def symbol(name, dimensions, value=0., mode='function'):
"""Short-cut for symbol creation to test "function"
and "indexed" API."""
assert(mode in ['function', 'indexed'])
s = DenseData(name=name, dimensions=dimensions)
s.data[:] = value
return s.indexify() if mode == 'indexed' else s
def test_symbol_cache_aliasing_reverse():
"""Test to assert that removing he original u[x, y] instance does
not impede our alisaing cache or leaks memory.
"""
# Ensure a clean cache to start with
clear_cache()
# FIXME: Currently not working, presumably due to our
# failure to cache new instances?
# assert(len(_SymbolCache) == 0)
# Create first instance of u and fill its data
u = DenseData(name='u', shape=(3, 4))
u.data[:] = 6.
u_ref = weakref.ref(u.data)
# Create derivative and delete orignal u[x, y]
dx = u.dx
del u
clear_cache()
# We still have a references to u
# FIXME: Unreliable cache sizes
# assert len(_SymbolCache) == 1
# Ensure u[x + h, y] still holds valid data
assert np.allclose(dx.args[0].args[1].data, 6.)
del dx
clear_cache()
# FIXME: Unreliable cache sizes
# assert len(_SymbolCache) == 0 # We still have a reference to u_h
assert u_ref() is None
def test_data_movement_1D(self):
u = DenseData(name='yu1D', shape=(16,), dimensions=(x,), space_order=0)
u.data
assert type(u._data_object) == YaskGrid
u.data[1] = 1.
assert u.data[0] == 0.
assert u.data[1] == 1.
assert all(i == 0 for i in u.data[2:])
def test_cache_after_indexification():
"""Test to assert that the SymPy cache retrieves the right Devito data object
after indexification.
"""
u0 = DenseData(name='u', shape=(4, 4, 4), space_order=0)
u1 = DenseData(name='u', shape=(4, 4, 4), space_order=1)
u2 = DenseData(name='u', shape=(4, 4, 4), space_order=2)
for i in [u0, u1, u2]:
assert i.indexify().base.function.space_order ==\
(i.indexify() + 1.).args[1].base.function.space_order
def test_constant_time_dense(self):
"""Test arithmetic between different data objects, namely ConstantData
and DenseData."""
const = ConstantData(name='truc', value=2.)
a = DenseData(name='a', shape=(20, 20))
a.data[:] = 2.
eqn = Eq(a, a + 2.*const)
op = Operator(eqn)
op.apply(a=a, truc=const)
assert(np.allclose(a.data, 6.))
# Applying a different constant still works
op.apply(a=a, truc=ConstantData(name='truc2', value=3.))
assert(np.allclose(a.data, 12.))
def u(dims):
u = DenseData(name='yu3D',
shape=(16, 16, 16),
dimensions=(x, y, z),
space_order=0)
u.data # Trigger initialization
return u
def Born(self,
save=False,
cache_blocking=None,
auto_tuning=False,
dse='advanced',
dle='advanced',
compiler=None,
free_surface=False):
"""Linearized modelling of one or multiple point source from
an input model perturbation.
"""
nt, nrec = self.data.shape
nsrc = self.source.shape[1]
ndim = len(self.model.shape)
h = self.model.get_spacing()
dtype = self.model.dtype
nbpml = self.model.nbpml
# Create source symbol
src = PointSource(name='src',
ntime=self.source.nt,
coordinates=self.source.receiver_coords)
# Create receiver symbol
Linrec = Receiver(name='Lrec',
ntime=self.data.nt,
coordinates=self.data.receiver_coords)
# Create the forward wavefield
u = TimeData(name="u",
shape=self.model.shape_domain,
time_dim=nt,
time_order=2,
space_order=self.s_order,
dtype=self.model.dtype)
du = TimeData(name="du",
shape=self.model.shape_domain,
time_dim=nt,
time_order=2,
space_order=self.s_order,
dtype=self.model.dtype)
dm = DenseData(name="dm",
shape=self.model.shape_domain,
dtype=self.model.dtype)
# Execute operator and return wavefield and receiver data
born = BornOperator(self.model,
u,
du,
src,
Linrec,
dm,
self.data,
time_order=self.t_order,
spc_order=self.s_order,
cache_blocking=cache_blocking,
dse=dse,
dle=dle,
free_surface=free_surface)
return born
def test_inject_from_field(shape, coords, result, npoints=19):
"""Test point injection from a second field along a line
through the middle of the grid.
