ModelConfig modelConfig;

    public ModelClassificator(Context context,
                              ModelConfig modelConfig) throws IOException {

frogermcs

public class ClassificationFrameProcessor implements FrameProcessor {

    private final ModelClassificator modelClassificator;
    private final ClassificationListener classificationListener;

    public ClassificationFrameProcessor(ModelClassificator modelClassificator,
                                        ClassificationListener classificationListener) {
        this.modelClassificator = modelClassificator;
        this.classificationListener = classificationListener;
    }

frogermcs

public class MainActivity extends AppCompatActivity
        implements ClassificationFrameProcessor.ClassificationListener {

    private CameraView cameraView;
    private TextView tvClassification;
    private ClassificationFrameProcessor classificationFrameProcessor;

    @Override
    protected void onCreate(Bundle savedInstanceState) {
        super.onCreate(savedInstanceState);

frogermcs

Output, CoreML
(CPU) Prediction for Golden Retriever: golden retriever 0.611853480339
(GPU) Prediction for laptop: notebook 0.515091240406

Output, TensorFlow
Prediction for Golden Retriever: golden retriever 0.61186796
Prediction for laptop: notebook 0.51475537

frogermcs

TF_INPUT_TENSOR = 'input:0'
TF_OUTPUT_TENSOR = 'MobilenetV2/Predictions/Reshape_1:0'

with tf.Session(graph = g) as sess:
    tf_laptop_out = sess.run(TF_OUTPUT_TENSOR, feed_dict={TF_INPUT_TENSOR: img_laptop_tf})
    tf_golden_out = sess.run(TF_OUTPUT_TENSOR, feed_dict={TF_INPUT_TENSOR: img_golden_tf})
tf_laptop_out = tf_laptop_out.flatten()    
tf_golden_out = tf_golden_out.flatten()    
laptop_idx = np.argmax(tf_laptop_out)
golden_idx = np.argmax(tf_golden_out)

frogermcs

# Load TensorFlow frozen model
TF_FROZEN_MODEL = "mobilenet_v2_1.0_224_frozen.pb"
with open(TF_FROZEN_MODEL, 'rb') as f:
    serialized_model = f.read()
tf.reset_default_graph()
graph_definition = tf.GraphDef()
graph_definition.ParseFromString(serialized_model)

with tf.Graph().as_default() as g:
    tf.import_graph_def(graph_definition, name='')

frogermcs

# Prepare images for TensorFlow requirements

#Convert to expected type: float
img_laptop_tf = np.array(img_laptop).astype(np.float32) 
# Setup expected input shape: [1,224,224,3]
img_laptop_tf = np.expand_dims(img_laptop_tf, axis = 0) 
# Convert to expected values ranges: [0, 1]
img_laptop_tf = (1.0/255.0) * img_laptop_tf 

print( 'Image shape:', img_laptop_tf.shape)