def forward(self, focus):
focus = focus['__train__']
frames = util.extract_neighbor(self.frame_source,
focus,
np.arange(focus.shape[0]),
nb_size=self.nb_size)
frame_dev = self._deviation_1(frames)
focus_dev = self._deviation_2(focus)
frame_var = self._variance(frames)
belief_dev = 0
if self.belief_dev_coeff != 0:
pre_frames = util.extract_neighbor(
self.frame_source,
focus[1:],
np.arange(focus.shape[0] - 1),
nb_size=self.nb_size) # I_{t}(x_{t+1})
post_frames = util.extract_neighbor(
self.frame_source,
focus[:-1],
np.arange(focus.shape[0] - 1) + 1,
nb_size=self.nb_size) # I_{t+1}(x_{t})
belief_dev = self._belief_deviation(focus, frames, pre_frames,
post_frames)
if self.verbose:
logger.info(
'frame_dev= %f, focus_dev= %f, frame_var= %f, belief_dev= %f' %
(frame_dev, focus_dev, frame_var, belief_dev))
return self.frame_dev_coeff*frame_dev \
+ self.focus_dev_coeff*focus_dev \
- self.frame_var_coeff*frame_var \
- self.belief_dev_coeff*belief_dev
def save_focus_img(dataset, all_focus, save_path, changepoints=[]):
focus_img = util.extract_neighbor(dataset,
all_focus,
range(len(all_focus)),
nb_size=(15, 20))
for i, img in enumerate(focus_img):
# focused neighbor image
file_name = 'focus_img_%d.png' % (i)
file_path = os.path.join(save_path, file_name)
imsave(file_path, img)
print('saved under', save_path)
pos_cnt = np.zeros(dataset.get_shape()[-2:])
cp_mask = np.zeros(len(all_focus), dtype=bool)
cp_mask[changepoints] = True
save_subpath = util.get_dir(os.path.join(save_path, 'marker'))
for i, img in enumerate(dataset.get_frame(0, len(all_focus))):
# for i, img in enumerate(util.remove_mean(dataset.get_frame(0, len(all_focus)), all_focus)):
# marker
file_name = 'marker_%d.png' % (i)
marker_file_path = os.path.join(save_subpath, file_name)
marker_pos = (all_focus[i] * dataset.get_shape()[2:]).astype(int)
img[0, marker_pos[0], :] = 1
img[0, :, marker_pos[1]] = 1
if cp_mask[i]:
img = 1 - img
pos_cnt[marker_pos[0], marker_pos[1]] += 1
imsave(marker_file_path, img[0])
print('focus by marker saved under', save_subpath)
if np.sum(pos_cnt) > 0:
pos_cnt = np.sqrt(1.0 - (pos_cnt / np.max(pos_cnt) - 1)**2)
pos_cnt_file_path = os.path.join(save_path, 'counter_pos.png')
imsave(pos_cnt_file_path, pos_cnt)
print('position count saved at', pos_cnt_file_path)
def plot_focus(dataset, indices, all_focus):
PX, PY = 2, len(all_focus)
focus_img = util.extract_neighbor(dataset,
all_focus,
indices,
nb_size=(15, 20))
for i in range(PY):
plt.subplot(PX, PY, i + 1)
plt.imshow(dataset.get_frame(indices[i])[0], interpolation=None)
plt.subplot(PX, PY, PY + i + 1)
plt.imshow(focus_img[i], interpolation=None)
plt.show()
def from_focus_model(focus_model_forward, dataset, nb_size, batch_size=100):
n_channel = 1 # dataset.get_shape()[-3] # TODO: properly do this
match_size = (n_channel, nb_size[0]*2+1, nb_size[1]*2+1)
img_stack = np.zeros((dataset.n_state,) + match_size)
weight_stack = np.zeros(dataset.n_state)
cur_n = 0
for l in range(0, dataset.n_state, batch_size):
print(l) # TODO: remove
r = min(l + batch_size, dataset.n_state)
frames = dataset.get_frame(l, r)
focus, extra = focus_model_forward(frames)
neighbor = util.extract_neighbor(dataset, focus, np.arange(l, r),
nb_size=nb_size)
img_stack[l:r, 0] = neighbor # properly channels?
weight_stack[l:r] = util.focus_intensity(focus[l:r], extra[l:r])
img_mean = np.average(img_stack, weights=weight_stack, axis=0)
img_var = np.average((img_stack - img_mean)**2, weights=weight_stack,
axis=0)**0.5 # TODO: this is biased
return ModelFocusMeanVar(img_mean, img_var)