def compute_scores(iterations):
labels_key = ArrayKey('LABELS')
gt_affs_key = ArrayKey('GT_AFFINITIES')
pred_affinities_key = ArrayKey('PREDICTED_AFFS')
sample_z_key = ArrayKey("SAMPLE_Z")
voxel_size = Coordinate((1, 1, 1))
input_shape = Coordinate((132, 132, 132)) * voxel_size
output_shape = Coordinate((44, 44, 44)) * voxel_size
sample_shape = Coordinate((1, 1, 6)) * voxel_size
print ("input_size: ", input_shape)
print ("output_size: ", output_shape)
request = BatchRequest()
# request.add(labels_key, output_shape)
request.add(gt_affs_key, output_shape, )
request.add(pred_affinities_key, input_shape)
# request.add(sample_z_key, sample_shape)
dataset_names = {
gt_affs_key: 'volumes/gt_affs',
pred_affinities_key: 'volumes/pred_affs'
}
array_specs = {
gt_affs_key: ArraySpec(interpolatable=True),
pred_affinities_key: ArraySpec(interpolatable=True)
}
pipeline = tuple(
Hdf5Source(
os.path.join(data_dir, sample + '.hdf'),
datasets = dataset_names,
array_specs = array_specs
) +
Pad(gt_affs_key, None) +
Pad(pred_affinities_key, None)
# Pad(merged_labels_key[i], None) for i in range(num_merges) # don't know why this doesn't work
for sample in samples
)
pipeline += SequentialProvider()
pipeline += Snapshot(
dataset_names={
gt_affs_key: 'volumes/gt_affs',
# pred_affinities_key: 'volumes/pred_affs',
# sample_z_key: 'volumes/sample_z',
},
output_filename='test_scores.hdf',
every=1,
dataset_dtypes={
gt_affs_key: np.float32,
# pred_affinities_key: np.float32,
# sample_z_key: np.float32,
})
print("Calculating Scores...")
with build(pipeline) as p:
aris = []
for i in range(iterations):
if i % 10 == 0:
print("iteration: ", i)
req = p.request_batch(request)
gt_affs = np.array(req[gt_affs_key].data)
pred_affs = threshold(__crop_center(np.array(req[pred_affinities_key].data), (44,44,44)))
aris.append(adjusted_rand_score(gt_affs.flatten(), pred_affs.flatten()))
print ("ari: ", aris[i])
# print ("aris: ", aris)
aris = np.array(aris)
maximum = np.max(aris)
minimum = np.min(aris)
std = np.std(aris)
mean = np.mean(aris)
upper_std = mean + std
lower_std = mean - std
print ("maximum: ", maximum)
print ("minimum: ", minimum)
print ("mean: ", mean)
print ("std: ", std)
print ("upper_std: ", upper_std)
print ("lower_std: ", lower_std)
with open("ari/" + setup_name + ".txt", "wb") as fp: #Pickling
pickle.dump(aris, fp)
print("Score calculation finished")
def predict(checkpoint, iterations):
print("checkpoint: ", checkpoint)
labels_key = ArrayKey('LABELS')
gt_affs_key = ArrayKey('GT_AFFINITIES')
raw_affs_key = ArrayKey('RAW_AFFINITIES')
raw_joined_affs_key = ArrayKey('RAW_JOINED_AFFINITIES')
raw_key = ArrayKey('RAW')
pred_affinities_key = ArrayKey('PREDICTED_AFFS')
sample_z_key = ArrayKey("SAMPLE_Z")
# broadcast_key = ArrayKey("BROADCAST")
# pred_logits_key = ArrayKey("PRED_LOGITS")
# sample_out_key = ArrayKey("SAMPLE_OUT")
# debug_key = ArrayKey("DEBUG")
voxel_size = Coordinate((1, 1, 1))
