def setUp(self):
gvf = 'mni_icbm152_t1_tal_nlin_sym_09c_atlas/atlas_csf.mnc'
self.fname = os.path.join(DATA_PATH, gvf)
self.img = minc.Label(self.fname)
self.tmp = create_tmp_filename(prefix='atlas_csf',
suffix='.mnc',
remove=False)
def setUp(self):
gvf = LABEL_FILE
self.fname = os.path.join(DATA_PATH, gvf)
self.img = minc.Label(self.fname)
self.tmp = create_tmp_filename(prefix='atlas_csf',
suffix='.mnc',
remove=False)
or options.load is not None) and options.image is not None:
if options.debug: print("Loading images...")
# convert to float as we go
#images= [ minc.Image(i).data.astype(np.float32) for i in options.image ]
image = minc.Image(options.image).data.astype(np.float32)
if options.debug: print("Done")
clf = None
if options.load is not None:
#TODO: load classifications
pass
else:
prior = minc.Label(options.prior)
labels = list(np.unique(prior.data))
counts = list(np.bincount(np.ravel(prior.data)))
if 0 in labels:
if options.debug: print("Label 0 will be discarded...")
labels.remove(0)
counts.pop(0) # assume it's first
num_classes = len(labels)
if options.debug:
print("Available labels:{} counts: {} ".format(
repr(labels), repr(counts)))
(images[0].shape[1] / 2.0) * 100)
images.append((c[2] - images[0].shape[2] / 2.0) /
(images[0].shape[1] / 2.0) * 100)
mask = None
if options.debug: print("Done")
man = None
training_X = None
Y = None
if options.debug: print("Creating training dataset for classifier")
if options.mask is not None:
mask = minc.Label(options.mask)
training_X = np.column_stack(
tuple(np.ravel(j[mask.data > 0]) for j in images))
else:
training_X = np.column_stack(tuple(np.ravel(j) for j in images))
if options.debug:
print("Fitting {}, dataset size:{} ...".format(
options.method, training_X.shape))
if options.method == "LLE":
man = manifold.LocallyLinearEmbedding(n_components=options.c,
n_neighbors=options.n,
eigen_solver='auto',
method='standard')
Y = man.fit_transform(training_X)
if options.coord:
# add features dependant on coordinates
c = np.mgrid[0:images[0].shape[0], 0:images[0].shape[1],
0:images[0].shape[2]]
# use with center at 0 and 1.0 at the edge, could have used preprocessing
images.append(
(c[0] - images[0].shape[0] / 2.0) / (images[0].shape[0] / 2.0))
images.append(
(c[1] - images[0].shape[1] / 2.0) / (images[0].shape[1] / 2.0))
images.append(
(c[2] - images[0].shape[2] / 2.0) / (images[0].shape[1] / 2.0))
mask = None
if options.mask is not None:
mask = minc.Label(options.mask)
if options.debug: print("Done")
clf = None
if options.load is not None:
clf = xgb.Booster(model_file=options.load)
else:
prior = minc.Label(options.prior)
labels = list(np.unique(prior.data))
counts = list(np.bincount(np.ravel(prior.data)))
if 0 in labels:
if options.debug: print("Label 0 will be discarded...")
labels.remove(0)
train=json.load(f)
else:
with open(options.train_csv,'rb') as f:
train=list(csv.reader(f))
training_images=[]
training_output=[]
training_err=[]
#go over training samples
clf=None
#scaler=preprocessing.StandardScaler().fit(X)
for (i,inp) in enumerate(train):
mask =minc.Label( inp[-3] ).data
ground=minc.Label( inp[-2] ).data
auto =minc.Label( inp[-1] ).data
# normalize input features to zero mean and unit std
if options.normalize:
images=[ preprocessing.scale(minc.Image(k).data) for k in inp[0:-3] ]
else:
images=[ minc.Image(k).data for k in inp[0:-3] ]
# store training data
training_images.append( prepare_features( options, images, None, auto, mask ) )
# perform direct learning right now
training_output.append( ground[mask>0] )
if __name__ == "__main__":
history=minc.format_history(sys.argv)
options = parse_options()
# load prior and input image
if options.debug: print("Loading images...")
# convert to float as we go
ref_image=minc.Image(options.reference).data.astype(np.float32)
image=minc.Image(options.image).data.astype(np.float32)
if options.debug: print("Done")
mm=(image>0)
if options.mask is not None:
mask = minc.Label(options.mask)
mm = np.logical_and(image>0 , mask.data>0 )
rmm=(ref_image>0)
if options.refmask is not None:
refmask = minc.Label(options.refmask)
rmm = np.logical_and(ref_image>0 , refmask.data>0 )
#print ref_image[rmm]
rmin=np.amin(ref_image[rmm])
rmax=np.amax(ref_image[rmm])
print("Ref Range {} - {}".format(rmin,rmax))
imin=np.amin(image[mm])
imax=np.amax(image[mm])
print("Image Range {} - {}".format(imin,imax))
if n_i == -1:
n_i = len(i) - 1
elif n_i != (len(i) - 1):
raise "Inconsistent number of images:{}".format(repr(i))
if n_i == -1:
raise "No input images!"
if options.debug:
print("Loading {} images ...".format(n_i * len(train_library)))
images = [[
minc.Image(os.path.join(prefix, j[i])).data for i in range(n_i)
] for j in train_library]
segs = [
minc.Label(os.path.join(prefix, j[n_i])).data
for j in train_library
]
priors = []
# TODO: check shape of all images for consistency
_shape = images[0][0].shape
if options.coord:
# add features dependant on coordinates
c = np.mgrid[0:_shape[0], 0:_shape[1], 0:_shape[2]]
# use with center at 0 and 1.0 at the edge, could have used preprocessing
priors.append((c[0] - _shape[0] / 2.0) / (_shape[0] / 2.0))
priors.append((c[1] - _shape[1] / 2.0) / (_shape[1] / 2.0))
priors.append((c[2] - _shape[2] / 2.0) / (_shape[1] / 2.0))
