def compute_mws_segmentation(input_, offsets, n_attractive_channels, strides,
randomize_strides):
vol_shape = input_.shape[1:]
mst = mws.MutexWatershed(np.array(vol_shape), offsets,
n_attractive_channels, strides)
if randomize_strides:
mst.compute_randomized_bounds()
sorted_edges = np.argsort(input_.ravel())
mst.repulsive_mst_cut(sorted_edges)
segmentation = mst.get_flat_label_image().reshape(vol_shape)
return segmentation
def mws_segmentation_old(affs, stride=np.array([1, 1]),
randomize_bounds=False):
t0 = time.time()
vol_shape = affs.shape[1:]
sorted_edges = np.argsort(affs.ravel())
# run the mst watershed
seperating_channel = 2
mst = mws.MutexWatershed(np.array(vol_shape), OFFSETS,
seperating_channel, stride)
if randomize_bounds:
mst.compute_randomized_bounds()
mst.repulsive_mst_cut(sorted_edges)
segmentation = mst.get_flat_label_image().reshape(vol_shape)
return segmentation, time.time() - t0
def run_mws(affinities,
offsets, stride,
seperating_channel=2,
invert_dam_channels=True,
bias_cut=0.,
randomize_bounds=True):
assert len(affinities) == len(offsets), "%s, %i" % (str(affinities.shape), len(offsets))
affinities_ = np.require(affinities.copy(), requirements='C')
if invert_dam_channels:
affinities_[seperating_channel:] *= -1
affinities_[seperating_channel:] += 1
affinities_[:seperating_channel] += bias_cut
sorted_edges = np.argsort(affinities_.ravel())
# run the mst watershed
vol_shape = affinities_.shape[1:]
mst = mws.MutexWatershed(np.array(vol_shape),
offsets,
seperating_channel,
stride)
if randomize_bounds:
mst.compute_randomized_bounds()
mst.repulsive_ucc_mst_cut(sorted_edges, 0)
segmentation = mst.get_flat_label_image().reshape(vol_shape)
return segmentation
def run_mws(affinities, offsets, stride, var=None, cons_aff=False):
assert len(affinities) == len(offsets), "%s, %i" % (str(
affinities.shape), len(offsets))
affinities_ = np.require(affinities.copy(), requirements='C')
vol_shape = affinities_.shape[2:]
print(vol_shape)
affinities_ = affinities_.ravel()
if var is not None:
var_ = var.ravel()
# get ids and values, sorted by value
tmp = np.array(
list(zip(var_, affinities_, list(range(len(affinities_))))))
tmp5 = tmp[tmp[:, 0].argsort()[::-1]].transpose()
_, sorted_edgesAff, sorted_edgesAffIds = tmp5
elif cons_aff:
affs = []
reps = []
print("splitting affs,reps")
# for idx, a in enumerate(affinities_):
# if a < 0:
# reps.append((a, idx))
# elif a > 0:
# affs.append((a, idx))
affs = zip(affinities_[affinities_ > 0], np.where(affinities_ > 0)[0])
affs = np.array(list(affs))
print(affs.shape)
print("got affs")
reps = zip(-affinities_[affinities_ < 0], np.where(affinities_ < 0)[0])
# reps = zip(-affinities_[affinities_<=0], np.where(affinities_<=0)[0])
reps = np.array(list(reps))
print(reps.shape)
print("done")
print(len(reps), len(affs))
print("sorting")
affs_t = affs[affs[:, 0].argsort()[::-1]].transpose()
sorted_edgesAff, sorted_edgesAffIds = affs_t
reps_t = reps[reps[:, 0].argsort()[::-1]].transpose()
sorted_edgesRep, sorted_edgesRepIds = reps_t
print("done")
else:
print("getting reps")
repulsions_ = 1.0 - affinities_
repulsions_ = repulsions_.ravel()
print("sorting affs")
