def on_train_begin(self, num_epochs, logs={}):
self.progress = 0
self.increment = rescale(1, self.num_epochs)
self.i = 1
self.x = []
self.losses = []
self.val_losses = []
self.logs = []
plt.figure()
plt.xlim(1, self.num_epochs)
plt.ylim(0, 1)
plt.plot(0, 0, label="loss")
plt.plot(0, 0, label="val_loss")
plt.title("Training and Validation loss [Epoch {}]".format(0))
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.savefig('temp/Epoch-{}.png'.format(0))
pixmap = QPixmap('temp/Epoch-{}.png'.format(0))
scaled_pixmap = pixmap.scaled(self.label.size(),
QtCore.Qt.KeepAspectRatio)
self.label.setPixmap(scaled_pixmap)
plt.close()
def start_reconstruction(self):
os.makedirs(self.reconstructed_path, exist_ok=True)
dcm_filenames = sorted(list(os.listdir(self.sinogram_path)))
file, header = nrrd.read(os.path.join(self.sinogram_path,
dcm_filenames[-1]),
index_order='C')
last_size = len(file)
length = (len(dcm_filenames) - 1) * 128 + last_size
progress = 0
increment = rescale(1, length)
it = 0
for dcm_file in tqdm(dcm_filenames):
if dcm_file.endswith('.nrrd'):
if self._case == 0:
recon_filename = os.path.join(
self.reconstructed_path,
'noisy_astra_{:03d}.nrrd'.format(it + 101))
nrrd.write(recon_filename,
self.reconstruct_astra(
os.path.join(self.sinogram_path, dcm_file),
progress, increment),
index_order='C')
if self._case == 1:
recon_filename = os.path.join(
self.reconstructed_path,
'noisy_odl_{:03d}.nrrd'.format(it + 101))
nrrd.write(recon_filename,
self.reconstruct_odl(
os.path.join(self.sinogram_path, dcm_file),
progress, increment),
index_order='C')
it += 1
progress += 100 / len(dcm_filenames)
self.progressbar.emit(100)
def start_sinogram_generation(self):
file_list = list_files(self.files_path)
os.makedirs(self.sinogram_path, exist_ok=True)
os.makedirs(self.ground_truth_path, exist_ok=True)
lidc_idri_gen_len = len(file_list)
vol_geom = astra.create_vol_geom(512, 512)
angles = np.linspace(0, np.pi, 1000, False)
proj_geom = astra.create_proj_geom('parallel', 1.0, 727, angles)
projector_id = astra.create_projector('cuda', proj_geom, vol_geom)
reco_space = odl.uniform_discr(min_pt=self.MIN_PT, max_pt=self.MAX_PT,
shape=self.RECO_IM_SHAPE, dtype=np.float32)
space = odl.uniform_discr(min_pt=self.MIN_PT, max_pt=self.MAX_PT, shape=self.IM_SHAPE,
dtype=np.float32)
reco_geometry = odl.tomo.parallel_beam_geometry(
reco_space, num_angles=self.NUM_ANGLES)
geometry = odl.tomo.parallel_beam_geometry(
space, num_angles=self.NUM_ANGLES, det_shape=reco_geometry.detector.shape)
# IMPL = 'astra_cpu'
# reco_ray_trafo = odl.tomo.RayTransform(reco_space, reco_geometry)
ray_trafo = odl.tomo.RayTransform(space, geometry)
rs = np.random.RandomState(3)
n_files = ceil(lidc_idri_gen_len / self.NUM_SAMPLES_PER_FILE)
slices = self.lidc_idri_gen(file_list)
it1 = 0
it2 = self.NUM_SAMPLES_PER_FILE
progress = 0
increment = rescale(1, n_files)
for filenumber in tqdm(range(n_files)):
obs_filename = os.path.join(
self.sinogram_path, 'sinogram_{:03d}.nrrd'.format(filenumber))
ground_truth_filename = os.path.join(
self.ground_truth_path, 'ground_truth_{:03d}.nrrd'.format(filenumber))
observation_dataset = []
ground_truth_dataset = []
for data in tqdm(slices[it1:it2, ...]):
data = np.flipud(data)
# ground_truth_dataset.append(data)
data = self.ff(ray_trafo, data)
data /= self.PHOTONS_PER_PIXEL
np.maximum(0.1 / self.PHOTONS_PER_PIXEL, data, out=data)
np.log(data, out=data)
data /= (-self.MU_MAX)
observation_dataset.append(data)
if self.FLAG:
print('terminating process')
return
it1 += self.NUM_SAMPLES_PER_FILE
it2 += self.NUM_SAMPLES_PER_FILE
if it2 > len(slices):
it2 = len(slices)
# ground_truth_dataset = np.array(ground_truth_dataset)
# ground_truth_dataset = np.rot90(ground_truth_dataset, -1, (1, 2))
nrrd.write(obs_filename, np.array(observation_dataset), index_order='C')