"""
a = unit_box(shape=shape)
spacing = a.data[tuple([1 for _ in shape])]
a.data[:] = 0.
b = DenseData(name='b', shape=a.data.shape)
b.data[:] = 1.
p = points(ranges=coords, npoints=npoints)
expr = p.inject(field=a, expr=b)
Operator(expr, subs={h: spacing})(a=a, b=b, t=1)
indices = [slice(4, 6, 1) for _ in coords]
indices[0] = slice(1, -1, 1)
assert np.allclose(a.data[indices], result, rtol=1.e-5)
def test_fd_space(derivative, space_order):
"""
This test compare the discrete finite-difference scheme against polynomials
For a given order p, the fiunite difference scheme should
be exact for polynomials of order p
:param derivative: name of the derivative to be tested
:param space_order: space order of the finite difference stencil
"""
clear_cache()
# dummy axis dimension
nx = 100
xx = np.linspace(-1, 1, nx)
dx = xx[1] - xx[0]
# Symbolic data
u = DenseData(name="u",
shape=(nx, ),
space_order=space_order,
dtype=np.float32)
du = DenseData(name="du",
shape=(nx, ),
space_order=space_order,
dtype=np.float32)
# Define polynomial with exact fd
h, y = symbols('h y')
coeffs = np.ones((space_order, ), dtype=np.float32)
polynome = sum([coeffs[i] * y**i for i in range(0, space_order)])
polyvalues = np.array([polynome.subs(y, xi) for xi in xx], np.float32)
# Fill original data with the polynomial values
u.data[:] = polyvalues
# True derivative of the polynome
Dpolynome = diff(diff(polynome)) if derivative == 'dx2' else diff(polynome)
Dpolyvalues = np.array([Dpolynome.subs(y, xi) for xi in xx], np.float32)
# FD derivative, symbolic
u_deriv = getattr(u, derivative)
# Compute numerical FD
stencil = Eq(du, u_deriv)
op = Operator(stencil, subs={h: dx})
op.apply()
# Check exactness of the numerical derivative except inside space_brd
space_border = space_order
error = abs(du.data[space_border:-space_border] -
Dpolyvalues[space_border:-space_border])
assert np.isclose(np.mean(error), 0., atol=1e-3)
def _new_operator1(shape, **kwargs):
infield = DenseData(name='infield', shape=shape, dtype=np.int32)
infield.data[:] = np.arange(reduce(mul, shape), dtype=np.int32).reshape(shape)
outfield = DenseData(name='outfield', shape=shape, dtype=np.int32)
stencil = Eq(outfield.indexify(), outfield.indexify() + infield.indexify()*3.0)
# Run the operator
op = Operator(stencil, **kwargs)
op(infield=infield, outfield=outfield)
return outfield, op
def Gradient(self,
cache_blocking=None,
auto_tuning=False,
dle='advanced',
dse='advanced',
compiler=None,
free_surface=False):
"""FWI gradient from back-propagation of a shot record
and input forward wavefield
"""
nt, nrec = self.data.shape
ndim = len(self.model.shape)
h = self.model.get_spacing()
dtype = self.model.dtype
nbpml = self.model.nbpml
# Create receiver symbol
rec = Receiver(name='rec',
ntime=self.data.nt,
coordinates=self.data.receiver_coords)
# Gradient symbol
grad = DenseData(name="grad",
shape=self.model.shape_domain,
dtype=self.model.dtype)
# Create the forward wavefield
v = TimeData(name="v",
shape=self.model.shape_domain,
time_dim=nt,
time_order=2,
space_order=self.s_order,
dtype=self.model.dtype)
u = TimeData(name="u",
shape=self.model.shape_domain,
time_dim=nt,
time_order=2,
space_order=self.s_order,
dtype=self.model.dtype,
save=True)
# Execute operator and return wavefield and receiver data
gradop = GradientOperator(self.model,
v,
grad,
rec,
u,
self.data,
time_order=self.t_order,
spc_order=self.s_order,
cache_blocking=cache_blocking,
dse=dse,
dle=dle,
tsave=self.tsave,
free_surface=free_surface)
return gradop
