def log_image(self, print_calling_method=False):
"""
Log basic stats for this ImagePair.
"""
self.logger(
"%s\n"
"--- shape: %s\n"
"--- real shape: %s\n"
"--- pixdim: %s" %
(self.id, self.shape, np.round(get_real_image_size(self),
3), np.round(get_pix_dim(self), 3)),
print_calling_method=print_calling_method)
def audit(self):
shapes = []
channels = []
real_sizes = []
pixdims = []
memory = []
for im_path in self.nii_paths:
# Load the nii file without loading image data
im = nib.load(im_path)
# Get image voxel shape
shape = im.shape
shapes.append(shape[:3])
try:
c = shape[3]
except IndexError:
c = 1
channels.append(c)
# Get image real shape
real_sizes.append(get_real_image_size(im))
# Get pixel dims
pixdims.append(get_pix_dim(im))
# Calculate memory in bytes to store image
memory.append(im.get_data_dtype().itemsize * np.prod(shape))
n_classes = None
if self.nii_lab_paths is not None:
# Labels exists, thus we need the n_classes attribute
if self.hparms is not None:
# Attempt to get it from a potentially specified hparams obj
n_classes = self.hparms.get_from_anywhere("n_classes")
if n_classes is None:
# If still none, infer it
n_classes = _audit_classes(self.nii_lab_paths, self.logger)
info = {
"shapes": shapes,
"real_sizes": real_sizes,
"pixdims": pixdims,
"memory_bytes": memory,
"n_channels": channels,
"n_classes": n_classes,
}
return info
def audit(self):
shapes = []
channels = []
real_sizes = []
pixdims = []
memory = []
for im_path in self.nii_paths:
# Load the nii file without loading image data
im = nib.load(im_path)
# Get image voxel shape
shape = im.shape
shapes.append(shape[:3])
try:
c = shape[3]
except IndexError:
c = 1
channels.append(c)
# Get image real shape
real_sizes.append(get_real_image_size(im))
# Get pixel dims
pixdims.append(get_pix_dim(im))
# Calculate memory in bytes to store image
memory.append(im.get_data_dtype().itemsize * np.prod(shape))
if self.nii_lab_paths is not None:
self.logger("Auditing number of target classes. This may take "
"a while as data must be read from disk."
"\n-- Note: avoid this by manually setting the "
"n_classes attribute in train_hparams.yaml.")
# Select up to 50 random images and find the unique classes
lab_paths = np.random.choice(self.nii_lab_paths,
min(50, len(self.nii_lab_paths)),
replace=False)
classes = []
for l in lab_paths:
classes.append(np.unique(nib.load(l).get_data()))
classes = np.unique(classes)
n_classes = classes.shape[0]
# Make sure the classes start from 0 and step continuously by 1
c_min, c_max = np.min(classes), np.max(classes)
if c_min != 0:
raise ValueError(
"Invalid class audit - Class integers should"
" start from 0, found %i (classes found: %s)" %
(c_min, classes))
if n_classes != max(classes) + 1:
raise ValueError("Invalid class audit - Found %i classes, but"
" expected %i, as the largest class value"
" found was %i. Classes found: %s" %
(n_classes, c_max + 1, c_max, classes))
else:
n_classes = None
classes = None
info = {
"shapes": shapes,
"real_sizes": real_sizes,
"pixdims": pixdims,
"memory_bytes": memory,
"n_channels": channels,
"n_classes": n_classes,
"classes": classes
}
return info
def real_shape(self):
"""
Returns:
The real (physical, scanner-space span) shape of the image
"""
return get_real_image_size(self.image_obj)