def preprocess_Tmap(self, T1map, T2map):
print('%%%%%%%%%%%%%%%%%%%5: ', T1map.shape)
T1map = numpy.fliplr(numpy.flipud(T1map))
T2map = numpy.fliplr(numpy.flipud(T2map))
Tmap = numpy.stack((T1map, T2map), axis=0).transpose(2, 1, 0)
Tmap = util.preprocess_tissue_property(Tmap).transpose(2, 1, 0)
return Tmap.transpose(1, 2, 0)
def preprocess_Tmap(self, T1map, T2map):
Tmap = numpy.stack((T1map,T2map), axis=0).transpose(2,1,0)
Tmap = util.preprocess_tissue_property(Tmap).transpose(2,1,0)
return Tmap
def initialize(self, opt):
if opt.isTrain:
self.augmentation = opt.augmentation
else:
self.augmentation = False
# self.augmentation = False
self.augmentation_type = 'flip'
self.goal_type = 'dict'
self.mask_type = opt.mask
self.get_paths(opt)
self.A_imgs = []
self.B_imgs = []
self.masks = []
for i, A_path in enumerate(self.A_paths):
A_path1 = A_path
f = h5py.File(A_path1)
A_img = f['imMRF'][0:int(opt.input_nc / 2), :, :]
A_img = numpy.transpose(A_img)
A_img = numpy.concatenate((A_img['real'], A_img['imag']),
axis=2).astype('float32')
# A_img = numpy.concatenate((A_img[:,:,0:int(opt.input_nc/2)],A_img[:,:,2304:2304+int(opt.input_nc/2)]), axis=2)
f.close()
# normalization
if opt.data_norm == 'non':
print("no normalization")
pass
else:
start = time.time()
t = numpy.mean(A_img**2, axis=2) * 2
t = t[:, :, numpy.newaxis]
A_img = A_img / (t**0.5)
A_img = A_img / 36
end = time.time()
util.print_log(
'Time for normalizing energy: %.5fs ' % (end - start),
opt.file_name)
# magnitude
if opt.magnitude:
ntp = int(opt.input_nc / 2)
A_img_mag = numpy.copy(A_img[:, :, 0:ntp])
for k in range(ntp):
A_img_mag[:, :, k] = (A_img[:, :, k]**2 +
A_img[:, :, ntp + k]**2)**0.5
A_img = A_img_mag
# set background signals as zero
# f = h5py.File(self.mask_large_paths[i])
# mask_large = numpy.transpose(f['mask']).astype('float32')
# f.close()
# A_img = A_img * mask_large[:,:,numpy.newaxis]
# load mask
if self.mask_type == 'tight':
f = h5py.File(self.mask_tight_paths[i])
else:
f = h5py.File(self.mask_large_paths[i])
mask = numpy.transpose(f['mask']).astype('float32')
f.close()
# clear mask
# mask = mask * 0 + 1
# load ground truth
if self.goal_type == 'densedict':
f = h5py.File(self.goal_densedict_paths[i])
else:
f = h5py.File(self.goal_paths[i])
T1 = numpy.transpose(f['t1big']).astype('float32')
T1 = T1[:, :, numpy.newaxis]
T2 = numpy.transpose(f['t2big']).astype('float32')
T2 = T2[:, :, numpy.newaxis]
B_img = numpy.concatenate((T1, T2), axis=2)
B_img = util.preprocess_tissue_property(B_img)
if self.goal_type == 'densedict':
B_img = numpy.flip(B_img, (0, 1)).copy()
f.close()
mask = mask[:, :, numpy.newaxis]
assert A_img.ndim == 3 and B_img.ndim == 3, "# of dim is not 3 for training image"
if opt.set_type == 'train':
A_img = A_img[40:216, 52:236, :]
B_img = B_img[40:216, 52:236, :]
mask = mask[40:216, 52:236, :]
A_img = torch.from_numpy(A_img)
B_img = torch.from_numpy(B_img)
mask = torch.from_numpy(mask)
if opt.input_nc == 2304 * 2:
print('half float16')
A_img = A_img.half()
B_img = B_img.half()
mask = mask.half()
A_img = A_img.permute(2, 0, 1)
B_img = B_img.permute(2, 0, 1)
mask = mask.permute(2, 0, 1)
self.A_imgs.append(A_img)
self.B_imgs.append(B_img)
self.masks.append(mask)
print("loaded image: %s" % A_path)
self.num_imgs = len(self.A_imgs)
if opt.patchSize != 0:
# self.num_patch = self.num_imgs*int((self.A_imgs[0].shape[1]*self.A_imgs[0].shape[2])/(opt.patchSize**2))
# self.num_patch = math.ceil(self.num_patch/opt.batchSize)*opt.batchSize
self.get_patch_pos(opt)