def testQuantizeSave(self):
model = QuantizeEmulate(test_utils.build_simple_dense_model(),
**self.params)
model.compile(
loss='categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])
_, keras_file = tempfile.mkstemp('.h5')
keras.models.save_model(model, keras_file)
def testMnistAccuracyinTFLite(self):
num_classes = 10
train_data, test_data, input_shape = test_utils.get_preprocessed_mnist_data(
num_classes=num_classes)
x_train, y_train = train_data
x_test, y_test = test_data
def linear(x):
return x
l = keras.layers
model = keras.Sequential([
QuantizeEmulate(l.Conv2D(32, 5, padding='same', activation='relu'),
input_shape=input_shape,
**self.params),
l.MaxPooling2D((2, 2), (2, 2), padding='same'),
QuantizeEmulate(l.Conv2D(64, 5, padding='same', activation='relu'),
**self.params),
l.MaxPooling2D((2, 2), (2, 2), padding='same'),
l.Flatten(),
QuantizeEmulate(l.Dense(1024, activation='relu'), **self.params),
l.Dropout(0.4),
# TODO(alanchiao): fuse softmax once we've handled it.
# Once we use QuantizeAwareActivation, pre/post activation should be
# handled. Adding dummy activation to force adding of quant operator.
QuantizeEmulate(l.Dense(num_classes, activation=linear),
**self.params),
l.Softmax(),
])
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
model.fit(x_train,
y_train,
batch_size=128,
epochs=1,
validation_data=(x_test, y_test))
tf_accuracy = model.evaluate(x_test, y_test, verbose=0)[1]
# High enough to validate that training is happening, with significantly
# better than 0.1 random accuracy.
self.assertGreater(tf_accuracy, 0.4)
_, keras_file = tempfile.mkstemp('.h5')
_, tflite_file = tempfile.mkstemp('.h5')
keras.models.save_model(model, keras_file)
utils.convert_keras_to_tflite(keras_file, tflite_file,
{'linear': linear})
tflite_accuracy = test_utils.eval_mnist_tflite(tflite_file,
is_quantized=True)
self.assertAlmostEqual(tf_accuracy, tflite_accuracy, delta=0.01)
def testQuantizeEmulateSequential(self):
model = keras.models.Sequential([
self.conv_layer,
self.dense_layer
])
quant_model = QuantizeEmulate(model, **self.params)
self._assert_quant_model(quant_model.layers)
def testQuantizeSaveAndRestore(self):
model = QuantizeEmulate(test_utils.build_simple_dense_model(),
**self.params)
model.compile(
loss='categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])
# Verify serialization correctness persists after training.
model.fit(
np.random.rand(20, 10),
keras.utils.to_categorical(np.random.randint(5, size=(20, 1)), 5),
batch_size=20,
)
_, keras_file = tempfile.mkstemp('.h5')
keras.models.save_model(model, keras_file)
loaded_model = keras.models.load_model(
keras_file,
custom_objects={'QuantizeEmulateWrapper': QuantizeEmulateWrapper})
self._check_models_match(model, loaded_model)
def testQuantizeEmulateList(self):
quant_layers = QuantizeEmulate([self.conv_layer, self.dense_layer],
QuantizationParams(8))
self._assert_quant_model(quant_layers)
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
y_train = tf.keras.utils.to_categorical(y_train, num_classes)
y_test = tf.keras.utils.to_categorical(y_test, num_classes)
l = tf.keras.layers
quant_params = {'num_bits': 8}
model = tf.keras.Sequential([
QuantizeEmulate(l.Conv2D(32, 5, padding='same', activation='relu'),
input_shape=input_shape,
**quant_params),
l.MaxPooling2D((2, 2), (2, 2), padding='same'),
QuantizeEmulate(l.Conv2D(64, 5, padding='same', activation='relu'),
**quant_params),
l.MaxPooling2D((2, 2), (2, 2), padding='same'),
l.Flatten(),
QuantizeEmulate(l.Dense(1024, activation='relu'), **quant_params),
l.Dropout(0.4),
QuantizeEmulate(l.Dense(num_classes), **quant_params),
# TODO(alanchiao): fuse softmax once we've handled it.
l.Softmax(),
])
# Dump graph to /tmp for verification on tensorboard.
graph_def = tf.get_default_graph().as_graph_def()
def testQuantizeEmulateList(self):
quant_layers = QuantizeEmulate([self.conv_layer, self.dense_layer],
**self.params)
self._assert_quant_model(quant_layers)