def test_activation_maximization(vgg):
deo = Model(vgg)
layer_list = [layer.name for layer in vgg.layers[1:4]]
for i, layer in enumerate(layer_list):
outs = deprocess_image(
deo.activation_maximization(layer, epochs=5, step_size=2)[0])
def test_max_filter(vgg):
deo = Model(vgg)
img = load_image('tests/img/cat.jpg')
layer_list = [layer.name for layer in vgg.layers[1:4]]
for i, layer in enumerate(layer_list):
deo.max_filter(layer, img)
def test_guided_backprop(vgg):
# load the vgg into the deoxys model
deo = Model(vgg)
# load an real image
img = load_image('tests/img/cat.jpg')
layer_list = [layer.name for layer in vgg.layers[1:4]]
for i, layer in enumerate(layer_list):
outs = deprocess_image(deo.guided_backprop(layer, img, mode='mean')[0])
def test_deconvnet(vgg):
# load the vgg into the deoxys model
deo = Model(vgg)
# load an real image
img = load_image('tests/img/cat.jpg')
layer_list = [layer.name for layer in vgg.layers[1:4]]
for i, layer in enumerate(layer_list):
deprocess_image(deo.deconv(layer, img, mode='min')[0])
def test_activation_map(vgg):
# load the vgg into the deoxys model
deo = Model(vgg)
# load an real image
img = load_image('tests/img/cat.jpg')
# View activation map, 1st filters
layer_list = [layer.name for layer in vgg.layers[1:4]]
for i, layer in enumerate(layer_list):
outs = deprocess_image(deo.activation_map(layer, img)[0][..., 0])
def load_image(path, target_size=(224, 224)):
x = image.load_img(path, target_size=target_size)
x = image.img_to_array(x)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
return x
if __name__ == '__main__':
vgg = vgg16.VGG16(weights='imagenet', include_top=True)
vgg.summary()
# load the vgg into the deoxys model
deo = Model(vgg)
# load an real image
img = load_image('../../test_img/cat.jpg')
# predict what is in that image
preds = deo.predict(img)
predicted_class = preds.argmax(axis=1)[0]
print("predicted top1 class:", predicted_class)
print('Predicted:', decode_predictions(preds, top=1)[0])
# View activation map, 1st filters
layer_list = [layer.name for layer in vgg.layers[1:19]]
nrow, ncol = 4, 5