tmp = np.array(list(zip(affinities_, list(range(len(affinities_))))))
tmp5 = tmp[tmp[:, 0].argsort()[::-1]].transpose()
sorted_edgesAff, sorted_edgesAffIds = tmp5
print("sorting reps")
tmp = np.array(list(zip(repulsions_, list(range(len(repulsions_))))))
tmp5 = tmp[tmp[:, 0].argsort()[::-1]].transpose()
sorted_edgesRep, sorted_edgesRepIds = tmp5
print(sorted_edgesAff.shape, sorted_edgesRep.shape)
var = None
# run the mst watershed
print("starting mws")
seperating_channel = 3 # ignored
mst = mws.MutexWatershed(np.array(vol_shape), offsets, seperating_channel,
stride)
mst.repulsive_ucc_mst_cut(sorted_edgesAff, sorted_edgesAffIds,
sorted_edgesRep, sorted_edgesRepIds,
len(sorted_edgesAff), len(sorted_edgesRep), 0,
var is not None)
segmentation = mst.get_flat_label_image().reshape(vol_shape)
return segmentation
def test_cuf():
C_ATTRACTIVE = 4
EDGE_LEN = 512
out_label = np.empty((30, EDGE_LEN, EDGE_LEN), dtype=np.int64)
if C_ATTRACTIVE == 2:
offsets = np.array([[-1, 0], [0, -1], [-27, 0], [0, -27]],
dtype=np.int)
elif C_ATTRACTIVE == 3:
offsets = np.array(
[[-1, 0], [0, -1], [-1, -1], [-27, 0], [0, -27], [-27, -27]],
dtype=np.int)
else:
offsets = np.array([[-1, 0], [0, -1], [-1, -1], [+1, -1], [-27, 0],
[0, -27], [-27, -27], [27, -27]],
dtype=np.int)
for z_slice in [15]:
with h5py.File(
"/mnt/localdata01/swolf/ws/isbi/isbi_for_steffen/isbi_test_100k.h5",
"r") as aff:
if C_ATTRACTIVE == 2:
affinities = aff["data"][:, z_slice, :EDGE_LEN, :EDGE_LEN][[
0, 1, 12, 13
]]
elif C_ATTRACTIVE == 3:
affinities = aff["data"][:, z_slice, :EDGE_LEN, :EDGE_LEN][[
0, 1, 2, 12, 13, 14
]]
else:
affinities = aff["data"][:, z_slice, :EDGE_LEN, :EDGE_LEN][[
0, 1, 2, 3, 12, 13, 14, 15
]]
affinities[C_ATTRACTIVE:] *= -1
affinities[C_ATTRACTIVE:] += 1.
weight_shape = affinities.shape
image_shape = weight_shape[1:]
label_image = np.zeros(image_shape, dtype=np.int)
label_image[:EDGE_LEN // 2] = 1
label_image[:, :EDGE_LEN // 2] += 2
label_image += 10
# affinities[:, :, C_ATTRACTIVE:] += affinities[:, :, :C_ATTRACTIVE]
# affinities[:, :, C_ATTRACTIVE:] /= 5
# affinities[:, :, C_ATTRACTIVE:] += 10.
# affinities[::4 ,::4, C_ATTRACTIVE:] -= 10.
# for k in range(affinities.shape[0]):
# bound = np.array([MST_calculator._check_bounds(i, k) for i in range(label_image.flatten().size)])
# plt.imshow(bound.reshape(image_shape), interpolation=None)
# plt.savefig('bound_{}.png'.format(k), dpi=500)
# plt.close()
f_affinities = affinities.flatten()
for epoch in range(1):
sorted_edges = np.argsort(f_affinities)
MST_calculator = mws.MutexWatershed(np.array(image_shape), offsets,
C_ATTRACTIVE,
np.array([1, 1, 1]))
# MST_calculator.clear_all()
MST_calculator.set_gt_labels(label_image.flatten())
start = time.time()
gradients = MST_calculator.get_gradients(sorted_edges,
f_affinities, 0)
# MST_calculator.repulsive_ucc_mst_cut(sorted_edges, 0)#[:int(EDGE_LEN*EDGE_LEN*C_ATTRACTIVE*0.5)])
end = time.time()
print("time ", end - start)
