def test_get_gradient(self):
model = keras.applications.MobileNetV2()
gradient_model = keras.models.Sequential()
gradient_model.add(model)
gradient_model.add(Categorical(7))
gradient_model.get_config()
get_gradient(gradient_model, np.random.random((1, 224, 224, 3)))
get_gradient(gradient_model, np.random.random((1, 224, 224, 3)),
targets=model.get_layer('bn_Conv1').trainable_weights[0])
def test_cut_model(self):
model = keras.applications.MobileNetV2()
head, tail = split_model_by_layer(model, 'block_5_add')
gradient_model = keras.models.Sequential()
gradient_model.add(tail)
gradient_model.add(Categorical(7))
gradients = get_gradient([head, gradient_model], np.random.random((1, 224, 224, 3)))
self.assertEqual(2, len(gradients))
def process(relu_type, target_class):
# Build model and get gradients
model = keras.applications.MobileNetV2()
if relu_type is not None:
model = replace_relu(model, relu_type=relu_type)
gradient_model = keras.models.Sequential()
gradient_model.add(model)
gradient_model.add(Categorical(target_class))
gradients = get_gradient(gradient_model, inputs)
# Visualize gradients
gradient = gradients.numpy()[0]
gradient = (gradient - np.min(gradient)) / (np.max(gradient) -
np.min(gradient) + 1e-4)
gradient = (gradient * 255.0).astype(np.uint8)
return Image.fromarray(gradient)