def voi(self, prob, label):
prob = np.int64(1 - prob)
f_image, n_labels = mahotas.label(prob)
f_image = np.int64(f_image.flatten())
f_label = np.int64(label.flatten())
return n_labels, partition_comparison.variation_of_information(
f_image, f_label)
def create_vi_plot(initial_segmentation,
new_segmentation,
vi_s,
filepath=None):
# create plot
new_segmentation_target = Util.propagate_max_overlap(
new_segmentation, initial_segmentation)
before_vi = partition_comparison.variation_of_information(
new_segmentation.ravel(), initial_segmentation.ravel())
target_vi = partition_comparison.variation_of_information(
new_segmentation_target.ravel(), initial_segmentation.ravel())
bins = np.arange(0, len(vi_s))
before_vi = [before_vi] * len(vi_s)
target_vi = [target_vi] * len(vi_s)
fig, ax = plt.subplots()
ax.plot(bins, target_vi, 'k--', label='Target VI')
ax.plot(bins, vi_s_, 'k', label='Variation of Information')
ax.plot(bins, before_vi, 'k:', label='VI before')
# ax.set_yscale('log')
# Now add the legend with some customizations.
legend = ax.legend(loc='upper center', shadow=True)
ax.set_ylim([-0.5, 1.5])
# The frame is matplotlib.patches.Rectangle instance surrounding the legend.
frame = legend.get_frame()
frame.set_facecolor('0.90')
# Set the fontsize
for label in legend.get_texts():
label.set_fontsize('large')
for label in legend.get_lines():
label.set_linewidth(1.5) # the legend line width
if filepath:
plt.savefig(filepath)
def apply(cls, label_array):
'''
Apply the metric to a label_array with shape Y,X,Z.
'''
vi_sum = 0.
done = 0.
print 'Calulating for', label_array.shape[0], 'slices.'
for z in range(label_array.shape[0] - 1):
vi_sum += partition_comparison.variation_of_information(
label_array[z, :, :].astype(np.uint64).ravel(),
label_array[z + 1, :, :].astype(np.uint64).ravel())
done += 1
percentage = int(((done) / (label_array.shape[0] - 1)) * 100)
if (percentage % 10 == 0):
print percentage, '% done'
vi_sum /= label_array.shape[0]
return vi_sum
for thresh in thresh_range:
below_thresh = smooth_output < np.uint16(max_smooth * thresh)
#below_thresh = mahotas.morph.close(below_thresh.astype(np.bool), disc)
#below_thresh = mahotas.morph.open(below_thresh.astype(np.bool), disc)
seeds,nseeds = mahotas.label(below_thresh)
if nseeds == 0:
continue
ws = np.uint32(mahotas.cwatershed(smooth_output, seeds))
voi_score = partition_comparison.variation_of_information(target_segs.ravel(), ws.ravel())
thresh_voi_results.append(voi_score)
print 's={0:0.2f}, t={1:0.2f}, voi_score={2:0.4f}.'.format(smooth_sigma, thresh, voi_score)
dx, dy = np.gradient(ws)
result = np.logical_or(dx!=0, dy!=0)
figsize(20,20)
imshow(result, cmap=cm.gray)
plt.show()
if voi_score < best_score:
best_score = voi_score
best_sigma = smooth_sigma
# print out.shape
# print np.where(segs.ravel() != 0)
# print np.where(gt.ravel() != 0)
# print np.where(out.ravel() != 0)
# crop the data to only image data
x = 210
y = 60
z = 50 # we dont need since we have already cut the z
dim_x = dim_y = 400
segs = segs[:, y:y + dim_y, x:x + dim_x]
gt = gt[:, y:y + dim_y, x:x + dim_x]
#out = out[:,y:y+dim_y,x:x+dim_x]
print 'VI groundtruth vs. groundtruth:', pc.variation_of_information(
gt.astype(np.uint32).ravel(),
gt.astype(np.uint32).ravel())
print 'VI groundtruth vs. pipeline:', pc.variation_of_information(
gt.astype(np.uint32).ravel(), segs.ravel())
print 'VI groundtruth vs. proofread:', pc.variation_of_information(
gt.astype(np.uint32).ravel(), out.ravel())
print 'Adjusted RandIndex groundtruth vs. groundtruth:', metrics.adjusted_rand_score(
gt.astype(np.uint32).ravel(),
gt.astype(np.uint32).ravel())
print 'Adjusted RandIndex groundtruth vs. pipeline:', metrics.adjusted_rand_score(
gt.astype(np.uint32).ravel(), segs.ravel())
print 'Adjusted RandIndex groundtruth vs. proofread:', metrics.adjusted_rand_score(
gt.astype(np.uint32).ravel(), out.ravel())
import numpy as np
from sklearn import metrics
segs = tif.imread('/home/d/TMP/DOJO/mojo_400_5_orig/cut-train-segs.tif')
gt = tif.imread('/home/d/TMP/DOJO/mojo_400_5_orig/cut-train-labels.tif')
