def fit(self, directory, subject_index=None):
"""Ffitits internal statistics to some sample data.
"""
self.mean = np.zeros((1,) + self.image_shape, dtype=K.floatx())
self.std = np.zeros((1,) + self.image_shape, dtype=K.floatx())
nb_train_samples = len(subject_index)
paths_train = []
for root, dirs, files in os.walk(directory):
for file in files:
if file.endswith(".npz"):
paths_train.append(os.path.join(root, file))
filenames = sort_nicely(paths_train)
for i in range(nb_train_samples):
fname = filenames[subject_index[i]]
img, label = load_img(fname)
self.mean += np.true_divide(img,nb_train_samples)
for i in range(nb_train_samples):
fname = filenames[subject_index[i]]
img, label = load_img(fname)
self.std += np.true_divide(np.square(img-self.mean),nb_train_samples)
self.std = np.sqrt(self.std)
def __init__(self, directory, image_data_generator,
subject_index=None,
image_shape=(256, 256, 256),
nb_class = None,
labels_permuted=None,
batch_size=32, shuffle=True, seed=None):
self.directory = directory
self.image_data_generator = image_data_generator
self.nb_class = nb_class
self.image_shape = image_shape
self.labels_permuted = labels_permuted
# first, count the number of samples and classes
self.nb_sample = len(subject_index)
self.filenames = []
self.subject_index = subject_index
paths_train = []
for root, dirs, files in os.walk(directory):
for file in files:
if file.endswith(".npz"):
paths_train.append(os.path.join(root, file))
self.filenames = sort_nicely(paths_train)
print('Found %d neuroimages in the directory.' % (self.nb_sample))
super(DirectoryIterator, self).__init__(self.nb_sample, batch_size, shuffle, seed)
def create_kernel(config_module):
paths = config_module.path_files
input_data_type = config_module.input_data_type
experiment_name = config_module.experiment_name
labels_file = paths["labels_file"]
data_dir = paths["raw_images_dir"]
mask_file = paths["mask_file"]
print("Creating precomputed linear kernels for experiment: ",
experiment_name)
save_dir = "./results/" + experiment_name + "/SVM/precomputed_kernel/"
if not os.path.exists(save_dir):
os.makedirs(save_dir)
kernel_file = "precomputed_kernel.npz"
print("Reading labels from %s" % labels_file)
labels = np.genfromtxt(labels_file, delimiter=',', dtype='int8')
print(" # of labels samples: %d " % len(labels))
if mask_file is not "":
mask = nib.load(mask_file)
mask = mask.get_data()
mask = np.asarray(mask, dtype='float32')
mask = np.nan_to_num(mask)
print("Reading images with format {} from: %s".format(
input_data_type, data_dir))
img_paths = glob.glob(data_dir + "/*" + input_data_type)
img_paths = sort_nicely(img_paths)
img_names = []
for img_name in img_paths:
img_names.append(os.path.splitext(os.path.basename(img_name))[0])
n_samples = len(labels)
if n_samples != len(img_paths):
raise ValueError('Different number of labels and images files')
print("Loading images")
print(" # of images samples: %d " % len(img_paths))
n_samples = len(img_paths)
print(n_samples)
K = np.float64(np.zeros((n_samples, n_samples)))
step_size = 30
# outer loop
for i in range(int(np.ceil(n_samples / np.float(step_size)))):
it = i + 1
max_it = int(np.ceil(n_samples / np.float(step_size)))
print(" outer loop iteration: %d of %d." % (it, max_it))
# generate indices and then paths for this block
start_ind_1 = i * step_size
stop_ind_1 = min(start_ind_1 + step_size, n_samples)
block_paths_1 = img_paths[start_ind_1:stop_ind_1]
# read in the images in this block
images_1 = []
for k, path in enumerate(block_paths_1):
img = nib.load(path)
img = img.get_data()
img = np.asarray(img, dtype='float64')
img = np.nan_to_num(img)
if mask_file is not "":
img_vec = img[mask > 0]
else:
img_vec = np.reshape(img, np.product(img.shape))
images_1.append(img_vec)
del img
images_1 = np.array(images_1)
for j in range(i + 1):
