def __call__(self, model, scanner):
dataset = scanner.dataset
# Test-time augmentation
if self.test_aug:
# For variance computation
if self.variance:
aug_out = dict()
for k, v in scanner.outputs.data.items():
aug_out[k] = list()
else:
aug_out = None
count = 0.0
for aug in self.test_aug:
# dec2bin
rule = np.array([int(x) for x in bin(aug)[2:].zfill(4)])
print("Test-time augmentation {}".format(rule))
# Augment dataset.
aug_dset = Dataset(spec=self.in_spec)
for k, v in dataset.data.items():
aug_dset.add_data(k, fwd_utils.flip(v._data, rule=rule))
# Forward scan
aug_scanner = self.make_forward_scanner(aug_dset)
outputs = self.forward(model, aug_scanner)
# Accumulate.
for k, v in scanner.outputs.data.items():
print("Accumulate to {}...".format(k))
output = outputs.get_data(k)
# Revert output.
dst = (1, 1, 1) if k == 'affinity' else None
reverted = fwd_utils.revert_flip(output,
rule=rule,
dst=dst)
v._data += reverted
# For variance computation
if self.variance:
aug_out[k].append(reverted)
count += 1
# Normalize.
for k, v in scanner.outputs.data.items():
print("Normalize {}...".format(k))
if self.precomputed:
v._data[...] /= count
else:
v._norm._data[...] = count
return (scanner.outputs, aug_out)
return (self.forward(model, scanner), None)
def make_forward_scanner(opt, data_name=None):
# Cloud-volume
if opt.gs_input:
try:
from deepem.test import cv_utils
img = cv_utils.cutout(opt, opt.gs_input, dtype='uint8')
# Optional input histogram normalization
if opt.gs_input_norm:
assert len(opt.gs_input_norm) == 2
low, high = opt.gs_input_norm
img = normalize_per_slice(img, lowerfract=low, upperfract=high)
# [0, 255] -> [0.0, 1.0]
img = (img/255.).astype('float32')
# Optional input mask
if opt.gs_input_mask:
try:
msk = cv_utils.cutout(opt, opt.gs_input_mask, dtype='uint8')
img[msk > 0] = 0
except:
raise
except ImportError:
raise
else:
assert data_name is not None
print(data_name)
# Read an EM image.
if opt.dummy:
img = np.random.rand(*opt.dummy_inputsz[-3:]).astype('float32')
else:
fpath = os.path.join(opt.data_dir, data_name, opt.input_name)
img = emio.imread(fpath)
img = (img/255.).astype('float32')
# Border mirroring
if opt.mirror:
pad_width = [(x//2,x//2) for x in opt.mirror]
img = np.pad(img, pad_width, 'reflect')
# ForwardScanner
dataset = Dataset(spec=opt.in_spec)
dataset.add_data('input', img)
return ForwardScanner(dataset, opt.scan_spec, **opt.scan_params)
def build_dataset(self, tag, data):
img = data['img']
mit = data['mit']
loc = data['loc']
msk = self.get_mask(data)
# Create Dataset.
dset = Dataset(tag=tag)
dset.add_data(key='input', data=img)
dset.add_data(key='mitochondria', data=mit)
dset.add_mask(key='mitochondria_mask', data=msk)
return dset
def build_dataset(self, tag, data):
img = data['img']
seg = data['seg']
loc = data['loc']
msk = self.get_mask(data)
# Create Dataset.
dset = Dataset(tag=tag)
dset.add_data(key='input', data=img)
dset.add_data(key='affinity', data=seg)
dset.add_mask(key='affinity_mask', data=msk, loc=loc)
return dset
def build_dataset(self, tag, data):
img = data['img']
psd = data['psd']
psd_msk = data['psd_msk']
loc = data['loc']
msk = self.get_mask(data)
# Create Dataset.
dset = Dataset(tag=tag)
dset.add_data(key='input', data=img)
dset.add_data(key='synapse', data=psd)
dset.add_mask(key='synapse_mask', data=psd_msk, loc=loc)
return dset
def build_dataset(self, datadir, vol):
img = read_h5(os.path.join(datadir, vol + "_img.h5"))
soma = read_h5(os.path.join(datadir, vol + "_lbl.h5")).astype("float32")
#Preprocessing
img = (img / 255.).astype("float32")
soma[soma != 0] = 1
# Create Dataset.
dset = Dataset()
dset.add_data(key='input', data=img)
dset.add_data(key='soma_label', data=soma)
return dset
def build_dataset(self, datadir, vol):
img = read_h5(os.path.join(datadir, vol + "_img.h5"))
clf = read_h5(os.path.join(datadir, vol + "_syn.h5")).astype("float32")
#Preprocessing
img = (img / 255.).astype("float32")
clf[clf != 0] = 1
# Create Dataset.
dset = Dataset()
dset.add_data(key='input', data=img)
dset.add_data(key='cleft_label', data=clf)
return dset
def build_dataset(self, datadir, vol):
# Reading either hdf5 or tif training data; raw image has to be consistent with label image
if os.path.isfile(os.path.join(datadir, vol + "_img.h5")):
img = read_img(os.path.join(datadir, vol + "_img.h5"))
soma = read_img(os.path.join(datadir, vol + "_lbl.h5")).astype("float32")
elif os.path.isfile(os.path.join(datadir, vol + "_img.tif")):
img = read_img(os.path.join(datadir, vol + "_img.tif"))
soma = read_img(os.path.join(datadir, vol + "_lbl.tif")).astype("float32")
#Preprocessing
img = (img / 255.).astype("float32")
soma[soma != 0] = 1
# Create Dataset.
dset = Dataset()
dset.add_data(key='input', data=img)
dset.add_data(key='soma_label', data=soma)
return dset
def build_dataset(self, tag, data):
img = data['img']
seg = data['seg']
loc = data['loc']
msk = self.get_mask(data)
glia = data['glia']
gmsk = data['gmsk'] if 'gmsk' in data else msk
# Create Dataset.
dset = Dataset(tag=tag)
dset.add_data(key='input', data=img)
dset.add_data(key='affinity', data=seg)
dset.add_mask(key='affinity_mask', data=msk, loc=loc)
dset.add_data(key='glia', data=glia)
dset.add_mask(key='glia_mask', data=gmsk)
return dset
def build_dataset(self, tag, data):
img = data['img']
seg = data['seg']
mye = data['mye']
blv = data['blv']
loc = data['loc']
msk = self.get_mask(data)
# Create Dataset.
dset = Dataset(tag=tag)
dset.add_data(key='input', data=img)
dset.add_data(key='affinity', data=seg)
dset.add_mask(key='affinity_mask', data=msk, loc=loc)
dset.add_data(key='myelin', data=mye)
dset.add_mask(key='myelin_mask', data=msk)
dset.add_data(key='blood_vessel', data=blv)
dset.add_mask(key='blood_vessel_mask', data=msk)
return dset
def build_dataset(self, tag, data):
img = data['img']
seg = data['seg']
psd = data['psd']
psd_msk = data['psd_msk']
mye = data['mye']
loc = data['loc']
msk = self.get_mask(data)
# Create Dataset.
dset = Dataset(tag=tag)
dset.add_data(key='input', data=img)
dset.add_data(key='affinity', data=seg)
dset.add_mask(key='affinity_mask', data=msk, loc=loc)
dset.add_data(key='synapse', data=psd)
dset.add_mask(key='synapse_mask', data=psd_msk)
dset.add_data(key='myelin', data=mye)
dset.add_mask(key='myelin_mask', data=msk)
return dset