input_shape = Coordinate(config['input_shape']) * voxel_size
output_shape = Coordinate(config['output_shape']) * voxel_size
sample_shape = Coordinate((1, 1, 6)) * voxel_size
# debug_shape = Coordinate((1, 1, 5)) * voxel_size
print("input_size: ", input_shape)
print("output_size: ", output_shape)
request = BatchRequest()
request.add(labels_key, output_shape)
request.add(gt_affs_key, input_shape)
request.add(raw_affs_key, input_shape)
request.add(raw_joined_affs_key, input_shape)
request.add(raw_key, input_shape)
request.add(pred_affinities_key, output_shape)
# request.add(broadcast_key, output_shape)
request.add(sample_z_key, sample_shape)
# request.add(pred_logits_key, output_shape)
# request.add(sample_out_key, sample_shape)
# request.add(debug_key, debug_shape)
dataset_names = {
labels_key: 'volumes/labels',
}
array_specs = {labels_key: ArraySpec(interpolatable=False)}
pipeline = tuple(
Hdf5Source(os.path.join(data_dir, sample + '.hdf'),
datasets=dataset_names,
array_specs=array_specs) + Pad(labels_key, None)
# Pad(merged_labels_key[i], None) for i in range(num_merges) # don't know why this doesn't work
for sample in samples)
pipeline += (
# Pad(raw_key, size=None) +
# Crop(raw_key, read_roi) +
#Normalize(raw_key) +
SequentialProvider() +
AddAffinities(affinity_neighborhood=neighborhood,
labels=labels_key,
affinities=raw_affs_key) +
AddJoinedAffinities(input_affinities=raw_affs_key,
joined_affinities=raw_joined_affs_key) +
AddRealism(joined_affinities=raw_joined_affs_key,
raw=raw_key,
sp=0.25,
sigma=1,
contrast=0.7) +
GrowBoundary(labels_key, steps=1, only_xy=True) +
AddAffinities(affinity_neighborhood=neighborhood,
labels=labels_key,
affinities=gt_affs_key) +
PreCache(cache_size=32, num_workers=8) +
IntensityScaleShift(raw_key, 2, -1) + Predict(
checkpoint=os.path.join(setup_dir,
'train_net_checkpoint_%d' % checkpoint),
inputs={config['raw']: raw_key},
outputs={
config['pred_affs']: pred_affinities_key,
config['sample_z']: sample_z_key,
# config['broadcast']: broadcast_key,
# config['pred_logits']: pred_logits_key,
# config['sample_out']: sample_out_key,
# config['debug']: debug_key
},
graph=os.path.join(setup_dir, 'predict_net.meta')) +
IntensityScaleShift(array=raw_key, scale=0.5, shift=0.5) + Snapshot(
dataset_names={
# labels_key: 'volumes/labels',
gt_affs_key: 'volumes/gt_affs',
# raw_key: 'volumes/raw',
pred_affinities_key: 'volumes/pred_affs',
# broadcast_key: 'volumes/broadcast',
# sample_z_key: 'volumes/sample_z',
# pred_logits_key: 'volumes/pred_logits',
# sample_out_key: 'volumes/sample_out'
},
output_filename='prob_unet/' + setup_name +
'/prediction_A_{id}.hdf',
every=1,
dataset_dtypes={
# labels_key: np.uint16,
gt_affs_key: np.float32,
pred_affinities_key: np.float32,
# broadcast_key: np.float32,
# sample_z_key: np.float32,
# pred_logits_key: np.float32,
# sample_out_key: np.float32
})
# PrintProfilingStats(every=20)
)
print("Starting prediction...")
