def forward(self, x, x_baseline=None, class_idx=None, retain_graph=False):
if x_baseline is None:
x_baseline = torch.zeros_like(x)
else:
x_baseline = x_baseline.cuda()
assert x_baseline.size() == x.size()
saliency_map = torch.zeros_like(x) # [1, 3, H, W]
x_diff = x - x_baseline
stdev = self.stdev_spread / (x_diff.max() - x_diff.min())
for alpha in torch.linspace(0., 1., self.n_steps):
x_step = x_baseline + alpha * x_diff
noise = torch.normal(mean=torch.zeros_like(x_step), std=stdev)
x_step_plus_noise = x_step + noise
logit = self.model(x_step_plus_noise)
if class_idx is None:
score = logit[:, logit.max(1)[-1]].squeeze()
else:
score = logit[:, class_idx].squeeze()
self.model.zero_grad()
score.backward(retain_graph=retain_graph)
saliency_map += x_step_plus_noise.grad
saliency_map = saliency_map / self.n_steps
saliency_map = convert_to_gray(saliency_map) # [1, 1, H, W]
saliency_map_min, saliency_map_max = saliency_map.min(), saliency_map.max()
saliency_map = (saliency_map - saliency_map_min) / (saliency_map_max - saliency_map_min).data
return saliency_map
def gray_preprocessing(args):
images_paths = os.listdir(args.img_dir)
for name in images_paths:
image = Image.open(os.path.join(args.img_dir, name))
tf_image = tf.convert_to_tensor(np.array(image).astype(np.float32))
gray_image = utils.convert_to_gray(tf_image[None])[0]
blurred = Image.fromarray(np.uint8(tf.Session().run(gray_image)))
blurred.save(os.path.join(args.output_dir, name))
def forward(self, x, class_idx=None, retain_graph=False):
x.requires_grad_()
logit = self.model(x)
if class_idx is None:
score = logit[:, logit.max(1)[-1]].squeeze()
else:
score = logit[:, class_idx].squeeze()
self.model.zero_grad()
score.backward(retain_graph=retain_graph)
saliency_map = x.grad # [1, 3, H, W]
saliency_map = convert_to_gray(saliency_map.cpu().data) # [1, 1, H, W]
saliency_map_min, saliency_map_max = saliency_map.min(), saliency_map.max()
saliency_map = (saliency_map - saliency_map_min) / (saliency_map_max - saliency_map_min)
return saliency_map
def prepare_frame():
frame = capturer.get_frame()
if frame is not None:
frame = flip_frame(frame)
return convert_to_gray(frame)
def detect_face_in_frame(self, a_frame):
a_frame = convert_to_gray(a_frame)
a_frame = equalize(a_frame)
faces_list = self.face_classifier.detect_multiscale(a_frame)
return faces_list[0] if len(faces_list) == 1 else None