exp = tif.imread('/home/d/dong.tif')
exp[exp==0] = 2000
# crop the data to only image data
x = 210
y = 60
z = 50 # we dont need since we have already cut the z
dim_x = dim_y = 400
segs = segs[:,y:y+dim_y,x:x+dim_x]
gt = gt[:,y:y+dim_y,x:x+dim_x]
#out = out[:,y:y+dim_y,x:x+dim_x]
print 'VI groundtruth vs. groundtruth:', pc.variation_of_information(gt.astype(np.uint32).ravel(), gt.astype(np.uint32).ravel())
print 'VI groundtruth vs. pipeline:',pc.variation_of_information(gt.astype(np.uint32).ravel(), segs.ravel())
# print 'VI groundtruth vs. proofread:', pc.variation_of_information(gt.astype(np.uint32).ravel(), out.ravel())
print 'VI groundtruth vs. expert:', pc.variation_of_information(gt.astype(np.uint32).ravel(), exp.astype(np.uint32).ravel())
print 'Adjusted RandIndex groundtruth vs. groundtruth:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), gt.astype(np.uint32).ravel())
print 'Adjusted RandIndex groundtruth vs. pipeline:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), segs.ravel())
# print 'Adjusted RandIndex groundtruth vs. proofread:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), out.ravel())
print 'Adjusted RandIndex groundtruth vs. expert:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), exp.astype(np.uint32).ravel())
def vi(array1, array2):
'''
'''
return partition_comparison.variation_of_information(
array1.ravel(), array2.ravel())
# out = np.dstack([out, im]) # else: # out = im # out_is_there = True # print segs.shape # print gt.shape # print out.shape # print np.where(segs.ravel() != 0) # print np.where(gt.ravel() != 0) # print np.where(out.ravel() != 0) # crop the data to only image data x = 210 y = 60 z = 50 # we dont need since we have already cut the z dim_x = dim_y = 400 segs = segs[:,y:y+dim_y,x:x+dim_x] gt = gt[:,y:y+dim_y,x:x+dim_x] #out = out[:,y:y+dim_y,x:x+dim_x] print 'VI groundtruth vs. groundtruth:', pc.variation_of_information(gt.astype(np.uint32).ravel(), gt.astype(np.uint32).ravel()) print 'VI groundtruth vs. pipeline:',pc.variation_of_information(gt.astype(np.uint32).ravel(), segs.ravel()) print 'VI groundtruth vs. proofread:', pc.variation_of_information(gt.astype(np.uint32).ravel(), out.ravel()) print 'Adjusted RandIndex groundtruth vs. groundtruth:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), gt.astype(np.uint32).ravel()) print 'Adjusted RandIndex groundtruth vs. pipeline:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), segs.ravel()) print 'Adjusted RandIndex groundtruth vs. proofread:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), out.ravel())
def run(slice):
DATA_PATH = '/Volumes/DATA1/EMQM_DATA/ac3x75/'
GOLD_PATH = os.path.join(DATA_PATH,'gold/')
RHOANA_PATH = os.path.join(DATA_PATH,'rhoana/')
IMAGE_PATH = os.path.join(DATA_PATH,'input/')
PROB_PATH = os.path.join(DATA_PATH,'prob/')
PATCH_PATH = os.path.join(DATA_PATH,'test_rhoana/')
OUTPUT_PATH = os.path.join(DATA_PATH,'fix_rhoana/')
gold = _metrics.Util.read(GOLD_PATH+'*.tif')
rhoana = _metrics.Util.read(RHOANA_PATH+'*.tif')
images = _metrics.Util.read(IMAGE_PATH+'*.tif')
probs = _metrics.Util.read(PROB_PATH+'*.tif')
SLICE = slice
print 'Running on slice', SLICE
image = images[SLICE]
prob = probs[SLICE]
seg = _metrics.Util.normalize_labels(rhoana[SLICE])[0]
# fill and normalize gold
gold_zeros = _metrics.Util.threshold(gold[SLICE], 0)
gold_filled = _metrics.Util.fill(gold[SLICE], gold_zeros.astype(np.bool))
gold_filled_relabeled = skimage.measure.label(gold_filled).astype(np.uint64)
gold_normalized = _metrics.Util.normalize_labels(gold_filled_relabeled)[0]
# color both segmentations
cm = _metrics.Util.load_colormap('/Volumes/DATA1/ac3x75/mojo/ids/colorMap.hdf5')
colored_rhoana_normalized = cm[seg % len(cm)]
colored_gold_normalized = cm[gold_normalized % len(cm)]
print 'Find all patches'
t0 = time.time()
p = loop(image, prob, seg)
print time.time()-t0, 'seconds for', len(p), 'patches'
print 'Setting up CNN'
val_fn = setup_n()
print 'Creating Matrix'
m_new = create_matrix(val_fn, seg, p)
out = seg
sureness = 0.