with build(pipeline) as p:
for i in range(iterations):
req = p.request_batch(request)
# sample_z = req[sample_z_key].data
# broadcast_sample = req[broadcast_key].data
# sample_out = req[sample_out_key].data
# debug = req[debug_key].data
# print("debug", debug)
# print("sample_z: ", sample_z)
# print("sample_out:", sample_out)
# print("Z - 0")
# print(np.unique(broadcast_sample[0, 0, :, :, :]))
# print("Z - 1")
# print(np.unique(broadcast_sample[0, 1, :, :, :]))
# print("Z - 2")
# print(np.unique(broadcast_sample[0, 2, :, :, :]))
# print("Z - 3")
# print(np.unique(broadcast_sample[0, 3, :, :, :]))
# print("Z - 4")
# print(np.unique(broadcast_sample[0, 4, :, :, :]))
# print("Z - 5")
# print(np.unique(broadcast_sample[0, 5, :, :, :]))
print("Prediction finished")
def compute_scores(d, iterations):
samples = [
"prediction_%08i_A" % d,
"prediction_%08i_B" % d,
"prediction_%08i_C" % d,
"prediction_%08i_D" % d
]
labels_key = ArrayKey('LABELS')
gt_affs_key = ArrayKey('GT_AFFINITIES')
pred_affinities_key = ArrayKey('PREDICTED_AFFS')
sample_z_key = ArrayKey("SAMPLE_Z")
voxel_size = Coordinate((1, 1, 1))
input_shape = Coordinate((132, 132, 132)) * voxel_size
output_shape = Coordinate((44, 44, 44)) * voxel_size
sample_shape = Coordinate((1, 1, 6)) * voxel_size
request = BatchRequest()
# request.add(labels_key, output_shape)
request.add(
gt_affs_key,
output_shape,
)
request.add(pred_affinities_key, input_shape)
# request.add(sample_z_key, sample_shape)
dataset_names = {
gt_affs_key: 'volumes/gt_affs',
pred_affinities_key: 'volumes/pred_affs'
}
array_specs = {
gt_affs_key: ArraySpec(interpolatable=True),
pred_affinities_key: ArraySpec(interpolatable=True)
}
pipeline = tuple(
Hdf5Source(os.path.join(data_dir, sample + '.hdf'),
datasets=dataset_names,
array_specs=array_specs) + Pad(gt_affs_key, None) +
Pad(pred_affinities_key, None)
# Pad(merged_labels_key[i], None) for i in range(num_merges) # don't know why this doesn't work
for sample in samples)
pipeline += SequentialProvider()
# pipeline += Snapshot(
# dataset_names={
# gt_affs_key: 'volumes/gt_affs',
# pred_affinities_key: 'volumes/pred_affs',
# # sample_z_key: 'volumes/sample_z',
# },
# output_filename='test_scores.hdf',
# every=1,
# dataset_dtypes={
# gt_affs_key: np.float32,
# pred_affinities_key: np.float32,
# sample_z_key: np.float32,
# })
# print("Calculating Scores...")
with build(pipeline) as p:
aris = []
pred_affs = []
for i in range(iterations):
req = p.request_batch(request)
if i == 0:
gt_affs = np.array(req[gt_affs_key].data)
pred_affs.append(
threshold(
__crop_center(np.array(req[pred_affinities_key].data),
(44, 44, 44))))
# print(np.sum(gt_affs))
# print(np.sum(pred_affs[0]))
# print(np.sum(pred_affs[1]))
ari_A = adjusted_rand_score(gt_affs.flatten(), pred_affs[0].flatten())
ari_B = adjusted_rand_score(gt_affs.flatten(), pred_affs[1].flatten())
ari_C = adjusted_rand_score(gt_affs.flatten(), pred_affs[2].flatten())
ari_D = adjusted_rand_score(gt_affs.flatten(), pred_affs[3].flatten())
ari_YS = (ari_A + ari_B + ari_C + ari_D) / 4
d_YS = 1 - ari_YS
# print("gt_ari_avg: ", ari_YS)
ari_AB = adjusted_rand_score(pred_affs[0].flatten(),
pred_affs[1].flatten())
ari_AC = adjusted_rand_score(pred_affs[0].flatten(),
pred_affs[2].flatten())
ari_AD = adjusted_rand_score(pred_affs[0].flatten(),
pred_affs[3].flatten())
ari_BC = adjusted_rand_score(pred_affs[1].flatten(),
pred_affs[2].flatten())
ari_BD = adjusted_rand_score(pred_affs[1].flatten(),
pred_affs[3].flatten())
ari_CD = adjusted_rand_score(pred_affs[2].flatten(),
pred_affs[3].flatten())
ari_SS = (ari_AB + ari_AC + ari_AD + ari_BC + ari_BD + ari_CD) / 6
d_SS = 1 - ari_SS
GED = 2 * d_YS - d_SS
# print("pred_ari_avg: ", ari_SS)
# print("GED: ", 2*d_YS - d_SS)
return (ari_YS, d_SS, GED)