def test_dimension_size_infer(self, nt=100):
"""Test that the dimension sizes are being inferred correctly"""
i, j, k = dimify('i j k')
shape = tuple([d.size for d in [i, j, k]])
a = DenseData(name='a', shape=shape).indexed
b = TimeData(name='b', shape=shape, save=True, time_dim=nt).indexed
eqn = Eq(b[time, x, y, z], a[x, y, z])
op = Operator(eqn)
_, op_dim_sizes = op.arguments()
assert(op_dim_sizes[time.name] == nt)
def test_clear_cache(nx=1000, ny=1000):
clear_cache()
cache_size = len(_SymbolCache)
for i in range(10):
assert (len(_SymbolCache) == cache_size)
DenseData(name='u', shape=(nx, ny), dtype=np.float64, space_order=2)
assert (len(_SymbolCache) == cache_size + 1)
clear_cache()
def test_equations_mixed_densedata_timedata(self, shape, dimensions):
"""
Test that equations using a mixture of DenseData and TimeData objects
are embedded within the same time loop.
"""
a = TimeData(name='a', shape=shape, time_order=2, dimensions=dimensions,
space_order=2, time_dim=6, save=False)
p_aux = Dimension(name='p_aux', size=10)
b = DenseData(name='b', shape=shape + (10,), dimensions=dimensions + (p_aux,),
space_order=2)
b.data[:] = 1.0
b2 = DenseData(name='b2', shape=(10,) + shape, dimensions=(p_aux,) + dimensions,
space_order=2)
b2.data[:] = 1.0
eqns = [Eq(a.forward, a.laplace + 1.),
Eq(b, time*b*a + b)]
eqns2 = [Eq(a.forward, a.laplace + 1.),
Eq(b2, time*b2*a + b2)]
subs = {x.spacing: 2.5, y.spacing: 1.5, z.spacing: 2.0}
op = Operator(eqns, subs=subs, dle='noop')
trees = retrieve_iteration_tree(op)
assert len(trees) == 2
assert all(trees[0][i] is trees[1][i] for i in range(3))
op2 = Operator(eqns2, subs=subs, dle='noop')
trees = retrieve_iteration_tree(op2)
assert len(trees) == 2
assert all(trees[0][i] is trees[1][i] for i in range(3))
# Verify both operators produce the same result
op()
op2()
assert(np.allclose(b2.data[2, ...].reshape(-1) -
b.data[..., 2].reshape(-1), 0.))
def test_symbol_cache_aliasing():
"""Test to assert that our aiasing cache isn't defeated by sympys
non-aliasing symbol cache.
For further explanation consider the symbol u[x, y] and it's first
derivative in x, which includes the symbols u[x, y] and u[x + h, y].
The two functions are aliased in devito's caching mechanism to allow
multiple stencil indices pointing at the same data object u, but
SymPy treats these two instances as separate functions and thus is
allowed to delete one or the other when the cache is cleared.
The test below asserts that u[x + h, y] is deleted, the data on u
is still intact through our own caching mechanism."""
# Ensure a clean cache to start with
clear_cache()
# FIXME: Currently not working, presumably due to our
# failure to cache new instances?
# assert(len(_SymbolCache) == 0)
# Create first instance of u and fill its data
u = DenseData(name='u', shape=(3, 4))
u.data[:] = 6.
u_ref = weakref.ref(u.data)
# Create u[x + h, y] and delete it again
dx = u.dx # Contains two u symbols: u[x, y] and u[x + h, y]
del dx
clear_cache()
# FIXME: Unreliable cache sizes
# assert len(_SymbolCache) == 1 # We still have a reference to u
assert np.allclose(u.data, 6.) # u.data is alive and well
# Remove the final instance and ensure u.data got deallocated
del u
clear_cache()
assert u_ref() is None
def test_arithmetic_deep(expr, result):
"""Tests basic point-wise arithmetic on multi-dimensional data"""
i = Dimension(name='i', size=3)
j = Dimension(name='j', size=4)
k = Dimension(name='k', size=5)
l = Dimension(name='l', size=6)
a = DenseData(name='a', dimensions=(i, j, k, l))
a.data[:] = 2.
