def test_gen_index_combs():
"""Check whether correct number of combinations are generated."""
N = MRAIDenseNeuralNetwork()
x = np.arange(10)
y = np.arange(10)
combs = N.gen_index_combs([x, y])
assert combs.shape[0] == 100
assert combs.shape[1] == 2
def test_subsample_rows():
"""Correct shape and contents."""
A = np.arange(24).reshape((12, 2))
N = MRAIDenseNeuralNetwork()
a = N.subsample_rows(A, num_draw=5)
assert a.shape[0] == 5
assert a.shape[1] == 2
assert len(np.unique(a) == 5)
def test_matrix2sparse():
"""Test all pixels are mapped to sparse."""
A = np.arange(24).reshape((6, 4))
N = MRAIDenseNeuralNetwork()
sA = N.matrix2sparse(A)
assert sA.shape[0] == 24
assert sA.shape[1] == 3
assert len(np.setdiff1d(sA[:, 0], np.arange(6))) == 0
assert len(np.setdiff1d(sA[:, 1], np.arange(4))) == 0
assert len(np.setdiff1d(sA[:, 2], np.arange(24))) == 0
def test_sample_pairs():
"""Produces correct output shape."""
# Network
N = MRAIDenseNeuralNetwork(patch_size=(9, 9), dense_size=[2], num_draw=2)
# Source array
X = rn.randn(32, 32)
Y = np.round(np.linspace(0, 3, 32**2)).reshape((32, 32))
y = N.matrix2sparse(Y, edge=(4, 4))
# Target array
Z = rn.randn(32, 32)
U = np.round(np.linspace(0, 3, 32**2)).reshape((32, 32))
u = N.matrix2sparse(U, edge=(4, 4))
# Sample pairs
P, S = N.sample_pairs(X, y, Z, u, num_draw=(2, 1))
# Extract pairs
A, B, a, b = P
# Check for correct shapes
assert len(A.shape) == 4
assert A.shape[1] == A.shape[2]
assert A.shape[3] == 1
assert A.shape[0] == a.shape[0]
assert len(B.shape) == 4
assert B.shape[1] == B.shape[2]
assert B.shape[3] == 1
assert B.shape[0] == b.shape[0]
# Check for correct contents
assert len(np.setdiff1d(np.unique(a), [0, 1])) == 0
assert len(np.setdiff1d(np.unique(b), [0, 1])) == 0
assert len(np.setdiff1d(np.unique(S), [0, 1])) == 0
def test_init_patch_size():
"""Valid patch size."""
with pytest.raises(ValueError):
N = MRAIDenseNeuralNetwork(patch_size=(-1, 1))
def test_l2_norm():
"""Test non-negative norm."""
N = MRAIDenseNeuralNetwork()
with tf.Session().as_default():
norms = N.l2_norm([rn.randn(100, 1), rn.randn(100, 1)]).eval()
assert np.all(np.array(norms) >= 0)
def test_contrastive_loss():
"""Test non-negative contrastive loss."""
N = MRAIDenseNeuralNetwork()
with tf.Session().as_default():
assert N.contrastive_loss(label=1, distance=rn.randn(1)).eval() >= 0
assert N.contrastive_loss(label=0, distance=rn.randn(1)).eval() >= 0
def test_init_dense_size():
"""Valid dense size."""
with pytest.raises(ValueError):
N = MRAIDenseNeuralNetwork(dense_size=[])
# Load source MRI-scan and corresponding segmentation
X = np.load('./data/subject01_GE2D_1.5T.npy')
Y = np.load('./data/subject01_segmentation.npy')
# Load target MRI-scan and corresponding segmentation
Z = np.load('./data/subject02_GE2D_3.0T.npy')
U = np.load('./data/subject02_segmentation.npy')
# Note that U is missing a lot of labels
print('Proportion missing labels = ' + str(np.mean(~np.isnan(U.ravel()))))
'''Set up MRAI network'''
# Initialize and compile a small neural network
N = MRAIDenseNeuralNetwork(patch_size=(31, 31),
dense_size=[16, 8],
batch_size=128,
num_epochs=4,
num_draw=10,
margin=10)
'''Train the net'''
# Call training procedure on source and target data
N.train(X, Y, Z, U, num_targets=1)
'''Map images to MRAI representation.'''
# Extract all source patches and feed them through network.
PX = extract_all_patches(X[0], patch_size=(31, 31), add_4d=True)
HX = N.feedforward(PX, scan_ID=0)
# Map label image to sparse array format
sY = N.matrix2sparse(Y[0])