def _populate_patches_and_labels(self, xkeys, xys, mode='train'):
np.random.seed(self.seed)
# estimate the size and create h5py dataset
total_frames = self._get_total_frames(xys)
patches_cache, labels_cache = self._create_patch_and_label_dataset(
total_frames)
assert 'dis_y' in xkeys
assert 'dis_u' in xkeys
assert 'dis_v' in xkeys
yss = xys['dis_y'] # yss: Y * frames * videos
uss = xys['dis_u']
vss = xys['dis_v']
if mode == 'train':
assert 'label' in xys
labels = xys['label']
elif mode == 'test':
labels = [None for _ in range(len(yss))]
else:
assert False
assert len(yss) == len(uss) == len(vss) == len(labels)
patch_idx = 0
for ys, us, vs, label in zip(yss, uss, vss, labels): # iterate videos
assert len(ys) == len(us) == len(vs)
for y, u, v in zip(ys, us, vs): # iterate frames
yuvimg = dstack_y_u_v(y, u, v)
img = create_hp_yuv_4channel(yuvimg)
h, w, c = img.shape
adj_h = h - self.patch_height
adj_w = w - self.patch_width
iv, jv = np.meshgrid(np.arange(adj_h),
np.arange(adj_w),
sparse=False,
indexing='ij')
iv = iv.reshape(-1)
jv = jv.reshape(-1)
idx = np.random.permutation(adj_h * adj_w)
iv = iv[idx]
jv = jv[idx]
patches_found = 0
for (y, x) in zip(iv, jv):
patches_cache[patch_idx] = img[y:y + self.patch_height,
x:x + self.patch_width]
if mode == 'train':
labels_cache[patch_idx] = label
patches_found += 1
patch_idx += 1
if patches_found >= self.patches_per_frame:
break
return patches_cache, labels_cache
def _populate_patches_and_labels(self, xkeys, xys, mode='train'):
np.random.seed(self.seed)
# estimate the size and create h5py dataset
total_frames = self._get_total_frames(xys)
patches_cache, labels_cache = self._create_patch_and_label_dataset(
total_frames)
assert 'dis_y' in xkeys
assert 'dis_u' in xkeys
assert 'dis_v' in xkeys
yss = xys['dis_y'] # yss: Y * frames * videos
uss = xys['dis_u']
vss = xys['dis_v']
if mode == 'train':
assert 'label' in xys
labels = xys['label']
elif mode == 'test':
labels = [None for _ in range(len(yss))]
else:
assert False
assert len(yss) == len(uss) == len(vss) == len(labels)
patch_idx = 0
for ys, us, vs, label in zip(yss, uss, vss, labels): # iterate videos
assert len(ys) == len(us) == len(vs)
for y, u, v in zip(ys, us, vs): # iterate frames
yuvimg = dstack_y_u_v(y, u, v)
img = create_hp_yuv_4channel(yuvimg)
h, w, c = img.shape
adj_h = h - self.patch_height
adj_w = w - self.patch_width
iv, jv = np.meshgrid(np.arange(adj_h), np.arange(adj_w),
sparse=False, indexing='ij')
iv = iv.reshape(-1)
jv = jv.reshape(-1)
idx = np.random.permutation(adj_h * adj_w)
iv = iv[idx]
jv = jv[idx]
patches_found = 0
for (y, x) in zip(iv, jv):
patches_cache[patch_idx] = img[y: y + self.patch_height,
x: x + self.patch_width]
if mode == 'train':
labels_cache[patch_idx] = label
patches_found += 1
patch_idx += 1
if patches_found >= self.patches_per_frame:
break
return patches_cache, labels_cache