ai = a.indexify()
b = DenseData(name='b', dimensions=(j, k))
b.data[:] = 3.
bi = b.indexify()
eqn = eval(expr)
StencilKernel(eqn)(ai.base.function, bi.base.function)
assert np.allclose(ai.base.function.data, result, rtol=1e-12)
def test_indexed_open_loops(self, expr, result):
"""Test point-wise arithmetic with stencil offsets and open loop
boundaries in indexed expression format"""
i, j, l = dimify('i j l')
pushed = [d.size for d in [j, l]]
j.size = None
l.size = None
a = DenseData(name='a', dimensions=(i, j, l), shape=(3, 5, 6)).indexed
fa = a.function
fa.data[0, :, :] = 2.
eqn = eval(expr)
Operator(eqn)(a=fa)
assert np.allclose(fa.data[1, 1:-1, 1:-1], result[1:-1, 1:-1], rtol=1e-12)
j.size, l.size = pushed
def GradientOperator(model, source, receiver, time_order=2, space_order=4, **kwargs):
"""
Constructor method for the gradient operator in an acoustic media
:param model: :class:`Model` object containing the physical parameters
:param source: :class:`PointData` object containing the source geometry
:param receiver: :class:`PointData` object containing the acquisition geometry
:param time_order: Time discretization order
:param space_order: Space discretization order
"""
m, damp = model.m, model.damp
# Gradient symbol and wavefield symbols
grad = DenseData(name='grad', shape=model.shape_domain,
dtype=model.dtype)
u = TimeData(name='u', shape=model.shape_domain, save=True,
time_dim=source.nt, time_order=time_order,
space_order=space_order, dtype=model.dtype)
v = TimeData(name='v', shape=model.shape_domain, save=False,
time_order=time_order, space_order=space_order,
dtype=model.dtype)
rec = Receiver(name='rec', ntime=receiver.nt, ndim=receiver.ndim,
npoint=receiver.npoint)
if time_order == 2:
biharmonic = 0
dt = model.critical_dt
gradient_update = Eq(grad, grad - u.dt2 * v)
else:
biharmonic = v.laplace2(1/m)
biharmonicu = - u.laplace2(1/(m**2))
dt = 1.73 * model.critical_dt
gradient_update = Eq(grad, grad - (u.dt2 - s**2 / 12.0 * biharmonicu) * v)
# Derive stencil from symbolic equation
stencil = 1.0 / (2.0 * m + s * damp) * \
(4.0 * m * v + (s * damp - 2.0 * m) *
v.forward + 2.0 * s ** 2 * (v.laplace + s**2 / 12.0 * biharmonic))
eqn = Eq(v.backward, stencil)
# Add expression for receiver injection
ti = v.indices[0]
receivers = rec.inject(field=v, u_t=ti - 1, offset=model.nbpml,
expr=rec * dt * dt / m, p_t=time)
return Operator([eqn] + [gradient_update] + receivers,
subs={s: dt, h: model.get_spacing()},
time_axis=Backward, name='Gradient', **kwargs)
def test_equations_emulate_bc(self, t0):
"""
Test that bc-like equations get inserted into the same loop nest
as the "main" equations.
"""
shape = (3, 3, 3)
a = DenseData(name='a', shape=shape).indexed
b = TimeData(name='b', shape=shape, save=True, time_dim=6).indexed
main = Eq(b[time + 1, x, y, z], b[time - 1, x, y, z] + a[x, y, z] + 3.*t0)
bcs = [Eq(b[time, 0, y, z], 0.),
Eq(b[time, x, 0, z], 0.),
Eq(b[time, x, y, 0], 0.)]
op = Operator([main] + bcs, dse='noop', dle='noop')
trees = retrieve_iteration_tree(op)
assert len(trees) == 4
assert all(id(trees[0][0]) == id(i[0]) for i in trees)
def gradient(self, rec, u, v=None, grad=None, m=None, **kwargs):
"""
Gradient modelling function for computing the adjoint of the
Linearized Born modelling function, ie. the action of the
Jacobian adjoint on an input data.
