def prep_input(im, acc=4):
"""Undersample the batch, then reformat them into what the network accepts.
Parameters
----------
gauss_ivar: float - controls the undersampling rate.
higher the value, more undersampling
"""
mask = cs.cartesian_mask(im.shape, acc, sample_n=8)
im_und, k_und = cs.undersample(im, mask, centred=False, norm='ortho')
im_gnd_l = to_lasagne_format(im)
im_und_l = to_lasagne_format(im_und)
k_und_l = to_lasagne_format(k_und)
mask_l = to_lasagne_format(mask, mask=True)
return im_und_l, k_und_l, mask_l, im_gnd_l
def prep_input(im, acc=4):
"""Undersample the batch, then reformat them into what the network accepts.
Parameters
----------
gauss_ivar: float - controls the undersampling rate.
higher the value, more undersampling
"""
mask = cs.cartesian_mask(im.shape, acc, sample_n=8)
im_und, k_und = cs.undersample(im, mask, centred=False, norm='ortho')
im_gnd_l = to_lasagne_format(im)
im_und_l = to_lasagne_format(im_und)
k_und_l = to_lasagne_format(k_und)
mask_l = to_lasagne_format(mask, mask=True)
return im_und_l, k_und_l, mask_l, im_gnd_l
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
# Specify network (d2_c2 or d5_c5)
input_shape = (batch_size, 2, Nx, Ny)
# net_config, net, = build_d2_c2(input_shape)
net_config, net, = build_d5_c5(input_shape)
# Load D5-C5 with pretrained params !! only build_d5_c5 is selected !!
# D5-C5 with pre-trained parameters
# with np.load('./models/pretrained/d5_c5.npz') as f:
# param_values = [f['arr_{0}'.format(i)] for i in range(len(f.files))]
# lasagne.layers.set_all_param_values(net, param_values)
# Compute acceleration rate
dummy_mask = cs.cartesian_mask((10, Nx, Ny), acc, sample_n=8)
sample_und_factor = cs.undersampling_rate(dummy_mask)
print('Undersampling Rate: {:.2f}'.format(sample_und_factor))
# Compile function
train_fn, val_fn = compile_fn(net, net_config, args)
# Create dataset
# train, validate, test = create_dummy_data()
train, validate, test = load_training_valid_test_data()
print('Start Training...')
for epoch in xrange(num_epoch):
t_start = time.time()
# Training
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
# Specify network
input_shape = (batch_size, 2, Nx, Ny)
net_config, net, = build_d2_c2(input_shape)
# # Load D5-C5 with pretrained params
# net_config, net, = build_d5_c5(input_shape)
# D5-C5 with pre-trained parameters
# with np.load('./models/pretrained/d5_c5.npz') as f:
# param_values = [f['arr_{0}'.format(i)] for i in range(len(f.files))]
# lasagne.layers.set_all_param_values(net, param_values)
# Compute acceleration rate
dummy_mask = cs.cartesian_mask((10, Nx, Ny), acc, sample_n=8)
sample_und_factor = cs.undersampling_rate(dummy_mask)
print('Undersampling Rate: {:.2f}'.format(sample_und_factor))
# Compile function
train_fn, val_fn = compile_fn(net, net_config, args)
# Create dataset
train, validate, test = create_dummy_data()
print('Start Training...')
for epoch in xrange(num_epoch):
t_start = time.time()
# Training
train_err = 0
arr1 = sitk.GetArrayFromImage(img1)
arr2 = sitk.GetArrayFromImage(img2)
arr3 = sitk.GetArrayFromImage(img3)
arr4 = sitk.GetArrayFromImage(img4)
arr5 = sitk.GetArrayFromImage(img5)
arr6 = sitk.GetArrayFromImage(img6)
#combining array
arr = np.concatenate((arr1, arr2, arr3, arr4, arr5, arr6), axis=0)
np.save('image_array', arr)
#fft to numpy array
fft_arr = np.fft.fft2(arr)
fshift_out = np.fft.fftshift(fft_arr)
#cartesian mask
mask_1 = cs.cartesian_mask(arr.shape, 4, sample_n=10, centred=False)
im_und, k_und = cs.undersample(arr, mask_1, centred=False, norm='ortho')
#normalising the input and output data
features_IMG = tf.keras.utils.normalize(im_und, axis=1)
x_train = np.expand_dims(features_IMG, 3)
print(np.shape(x_train))
features_KS = tf.keras.utils.normalize(k_und, axis=1)
x_ks_train = np.expand_dims(features_KS, 3)
Target = tf.keras.utils.normalize(arr, axis=1)
y_train = np.expand_dims(Target, 3)
mask_data = np.expand_dims(mask_1[1], 3)
mask_ksp = mask_data.astype(int)
inv_mask = np.bitwise_not(mask_ksp)
print(np.shape(mask_data))
la = 0.001
warm_start = 'off'
warm_start_factor = 0
perp_start = 'rand'
perp_start_factor = 0.01
momentum = 0.9
learning_rate = 0.01
verbose = True
data = load_data()
#batch_size = data.shape[0];
Nt, Nx, Ny = data.shape
input_shape = (batch_size, 2, Nx, Ny)
mask = cs.cartesian_mask((batch_size, Nx, Ny),
undersampling_rate,
sample_n=8)
# mask = np.zeros([2,2,Nx,Ny]).astype('float32');
# mask[:,0,:,:]= dummy_mask;
# mask[:,1,:,:]= dummy_mask;
#
# runner_id = 25;
# runner = load_runner(runner_id);
# fname = 'data/data_runner_%d_' % (runner_id);
# runner.export_data(fname);
#
# mask = runner.mask[0];
net_config, net = load_network(input_shape, network_path)
val_fn, df = compile_functions(net, net_config)
# Configure directory info
project_root = '.'
save_dir = join(project_root, 'models/%s' % model_name)
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
# Specify network
input_shape = (batch_size, 2, Nx, Ny)
net_config, net, = build_UnetCascade(input_shape)
print(net_config)
# net_config, net, = build_d2_c2(input_shape)
# Compute acceleration rate
dummy_mask = cs.cartesian_mask((500, Nx, Ny),
gauss_ivar,
sample_high_freq=True,
sample_centre=True,
sample_n=8)
acc = dummy_mask.size / np.sum(dummy_mask)
print('Acceleration Rate: {:.2f}'.format(acc))
# Compile function
train_fn, val_fn = compile_fn(net, net_config, args)
# Create dataset
train, validate, test = create_data_knee()
print('data created')
#uncomment to start with already trained model
#name = '%s_epoch_%d.npz' % (model_name, 8550)
#param=np.load(join(save_dir, name))