surenesses = []
vi_s = []
images = []
before_VI = partition_comparison.variation_of_information(gold_normalized.ravel(), seg.ravel())
max_counter = 0
while m_new.max() >= sureness:
max_counter += 1
print 'Iteration', max_counter
print 'Merging for sureness s=', m_new.max()
surenesses.append(m_new.max())
m_new, out = merge(val_fn, image, out, prob, m_new)
vi = partition_comparison.variation_of_information(out.ravel(), gold_normalized.ravel())
vi_s.append(vi)
images.append(np.array(out))
print 'New VI', vi
print '-'*80
before_VI = partition_comparison.variation_of_information(gold_normalized.ravel(), seg.ravel())
index = vi_s.index(min(vi_s))
print 'Before VI', before_VI
print 'Lowest VI', vi_s[index]
print 'Iterations', index
print 'Sureness', surenesses[index]
# create plot
new_rhoana = propagate_max_overlap(seg, gold_normalized)
target_vi = partition_comparison.variation_of_information(new_rhoana.ravel(), gold_normalized.ravel())
bins = np.arange(0, len(vi_s))#np.arange(len(vi_s), 0, -1)
# bins /= 100
vi_s_ = vi_s #[x for (y,x) in sorted(zip(surenesses, vi_s))]
surenesses_ = bins #surenesses#[y for (y,x) in sorted(zip(surenesses, vi_s))]
original_vi = [before_VI]*len(vi_s)
target_vi = [target_vi]*len(vi_s)
fig, ax = plt.subplots()
ax.plot(surenesses_, target_vi, 'k--', label='Target VI')
ax.plot(surenesses_, vi_s_, 'k', label='Variation of Information')
ax.plot(surenesses_, original_vi, 'k:', label='Rhoana VI before')
# ax.set_yscale('log')
# Now add the legend with some customizations.
legend = ax.legend(loc='upper center', shadow=True)
ax.set_ylim([0.6,1.5])
# The frame is matplotlib.patches.Rectangle instance surrounding the legend.
frame = legend.get_frame()
frame.set_facecolor('0.90')
# Set the fontsize
for label in legend.get_texts():
label.set_fontsize('large')
for label in legend.get_lines():
label.set_linewidth(1.5) # the legend line width
plt.savefig(OUTPUT_PATH+os.sep+'graph_'+str(slice)+'.png')
mh.imsave(OUTPUT_PATH+os.sep+'best_'+str(slice)+'.tif', images[index])
mh.imsave(OUTPUT_PATH+os.sep+'last_'+str(slice)+'.tif', images[-1])
print 'All done.'
def partition_VI(putative, gold):
mask = gold != 0
VI_res = partition_comparison.variation_of_information(gold[mask], putative[mask].astype(gold.dtype))
return VI_res
exp = tif.imread('/home/d/dong.tif')
exp[exp == 0] = 2000
# crop the data to only image data
x = 210
y = 60
z = 50 # we dont need since we have already cut the z
dim_x = dim_y = 400
segs = segs[:, y:y + dim_y, x:x + dim_x]
gt = gt[:, y:y + dim_y, x:x + dim_x]
#out = out[:,y:y+dim_y,x:x+dim_x]
print 'VI groundtruth vs. groundtruth:', pc.variation_of_information(
gt.astype(np.uint32).ravel(),
gt.astype(np.uint32).ravel())
print 'VI groundtruth vs. pipeline:', pc.variation_of_information(
gt.astype(np.uint32).ravel(), segs.ravel())
# print 'VI groundtruth vs. proofread:', pc.variation_of_information(gt.astype(np.uint32).ravel(), out.ravel())
print 'VI groundtruth vs. expert:', pc.variation_of_information(
gt.astype(np.uint32).ravel(),
exp.astype(np.uint32).ravel())
print 'Adjusted RandIndex groundtruth vs. groundtruth:', metrics.adjusted_rand_score(
gt.astype(np.uint32).ravel(),
gt.astype(np.uint32).ravel())
print 'Adjusted RandIndex groundtruth vs. pipeline:', metrics.adjusted_rand_score(
gt.astype(np.uint32).ravel(), segs.ravel())
# print 'Adjusted RandIndex groundtruth vs. proofread:', metrics.adjusted_rand_score(gt.astype(np.uint32).ravel(), out.ravel())
print 'Adjusted RandIndex groundtruth vs. expert:', metrics.adjusted_rand_score(