# # sorted_edges = np.argsort(f_affinities)
# MST_calculator.clear_all()
# MST_calculator.set_gt_labels(label_image.flatten())
# start = time.time()
# # MST_calculator.get_gradients(sorted_edges, f_affinities, 0)
# MST_calculator.repulsive_ucc_mst_cut(sorted_edges, 0)#[:int(EDGE_LEN*EDGE_LEN*C_ATTRACTIVE*0.5)])
# end = time.time()
# # errors_edges = MST_calculator.get_smoothmax_gradients2(f_affinities, -1)
# print("time ", end - start)
# # error_image = np.zeros(sorted_edges.shape)
# # for k, e in errors_edges.items():
# # for x in e:
# # error_image[x] = k[2]
# # error_image = error_image.reshape(affinities.shape)
# # f_affinities -= 1000 *(errors_edges)
# # f_affinities = np.clip(f_affinities,0,1)
for k in range(affinities.shape[0]):
plt.imshow(np.sign(gradients).reshape(affinities.shape)[k],
interpolation=None)
plt.savefig('new_grads_{}_{}.png'.format(epoch, k), dpi=500)
plt.close()
# for k in range(affinities.shape[0]):
# plt.imshow(errors_edges.reshape(affinities.shape)[k], interpolation=None)
# plt.savefig('error_edges_{}_{}.png'.format(epoch, k), dpi=500)
# plt.close()
# fin_labels = MST_calculator.get_flat_label_image().reshape(image_shape)
# fin_labels += ((fin_labels.max()) * (fin_labels % 10))
# out_label[z_slice] = MST_calculator.get_flat_label_image().reshape(image_shape)
# plt.imshow(regionImageToCrackEdgeImage(fin_labels.astype(np.uint32)), interpolation=None)
# plt.savefig('seg_{}.png'.format(epoch), dpi=500)
# plt.close()
# for k in range(affinities.shape[0]):
# plti.mshow(f_affinities.reshape(affinities.shape)[k], interpolation=None)
# plt.savefig('aff_{}_{}.png'.format(epoch, k), dpi=500)
# plt.close()
exit()
plt.imshow(regionImageToCrackEdgeImage(label_image.astype(np.uint32)),
interpolation=None)
plt.savefig('gt_seg.png', dpi=500)
plt.close()
# out_label[z_slice] = MST_calculator.get_flat_label_image().reshape(image_shape)
plt.imshow(regionImageToCrackEdgeImage(
MST_calculator.get_flat_c_label_image().reshape(
image_shape).astype(np.uint32)),
interpolation=None)
plt.savefig('c_seg.png', dpi=500)
plt.close()
acts = MST_calculator.get_flat_applied_uc_actions().reshape(
(-1, EDGE_LEN, EDGE_LEN))
for k in range(acts.shape[0]):
plt.imshow(acts[k], interpolation=None)
plt.savefig('a_{}.png'.format(k), dpi=500)
plt.close()
acts = MST_calculator.get_flat_applied_c_actions().reshape(
(-1, EDGE_LEN, EDGE_LEN))
for k in range(acts.shape[0]):
plt.imshow(acts[k], interpolation=None)
plt.savefig('ca_{}.png'.format(k), dpi=500)
plt.close()
for k in range(affinities.shape[0]):
plt.imshow(affinities[k], interpolation=None)
plt.savefig('aff_{}.png'.format(k), dpi=500)
plt.close()
with h5py.File("out.h5", "w") as out:
out.create_dataset("data", data=out_label, compression='gzip')
def test_cuf():
C_ATTRACTIVE = 3
EDGE_LEN = 128
Z_EDGE_LEN = 4
offsets = np.array([[-1, 0, 0], [0, -1, 0], [0, 0, -1],
[-1, -1, 0], [-1, 0, -1], [-1, -1, -1], [-1, -1, 1],
[0, -9, 0], [0, 0, -9], [0, -9, -9], [0, 9, -9],
[0, -9, 4], [0, -4, -9], [0, 4, -9], [0, 9, -4],
[-2, 0, 0], [-2, 0, -9], [-2, -9, 0], [-2, 9, -9],
[-2, -9, -9], [-3, 0, 0], [0, -27, 0], [0, 0, -27]], dtype=np.int)
with h5py.File("/export/home/swolf/local//src/data/experiments/isbi_3d/struct_05/isbi_test_3d.h5", "r") as aff:
affinities = aff["data"][:, :Z_EDGE_LEN, :EDGE_LEN, :EDGE_LEN]
affinities[C_ATTRACTIVE:] *= -1
affinities[C_ATTRACTIVE:] += 1.
