def demo_attribute(img_nos: list = None, att: Attributer = None):
if att is None:
model_name = VGG
att = Attributer(model_name=model_name)
if img_nos is None:
img_nos = [11, 13, 15]
#img_nos = [6, 97, 278]
for img_no in img_nos:
# image handler for later (method attributions)
ih = ImageHandler(img_no=img_no, model_name=VGG)
# predictions
max_pred, max_p = att.predict_for_model(ih)
plt.figure(figsize=(15, 10))
plt.suptitle(
'Attributions for example {}, prediction = `{}`, probability = {:.2f}'
.format(img_no, max_pred, max_p))
# original image
plt.subplot(2, 4, 1)
plt.axis('off')
plt.title('ImageNet Example {}'.format(img_no))
plt.imshow(
plt.imread(get_image_file_name(IMG_BASE_PATH, img_no) + '.JPEG'))
# annotated image
plt.subplot(2, 4, 2)
plt.title('Annotated Example {}'.format(img_no))
plt.imshow(
plt.imread(
get_image_file_name(ANNOTATE_BASE_PATH, img_no) + '.JPEG'))
# processed image
plt.subplot(2, 4, 3)
plt.title('Reshaped Example')
plt.imshow(demo_resizer(img_no=img_no, target_size=ih.get_size()))
# processed image
plt.subplot(2, 4, 4)
plt.title('Annotation Mask')
plt.imshow(get_mask_for_eval(img_no=img_no, target_size=ih.get_size()),
cmap='seismic',
clim=(-1, 1))
attributions = att.attribute_panel(ih=ih,
methods=METHODS,
save=False,
visualise=False,
take_threshold=False,
take_absolute=False,
sigma_multiple=1)
# show attributions
for i, a in enumerate(attributions.keys()):
plt.subplot(2, 4, 5 + i)
plt.title(a)
plt.axis('off')
plt.imshow(ih.get_original_img(), cmap='gray', alpha=0.75)
plt.imshow(attributions[a],
cmap='seismic',
clim=(-1, 1),
alpha=0.8)
plt.show()
plt.clf()
plt.close()
def get_good_examples(self, cap: int = 1001):
good_examples = []
for i in range(1, cap):
ih = ImageHandler(i, self.curr_model_name)
max_pred, p = self.predict_for_model(ih,
top_n=1,
print_to_stdout=False)
if p > 0.9:
good_examples.append(i)
return good_examples
def attribute_panel_wrapper(model_name: str):
methods = [LIME, LIFT, GRAD]
att = Attributer(model_name)
for i in [11]: # range(6, 7):
ih = ImageHandler(img_no=i, model_name=model_name)
att.attribute_panel(ih=ih,
methods=methods,
save=True,
visualise=True,
take_threshold=True,
take_absolute=True,
sigma_multiple=1)
def get_image_handler_and_mask(self, img_no):
# this gets the image wrapped in the ImageHandler object, and the
# bounding box annotation mask for the image,
# ImageHandler is used to calculate attributions by each method, and the
# mask is used for evaluation
ih = ImageHandler(img_no=img_no, model_name=self.model_name)
# bounding box in the format of the model's input shape / attribution shape
annotation_mask = get_mask_for_eval(img_no=img_no,
target_size=ih.get_size(),
save=False,
visualise=False)
return ih, annotation_mask
def attributer_wrapper(method: str, model: str):
# draw_annotations([i for i in range(16, 300)])
# run some attributions
att = Attributer(model)
start_time = time.time()
for i in range(1, 101):
ih = ImageHandler(img_no=i, model_name=model)
att.attribute(ih=ih,
method=method,
save=False,
visualise=False,
take_threshold=False,
take_absolute=False,
sigma_multiple=1)
print('{} total seconds for {}'.format(str(time.time() - start_time),
method))
def performance_timer():
# performance timing
for model in [INCEPT]:
att = Attributer(model)
for method in [GRAD]:
print('Performance test for {} and {}:'.format(method, model))
start_time = time.time()
for i in range(1, 101):
ih = ImageHandler(img_no=i, model_name=model)
att.attribute(ih=ih,
method=method,
save=False,
visualise=False,
take_threshold=False,
take_absolute=False,
sigma_multiple=1)
print('{} total seconds for {}'.format(
str(time.time() - start_time), method))