:param recin: Receiver data as a numpy array
:param u: Symbol for full wavefield `u` (created with save=True)
:param v: (Optional) Symbol to store the computed wavefield
:param grad: (Optional) Symbol to store the gradient field
:returns: Gradient field and performance summary
"""
# Gradient symbol
if grad is None:
grad = DenseData(name='grad',
shape=self.model.shape_domain,
dtype=self.model.dtype)
# Create the forward wavefield
if v is None:
v = TimeData(name='v',
shape=self.model.shape_domain,
save=False,
time_order=self.time_order,
space_order=self.space_order,
dtype=self.model.dtype)
# Pick m from model unless explicitly provided
if m is None:
m = m or self.model.m
summary = self.op_grad.apply(rec=rec,
grad=grad,
v=v,
u=u,
m=m,
**kwargs)
return grad, summary
def __init__(self,
origin,
spacing,
shape,
vp,
nbpml=20,
dtype=np.float32,
epsilon=None,
delta=None,
theta=None,
phi=None):
self.origin = origin
self.spacing = spacing
self.shape = shape
self.nbpml = int(nbpml)
self.dtype = dtype
# Ensure same dimensions on all inpute parameters
assert (len(origin) == len(spacing))
assert (len(origin) == len(shape))
# Create square slowness of the wave as symbol `m`
if isinstance(vp, np.ndarray):
self.m = DenseData(name="m",
shape=self.shape_domain,
dtype=self.dtype)
else:
self.m = ConstantData(name="m", value=1 / vp**2, dtype=self.dtype)
# Set model velocity, which will also set `m`
self.vp = vp
# Create dampening field as symbol `damp`
self.damp = DenseData(name="damp",
shape=self.shape_domain,
dtype=self.dtype)
damp_boundary(self.damp.data, self.nbpml, spacing=self.get_spacing())
# Additional parameter fields for TTI operators
self.scale = 1.
if epsilon is not None:
if isinstance(epsilon, np.ndarray):
self.epsilon = DenseData(name="epsilon",
shape=self.shape_domain,
dtype=self.dtype)
self.epsilon.data[:] = self.pad(1 + 2 * epsilon)
# Maximum velocity is scale*max(vp) if epsilon > 0
if np.max(self.epsilon.data) > 0:
self.scale = np.sqrt(np.max(self.epsilon.data))
else:
self.epsilon = 1 + 2 * epsilon
self.scale = epsilon
else:
self.epsilon = 1
if delta is not None:
if isinstance(delta, np.ndarray):
self.delta = DenseData(name="delta",
shape=self.shape_domain,
dtype=self.dtype)
self.delta.data[:] = self.pad(np.sqrt(1 + 2 * delta))
else:
self.delta = delta
else:
self.delta = 1
if theta is not None:
if isinstance(theta, np.ndarray):
self.theta = DenseData(name="theta",
shape=self.shape_domain,
dtype=self.dtype)
self.theta.data[:] = self.pad(theta)
else:
self.theta = theta
else:
self.theta = 0
if phi is not None:
if isinstance(phi, np.ndarray):
self.phi = DenseData(name="phi",
shape=self.shape_domain,
dtype=self.dtype)
self.phi.data[:] = self.pad(phi)
else:
self.phi = phi
else:
self.phi = 0
def ai(x, y, name='a', value=2.):
a = DenseData(name=name, dimensions=(x, y))
a.data[:] = value
return a.indexify()
def densedata(name, shape, dimensions):
return DenseData(name=name, shape=shape, dimensions=dimensions)
def bi(x, y, name='b', value=3.):
b = DenseData(name=name, dimensions=(x, y))
b.data[:] = value
return b.indexify()
def BornOperator(model, source, receiver, time_order=2, space_order=4, **kwargs):
"""
Constructor method for the Linearized Born operator in an acoustic media
:param model: :class:`Model` object containing the physical parameters
