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)

frogermcs

LABELS_FILE = 'ImageNetLabels.txt' 
with open(LABELS_FILE) as f:
    labels = f.read().splitlines() 

frogermcs

Prediction for Golden Retriever: golden retriever 0.5627079010009766
Prediction for laptop: laptop 0.42153415083885193

frogermcs

INPUT_TENSOR = "input__0"
OUTPUT_TENSOR = "MobilenetV2__Predictions__Reshape_1__0"
PREDICTED_FEATURE_NAME = "classLabel"

# Prediction is run on CPU
coreml_output_golden = mlmodel.predict({INPUT_TENSOR: img_golden}, useCPUOnly=True)

#Prediction is run on GPU
coreml_output_laptop = mlmodel.predict({INPUT_TENSOR: img_laptop}, useCPUOnly=False)

frogermcs

img_golden = img_golden.resize([224,224], PIL.Image.ANTIALIAS)
img_laptop = img_laptop.resize([224,224], PIL.Image.ANTIALIAS)

frogermcs

# Load previously saved CoreML model of MobileNet v2
mlmodel = coremltools.models.MLModel('mobilenet_v2_1.0_224.mlmodel')

# Get spec from the model
spec = mlmodel.get_spec()
print(spec.description)

# Output 
# >> input {
# >>   name: "input__0"