weight_shape = affinities.shape
image_shape = weight_shape[1:]
label_image = np.zeros(image_shape, dtype=np.int)
label_image[:15] = 1
label_image[:, :EDGE_LEN//2] += 2
label_image[:, :, :EDGE_LEN//2] += 2
label_image += 10
# affinities[:, :, C_ATTRACTIVE:] += affinities[:, :, :C_ATTRACTIVE]
# affinities[:, :, C_ATTRACTIVE:] /= 5
# affinities[:, :, C_ATTRACTIVE:] += 10.
# affinities[::4 ,::4, C_ATTRACTIVE:] -= 10.
# for k in range(affinities.shape[0]):
# bound = np.array([MST_calculator._check_bounds(i, k) for i in range(label_image.flatten().size)])
# plt.imshow(bound.reshape(image_shape), interpolation=None)
# plt.savefig('bound_{}.png'.format(k), dpi=500)
# plt.close()
f_affinities = affinities.flatten()
sorted_edges = np.argsort(f_affinities)
MST_calculator = mws.MutexWatershed(np.array(image_shape), offsets, C_ATTRACTIVE, np.array([1,2,2]))
# MST_calculator.clear_all()
MST_calculator.set_gt_labels(label_image.flatten())
start = time.time()
gradients = MST_calculator.get_gradients(sorted_edges, f_affinities, 0)
# MST_calculator.repulsive_ucc_mst_cut(sorted_edges, 0)#[:int(EDGE_LEN*EDGE_LEN*C_ATTRACTIVE*0.5)])
end = time.time()
print("time ", end - start)
# # sorted_edges = np.argsort(f_affinities)
# MST_calculator.clear_all()
# MST_calculator.set_gt_labels(label_image.flatten())
# start = time.time()
# # MST_calculator.get_gradients(sorted_edges, f_affinities, 0)
# MST_calculator.repulsive_ucc_mst_cut(sorted_edges, 0)#[:int(EDGE_LEN*EDGE_LEN*C_ATTRACTIVE*0.5)])
# end = time.time()
# # errors_edges = MST_calculator.get_smoothmax_gradients2(f_affinities, -1)
# print("time ", end - start)
# # error_image = np.zeros(sorted_edges.shape)
# # for k, e in errors_edges.items():
# # for x in e:
# # error_image[x] = k[2]
# # error_image = error_image.reshape(affinities.shape)
# # f_affinities -= 1000 *(errors_edges)
# # f_affinities = np.clip(f_affinities,0,1)
# for k in range(affinities.shape[0]):
# plt.imshow(np.sign(gradients).reshape(affinities.shape)[k], interpolation=None)
# plt.savefig('new_grads_{}_{}.png'.format(epoch, k), dpi=500)
# plt.close()
# for k in range(affinities.shape[0]):
# plt.imshow(errors_edges.reshape(affinities.shape)[k], interpolation=None)
# plt.savefig('error_edges_{}_{}.png'.format(epoch, k), dpi=500)
# plt.close()
# fin_labels = MST_calculator.get_flat_label_image().reshape(image_shape)
# fin_labels += ((fin_labels.max()) * (fin_labels % 10))
# out_label[z_slice] = MST_calculator.get_flat_label_image().reshape(image_shape)
# plt.imshow(regionImageToCrackEdgeImage(fin_labels.astype(np.uint32)), interpolation=None)
# plt.savefig('seg_{}.png'.format(epoch), dpi=500)
# plt.close()
# for k in range(affinities.shape[0]):
# plti.mshow(f_affinities.reshape(affinities.shape)[k], interpolation=None)
# plt.savefig('aff_{}_{}.png'.format(epoch, k), dpi=500)
# plt.close()
exit()
import mutex_watershed as mws
import numpy as np
height = 4
ws = mws.MutexWatershed(np.array([height,10]), np.array([[-1,0],[0,-1], [-3,0],[0,-3]]), 2, 1)
labels = np.ones((height,10))
labels[:, :5] = 2
ws.set_gt_labels(labels)
mask = np.zeros(labels.shape, dtype=np.bool)
mask[:, 4:6] = 1
ws.set_masked_gt_image(mask)
sorted_w = np.argsort(np.random.rand(height*10*4))
s = 1
while not ws.is_finised():
ws.repulsive_ucc_mst_cut(sorted_w, s)
print(ws.get_flat_c_label_image().reshape((height, 10)), ws.get_flat_region_gt_label_image().reshape((height, 10)))
# print(ws.get_flat_applied_c_actions().reshape((4, height, 10))[[1,3]])
print("")
s += 1