:param source: :class:`PointData` object containing the source geometry
:param receiver: :class:`PointData` object containing the acquisition geometry
:param time_order: Time discretization order
:param space_order: Space discretization order
"""
m, damp = model.m, model.damp
# Create source and receiver symbols
src = PointSource(name='src', ntime=source.nt, ndim=source.ndim,
npoint=source.npoint)
rec = Receiver(name='rec', ntime=receiver.nt, ndim=receiver.ndim,
npoint=receiver.npoint)
# Create wavefields and a dm field
u = TimeData(name="u", shape=model.shape_domain, save=False,
time_order=time_order, space_order=space_order,
dtype=model.dtype)
U = TimeData(name="U", shape=model.shape_domain, save=False,
time_order=time_order, space_order=space_order,
dtype=model.dtype)
dm = DenseData(name="dm", shape=model.shape_domain,
dtype=model.dtype)
if time_order == 2:
biharmonicu = 0
biharmonicU = 0
dt = model.critical_dt
else:
biharmonicu = u.laplace2(1/m)
biharmonicU = U.laplace2(1/m)
dt = 1.73 * model.critical_dt
# Derive both stencils from symbolic equation
# first_eqn = m * u.dt2 - u.laplace + damp * u.dt
# second_eqn = m * U.dt2 - U.laplace - dm* u.dt2 + damp * U.dt
stencil1 = 1.0 / (2.0 * m + s * damp) * \
(4.0 * m * u + (s * damp - 2.0 * m) *
u.backward + 2.0 * s ** 2 * (u.laplace + s**2 / 12 * biharmonicu))
stencil2 = 1.0 / (2.0 * m + s * damp) * \
(4.0 * m * U + (s * damp - 2.0 * m) *
U.backward + 2.0 * s ** 2 * (U.laplace +
s**2 / 12 * biharmonicU - dm * u.dt2))
eqn1 = Eq(u.forward, stencil1)
eqn2 = Eq(U.forward, stencil2)
# Add source term expression for u
ti = u.indices[0]
source = src.inject(field=u, u_t=ti + 1, offset=model.nbpml,
expr=src * dt * dt / m, p_t=time)
# Create receiver interpolation expression from U
receivers = rec.interpolate(expr=U, u_t=ti, offset=model.nbpml)
return Operator([eqn1] + source + [eqn2] + receivers,
subs={s: dt, h: model.get_spacing()},
time_axis=Forward, name='Born', **kwargs)
def run(dimensions=(50, 50, 50),
tn=750.0,
time_order=2,
space_order=4,
nbpml=40,
dse='noop',
dle='noop'):
ndim = len(dimensions)
origin = tuple([0.] * ndim)
spacing = tuple([15.] * ndim)
f0 = .010
t0 = 0.0
# True velocity
true_vp = np.ones(dimensions) + .5
true_vp[..., int(dimensions[-1] / 2):] = 2.
# Smooth velocity - we use this as our initial m
initial_vp = smooth10(true_vp, dimensions)
# Model perturbation
model = Model(origin, spacing, true_vp.shape, true_vp, nbpml=nbpml)
m0 = np.float32(model.pad(initial_vp**-2))
dm = np.float32(model.m.data - m0)
dt = model.critical_dt
if time_order == 4:
dt *= 1.73
nt = int(1 + (tn - t0) / dt)
# Source geometry
time_series = np.zeros((nt, 1), dtype=np.float32)
time_series[:, 0] = source(np.linspace(t0, tn, nt), f0)
# Source location
location = np.zeros((1, ndim), dtype=np.float32)
location[0, :-1] = [
origin[i] + dimensions[i] * spacing[i] * .5 for i in range(ndim - 1)
]
location[0, -1] = origin[-1] + 2 * spacing[-1]
# Receivers locations
receiver_coords = np.zeros((dimensions[0], ndim), dtype=np.float32)
receiver_coords[:, 0] = np.linspace(0,
origin[0] +
(dimensions[0] - 1) * spacing[0],
num=dimensions[0])
receiver_coords[:, 1:] = location[0, 1:]
# Create source symbol
src = PointSource(name="src", data=time_series, coordinates=location)
# Receiver for true model
rec_t = Receiver(name="rec", ntime=nt, coordinates=receiver_coords)