it = j + 1
max_it = i + 1
print(" inner loop iteration: %d of %d." % (it, max_it))
# if i = j, then sets of image data are the same - no need to load
if i == j:
start_ind_2 = start_ind_1
stop_ind_2 = stop_ind_1
images_2 = images_1
# if i !=j, read in a different block of images
else:
start_ind_2 = j * step_size
stop_ind_2 = min(start_ind_2 + step_size, n_samples)
block_paths_2 = img_paths[start_ind_2:stop_ind_2]
images_2 = []
for k, path in enumerate(block_paths_2):
img = nib.load(path)
img = img.get_data()
img = np.asarray(img, dtype='float64')
img = np.nan_to_num(img)
if mask_file is not "":
img_vec = img[mask > 0]
else:
img_vec = np.reshape(img, np.product(img.shape))
images_2.append(img_vec)
del img
images_2 = np.array(images_2)
block_K = np.dot(images_1, np.transpose(images_2))
K[start_ind_1:stop_ind_1, start_ind_2:stop_ind_2] = block_K
K[start_ind_2:stop_ind_2,
start_ind_1:stop_ind_1] = np.transpose(block_K)
print("")
print("Saving Dataset")
print(" Kernel+Labels:" + kernel_file)
np.savez(save_dir + kernel_file, kernel=K, labels=labels, names=img_names)
print("Done")
def create_npy(config_module):
paths = config_module.path_files
input_data_type = config_module.input_data_type
experiment_name = config_module.experiment_name
labels_file = paths["labels_file"]
data_dir = paths["raw_images_dir"]
mask_file = paths["mask_file"]
print("Saving 3D images using .npz format for experiment: ",
experiment_name)
save_dir = "./results/" + experiment_name + "/CNN/img_npz/"
if not os.path.exists(save_dir):
os.makedirs(save_dir)
print("Reading labels from %s" % labels_file)
labels = np.genfromtxt(labels_file, delimiter=',', dtype='int8')
print(" # of labels samples: %d " % len(labels))
print("Reading images with format {} from: %s".format(
input_data_type, data_dir))
img_paths = glob.glob(data_dir + "/*" + input_data_type)
img_paths = sort_nicely(img_paths)
n_samples = len(labels)
if n_samples != len(img_paths):
raise ValueError('Different number of labels and images files')
# IF YOU WANT TO REMOVE BLANK VOXELS, UNCOMMENT THIS SECTION
print(
"Calculating 3d dimensions to remove blank voxels and reduce the 3d volume."
)
if mask_file:
mask = nib.load(mask_file)
mask = mask.get_data()
mask = np.asarray(mask, dtype='float32')
mask = np.nan_to_num(mask)
max_x, max_y, max_z, min_x, min_y, min_z = find_image_boundary(
mask_file)
print(max_x, max_y, max_z)
print(min_x, min_y, min_z)
else:
min_x = 1000
max_x = 0
min_y = 1000
max_y = 0
min_z = 1000
max_z = 0
for path in img_paths:
max_x_img, max_y_img, max_z_img, min_x_img, min_y_img, min_z_img = find_image_boundary(
path)
# X axis
if min_x > min_x_img:
min_x = min_x_img
if max_x < max_x_img:
max_x = max_x_img
# Y axis
if min_y > min_y_img:
min_y = min_y_img
if max_y < max_y_img:
max_y = max_y_img
# Z axis
if min_z > min_z_img:
min_z = min_z_img
if max_z < max_z_img:
max_z = max_z_img
print(max_x, max_y, max_z)
print(min_z, min_y, min_x)
print("Loading images")
print(" # of images samples: %d " % len(img_paths))
print("")
print("{:<5} {:100s} {:15s}\tCLASS\tMIN - MAX VALUES".format(
'#', 'FILENAME', 'DIMENSIONS'))
for k, path in enumerate(img_paths):
label = labels[k]
img = nib.load(path)
img = img.get_data()
img = np.asarray(img, dtype='float32')
img = np.nan_to_num(img)
if mask_file:
img = np.multiply(img, mask)
img = img[min_x:max_x, min_y:max_y, min_z:max_z]
print("{:<5} {:100s} ({:3}, {:3}, {:3})\t{:}\t{:6.4} - {:6.4}".format(
(k + 1), os.path.basename(os.path.normpath(path)), img.shape[0],
img.shape[1], img.shape[2], labels[k], np.min(img), np.max(img)))
img = np.true_divide(img, np.max(img))
img = np.reshape(img, (1, img.shape[0], img.shape[1], img.shape[2]))
f_name = os.path.splitext(os.path.basename(path))[0]
np.savez(save_dir + f_name, image=img, label=label)
del img
print("Done")