def _tile_figs(self, space_width):
n = int(np.ceil(np.sqrt(self.figs.shape[0])))
padding = (((0, n**2 - self.figs.shape[0]), (0, space_width),
(0, space_width)) # add some space between filters
+ ((0, 0), ) * (self.figs.ndim - 3)
) # don't pad the last dimension (if there is one)
if is_intarr(self.figs):
constant_values = 255
else:
constant_values = 1.
tiled_figs = np.pad(self.figs,
padding,
mode='constant',
constant_values=constant_values) # pad with white
# tile the filters into an image
tiled_figs = tiled_figs.reshape(
(n, n) + tiled_figs.shape[1:]).transpose(
(0, 2, 1, 3) + tuple(range(4, tiled_figs.ndim + 1)))
tiled_figs = tiled_figs.reshape((n * tiled_figs.shape[1],
n * tiled_figs.shape[3]) +
tiled_figs.shape[4:])
return tiled_figs[:-space_width, :-space_width,
...] # Delete the padding border
def _tile_figs(self, space_width):
padding = (((0, 0),
(0, space_width)) + ((0, 0), ) * (self.figs.ndim - 2))
if is_intarr(self.figs):
constant_values = 255
else:
constant_values = 1.
tiled_figs = np.pad(self.figs,
padding,
mode='constant',
constant_values=constant_values)
return tiled_figs.reshape((tiled_figs.shape[0] *
tiled_figs.shape[1], ) +
tiled_figs.shape[2:])[:-space_width, ...]
def _tile_figs(self, space_width):
padding = (((0, 0), (0, 0),
(0, space_width)) + ((0, 0), ) * (self.figs.ndim - 3))
if is_intarr(self.figs):
constant_values = 255
else:
constant_values = 1.
tiled_figs = np.pad(self.figs,
padding,
mode='constant',
constant_values=constant_values)
tiled_figs = tiled_figs.transpose((1, 0, 2) +
tuple(range(3, tiled_figs.ndim)))
tile_figs = tiled_figs.reshape(
(tiled_figs.shape[0], tiled_figs.shape[1] * tiled_figs.shape[2]) +
tiled_figs.shape[3:])
return tile_figs[:, :-space_width, ...]
def _tile_figs(self, space_width):
if self.row_num is None and self.col_num is None:
raise ValueError(
f"row_num={self.row_num} or col_num={self.col_num} should not be int, not None"
)
figs_num = self.figs.shape[0]
if self.row_num is None:
self.row_num = int(np.ceil(figs_num / self.col_num))
elif self.col_num is None:
self.col_num = int(np.ceil(figs_num / self.row_num))
n = self.row_num * self.col_num
if n < self.figs.shape[0]:
raise ValueError(f"row_num * col_num = {n} < self.figs.shape[0]")
padding = (((0, n - self.figs.shape[0]), (0, space_width),
(0, space_width)) # add some space between filters
+ ((0, 0), ) * (self.figs.ndim - 3)
) # don't pad the last dimension (if there is one)
if is_intarr(self.figs):
constant_values = 255
else:
constant_values = 1.
tiled_figs = np.pad(self.figs,
padding,
mode='constant',
constant_values=constant_values) # pad with white
# tile the filters into an image
tiled_figs = tiled_figs.reshape(
(self.row_num, self.col_num) + tiled_figs.shape[1:]).transpose(
(0, 2, 1, 3) + tuple(range(4, tiled_figs.ndim + 1)))
tiled_figs = tiled_figs.reshape((self.row_num * tiled_figs.shape[1],
self.col_num * tiled_figs.shape[3]) +
tiled_figs.shape[4:])
return tiled_figs[:-space_width, :-space_width,
...] # Delete the padding border
def resolve_loc_cue_conflict_by_priority(loc_cue: np.ndarray, priority: List[np.ndarray]) -> np.ndarray:
"""Solve conflict in candidate seed according to priority between channels.
Args:
loc_cue: (C, H, W) numpy array. Where (c, h, w) > 0 means channel c is one of the channel of the (h, w)
pixel in sample n
priority: Priority among channels. priority[index] = channels who prior to index'th channel
Returns: (C, H, W) numpy array.
"""
assert loc_cue.ndim == 3
assert is_intarr(loc_cue)
assert len(priority) == loc_cue.shape[0]
one_hot_seed = loc_cue.copy()
for ch_idx, prior_channels in enumerate(priority):
one_hot_seed[ch_idx][np.sum(loc_cue[prior_channels, ...], axis=0) > 0] = 0
assert np.all((np.sum(one_hot_seed, axis=0) <= 1) & np.sum(one_hot_seed, axis=0) >= 0)
return one_hot_seed