def key2BSTM(self, key):
trained = (self & key).fetch1()
voxel = key['vx'], key['vy'], key['vz']
b = RankDegenerateBernoulliProcess(voxel, quadratic_channels=key['quadratic_components'],
linear_channels=key['linear_components'],
common_channels=key['common_components'])
b.set_parameters(**trained)
return b
def test_convolution():
"""Testing separable convolution"""
channels = 2
flt_row, flt_col, flt_depth = (7, 5, 3)
in_channels = 1
b = RankDegenerateBernoulliProcess((flt_row, flt_col, flt_depth),
quadratic_channels=channels, linear_channels=channels)
X = np.random.randn(50, 40, 30, 1)
X_ = th.shared(np.require(X, dtype=floatX), borrow=True, name='stack')
c_, (xy_, z_) = b._build_separable_convolution(channels, X_, X.shape)
xy = np.random.randn(channels, flt_row, flt_col, in_channels)
z = np.random.randn(channels, 1, flt_depth, in_channels)
f = th.function([xy_, z_], c_)
for k in range(channels):
inter1 = [signal.convolve2d(e, xy[k, ...].squeeze(), mode='valid') for e in X.squeeze().transpose([2, 0, 1])]
inter1 = np.stack(inter1, axis=2)
inter2 = 0 * inter1[..., flt_depth - 1:]
for i, j in itertools.product(range(44), range(36)):
inter2[i, j, :] = signal.convolve(inter1[i, j, :], z[k].squeeze(), mode='valid')
assert_true(np.abs(inter2 - f(xy, z)[k, ...]).max() < 1e-10)
def _make_tuples(self, key):
X = Stacks().load(key)
voxel = (VoxelSize() & key).fetch1['vx', 'vy', 'vz']
b = RankDegenerateBernoulliProcess(voxel,
quadratic_channels=key['quadratic_components'],
linear_channels=key['linear_components'],
common_channels=key['common_components']
)
cells = (CellLocations() & key).fetch1['cells']
b.fit(X, cells, maxiter=100)
key.update(b.parameters)
key['train_cross_entropy'] = b.cross_entropy(X, cells)
key['train_auc_weighted'] = b.auc(X, cells, average='weighted')
key['train_auc'] = b.auc(X, cells, average='macro')
self.insert1(key)
self.cells = np.asarray(fid['cells']) if 'cells' in fid else None
self.P = np.asarray(fid['P']) if 'P' in fid else np.nan * np.empty(X.shape[:3])
def save(self, stackfile):
mode = 'r+' if os.path.exists(stackfile) else 'w'
with h5py.File(stackfile, mode) as fid:
fid.create_dataset('stack', self.stack.shape, dtype=float, data=self.stack)
fid.create_dataset('P', self.P.shape, dtype=float, data=self.P)
if self.cells is not None:
fid.create_dataset('cells', self.cells.shape, dtype=int, data=self.cells)
if __name__ == "__main__":
# s = Stack('data/smaller.hdf5', preprocessor=lambda x: preprocess(x).mean(axis=-1).squeeze())
# b = FullBernoulliProcess((9, 9, 7), quadratic_channels=3, linear_channels=3)
# s = Stack('data/sanity.hdf5', preprocessor=lambda x: x.mean(axis=-1).squeeze())
# s_test = Stack('data/sanity_test.hdf5', preprocessor=lambda x: x.mean(axis=-1).squeeze())
# b = RankDegenerateBernoulliProcess( (3,3,3), quadratic_channels=3, linear_channels=3, common_channels=3)
s = Stack('data/smaller.hdf5',
preprocessor=lambda x: histeq(unsharp_masking(medianfilter(center(x.squeeze()))), 500).mean(
axis=-1))
b = RankDegenerateBernoulliProcess((11, 11, 9), quadratic_channels=40, linear_channels=40, common_channels=20)
# s = Stack('data/2015-08-25_12-49-41_2015-08-25_13-02-18.h5',
# preprocessor=lambda x: histeq(unsharp_masking(medianfilter(center(x.squeeze()))),500).mean(axis=-1))
# b = RankDegenerateBernoulliProcess((19, 19, 17), quadratic_channels=40, linear_channels=40, common_channels=20)
b.fit(s.X, s.cells, maxiter=50)
b.visualize(s.X, s.cells)