# Receiver for smoothed model
rec_s = Receiver(name="rec", ntime=nt, coordinates=receiver_coords)
# Create the forward wavefield to use (only 3 timesteps)
# Once checkpointing is in, this will be the only wavefield we need
u_nosave = TimeData(name="u",
shape=model.shape_domain,
time_dim=nt,
time_order=time_order,
space_order=space_order,
save=False,
dtype=model.dtype)
# Forward wavefield where all timesteps are stored
# With checkpointing this should go away <----
u_save = TimeData(name="u",
shape=model.shape_domain,
time_dim=nt,
time_order=time_order,
space_order=space_order,
save=True,
dtype=model.dtype)
# Forward Operators - one with save = True and one with save = False
fw = ForwardOperator(model,
src,
rec_t,
time_order=time_order,
spc_order=space_order,
save=True,
dse=dse,
dle=dle)
fw_nosave = ForwardOperator(model,
src,
rec_t,
time_order=time_order,
spc_order=space_order,
save=False,
dse=dse,
dle=dle)
# Calculate receiver data for true velocity
fw_nosave.apply(u=u_nosave, rec=rec_t, src=src)
# Smooth velocity
# This is the pass that needs checkpointing <----
fw.apply(u=u_save, rec=rec_s, m=m0, src=src)
# Objective function value
F0 = .5 * linalg.norm(rec_s.data - rec_t.data)**2
# Receiver for Gradient
# Confusing nomenclature because this is actually the source for the adjoint
# mode
rec_g = Receiver(name="rec",
coordinates=rec_s.coordinates.data,
data=rec_s.data - rec_t.data)
# Gradient symbol
grad = DenseData(name="grad", shape=model.shape_domain, dtype=model.dtype)
# Reusing u_nosave from above as the adjoint wavefield since it is a temp var anyway
gradop = GradientOperator(model,
src,
rec_g,
time_order=time_order,
spc_order=space_order,
dse=dse,
dle=dle)
# Clear the wavefield variable to reuse it
# This time it represents the adjoint field
u_nosave.data.fill(0)
# Apply the gradient operator to calculate the gradient
# This is the pass that requires the checkpointed data
gradop.apply(u=u_save, v=u_nosave, m=m0, rec=rec_g, grad=grad)
# The result is in grad
gradient = grad.data
# <J^T \delta d, dm>
G = np.dot(gradient.reshape(-1), dm.reshape(-1))
# FWI Gradient test
H = [0.5, 0.25, .125, 0.0625, 0.0312, 0.015625, 0.0078125]
error1 = np.zeros(7)
error2 = np.zeros(7)
for i in range(0, 7):
# Add the perturbation to the model
mloc = m0 + H[i] * dm
# Set field to zero (we're re-using it)
u_nosave.data.fill(0)
# Receiver data for the new model
# Results will be in rec_s
fw_nosave.apply(u=u_nosave, rec=rec_s, m=mloc, src=src)
d = rec_s.data
# First order error Phi(m0+dm) - Phi(m0)
error1[i] = np.absolute(.5 * linalg.norm(d - rec_t.data)**2 - F0)
# Second order term r Phi(m0+dm) - Phi(m0) - <J(m0)^T \delta d, dm>
error2[i] = np.absolute(.5 * linalg.norm(d - rec_t.data)**2 - F0 -
H[i] * G)
# Test slope of the tests
p1 = np.polyfit(np.log10(H), np.log10(error1), 1)
p2 = np.polyfit(np.log10(H), np.log10(error2), 1)
assert np.isclose(p1[0], 1.0, rtol=0.1)
assert np.isclose(p2[0], 2.0, rtol=0.1)
def a(shape=(11, 11)):
x = np.linspace(0., 1., shape[0])
y = np.linspace(0., 1., shape[1])
a = DenseData(name='a', shape=shape)
a.data[:] = np.meshgrid(x, y)[1]
return a
def unit_box(name='a', shape=(11, 11)):
"""Create a field with value 0. to 1. in each dimension"""
a = DenseData(name=name, shape=shape)
dims = tuple([np.linspace(0., 1., d) for d in shape])
a.data[:] = np.meshgrid(*dims)[1]
return a