def dump_in_hdf5(db_path, output, title):
# TODO: use the number of subjects as the number of records of the table
images_dir = data_api.get_images_dir_path(db_path)
clinic_filename = data_api.get_clinic_file_path(db_path)
# Open the output file
h5file = tables.openFile(output, mode="w", title=title)
# Open the clinic file
csv_fd = open(clinic_filename)
data = pandas.io.parsers.read_csv(csv_fd)
n_subjects = data.shape[0]
# Load mask
mask_filename = data_api.get_mask_file_path(db_path)
print "Loading mask {mask_filename}".format(mask_filename=mask_filename)
babel_mask = nibabel.load(mask_filename)
mask = babel_mask.get_data()
binary_mask = mask != 0
useful_voxels = numpy.ravel_multi_index(numpy.where(binary_mask),
mask.shape)
n_useful_voxels = len(useful_voxels)
print "Mask loaded ({n_useful_voxels} useful voxels per image)".format(
n_useful_voxels=n_useful_voxels)
# Load grey matter images (X), apply mask and concatenate them
print "Loading {n_images} images, apply mask and flatten".format(
n_images=n_subjects)
image_filenames = [
os.path.join(images_dir, 'smwc1' + filename)
for filename in data.Images
]
masked_images = numpy.zeros((n_subjects, n_useful_voxels))
for (index, filename) in enumerate(image_filenames):
# Load (as numpy array)
image = nibabel.load(filename).get_data()
# Apply mask (returns a flat image)
masked_image = image[binary_mask]
# Store in X
masked_images[index, :] = masked_image
# Store data
data_api.write_images(h5file, masked_images)
h5file.close()
N_FOLDS_NESTED = 3
N_FOLDS_EVAL = 3
else:
N_FOLDS_NESTED = 5
N_FOLDS_EVAL = 10
OUT_DIR = os.path.join(DB_PATH, 'results', 'svm_feature_selection/')
if not os.path.exists(OUT_DIR):
os.makedirs(OUT_DIR)
WF_NAME_PATTERN = "svm_feature_selection_{images}"
#########################
# Oth step: access data #
#########################
csv_file_name = data_api.get_clinic_file_path(DB_PATH)
df = data_api.read_clinic_file(csv_file_name)
babel_mask = nibabel.load(data_api.get_mask_file_path(DB_PATH))
mask = babel_mask.get_data()
binary_mask = mask != 0
h5file = tables.openFile(LOCAL_PATH)
####################
# Create workflows #
####################
# Base workflow: SVM + feature selection
svms = pipelines = epac.Methods(*[
epac.Pipe(
type=str,
default=DEFAULT_WF_NAME,
help='Name of the workflow (default: %s)' %
(DEFAULT_WF_NAME))
args = parser.parse_args()
if TEST_MODE:
DB_PATH = '/volatile/DB/micro_subdepression/'
LOCAL_PATH = '/volatile/DB/cache/micro_subdepression.hdf5'
else:
DB_PATH = '/neurospin/brainomics/2013_imagen_subdepression'
LOCAL_PATH = '/volatile/DB/cache/imagen_subdepression.hdf5'
# Compute a priori probabilities
clinic_file_path = data_api.get_clinic_file_path(DB_PATH)
df = data_api.read_clinic_file(clinic_file_path)
N_SUBJECTS = float(df.shape[0])
counts = df['group_sub_ctl'].value_counts()
N_CONTROL = float(counts['control'])
P_CONTROL = N_CONTROL / N_SUBJECTS
N_SUBDEP = float(counts['sub'])
P_SUBDEP = N_SUBDEP / N_SUBJECTS
OUT_DIR = os.path.join(DB_PATH, 'results', 'svm')
WORKFLOW_PATH = os.path.join(OUT_DIR, args.wf_name)
if not os.path.exists(WORKFLOW_PATH):
raise Exception('{path} not found'.format(path=WORKFLOW_PATH))
svms_auto_cv = epac.map_reduce.engine.SomaWorkflowEngine.load_from_gui(
WORKFLOW_PATH)