def __init__(self, shots, shape, origin, spacing, m0, dm, noise=0.0, device='cpu', sequential=False):
super(wave_solver, self).__init__()
self.forward_born = ForwardBorn()
self.noise = noise
self.device = device
self.dm = dm
self.spacing = spacing
epsilon = np.sqrt(noise)*np.random.randn(shots.shape[0], shots.shape[1], shots.shape[2])
self.shots = shots + epsilon
self.model0 = Model(shape=shape, origin=origin, spacing=spacing, vp=1/np.sqrt(m0),
nbpml=40)
self.T = self.mute_top(dm, mute_end=10, length=5)
t0 = 0.
tn = 1500.0
dt = self.model0.critical_dt
nt = int(1 + (tn-t0) / dt)
self.time_range = np.linspace(t0,tn,nt)
self.f0 = 0.030
self.nsrc = 205
self.nsimsrc = 205
self.src = RickerSource(name='src', grid=self.model0.grid, f0=self.f0, time=self.time_range,
npoint=self.nsimsrc)
self.src.coordinates.data[:,0] = np.linspace(0, self.model0.domain_size[0], num=self.nsimsrc)
self.src.coordinates.data[:,-1] = 2.0*spacing[1]
nrec = 410
self.rec = Receiver(name='rec', grid=self.model0.grid, npoint=nrec, ntime=nt)
self.rec.coordinates.data[:, 0] = np.linspace(0, self.model0.domain_size[0], num=nrec)
self.rec.coordinates.data[:, 1] = 2.0*spacing[1]
self.seq_src_idx = 0
self.sequential = sequential
if iteration == 1:
x = np.zeros(shape=shape, dtype='float32')
else:
x = array_get(
bucket,
variable_path + 'chunk_1/' + variable_name + str(iteration - 1))
if num_chunks > 1:
for chunk in range(1, num_chunks):
x_chunk = array_get(
bucket, variable_path + 'chunk_' + str(chunk + 1) + '/' +
variable_name + str(iteration - 1))
x = np.concatenate((x, x_chunk), axis=0)
x = x.reshape(shape[0], shape[1], order='F')
# Set up model structures
model = Model(shape=shape, origin=origin, spacing=spacing, vp=np.sqrt(1 / m0))
# Time axis
t0 = 0.
dt_comp = model.critical_dt
nt_comp = int(1 + (tn - t0) / dt_comp) + 1
time_comp = np.linspace(t0, tn, nt_comp)
# Source
f0 = 0.020
src_coordinates = np.empty((1, ndims))
src_coordinates[0, 0] = sx
src_coordinates[0, 1] = sz
# Receiver for predicted data
nrec = len(gx)
if np.isscalar(vel):
return .9 * vel * np.ones(shape, dtype=np.float32)
out = np.copy(vel)
nz = shape[-1]
for a in range(5, nz - 6):
if len(shape) == 2:
out[:, a] = np.sum(vel[:, a - 5:a + 5], axis=1) / 10
else:
out[:, :, a] = np.sum(vel[:, :, a - 5:a + 5], axis=2) / 10
return out
# Set up model structures
rho0 = smooth10(rho, shape)
model = Model(shape=shape, origin=origin, spacing=spacing, vp=v, rho=rho0)
# Smooth background model
v0 = smooth10(v, shape)
dm = (1 / v)**2 - (1 / v0)**2
model0 = Model(shape=shape,
origin=origin,
spacing=spacing,
vp=v0,
dm=dm,
rho=rho0)
# Constant background model
v_const = np.empty(shape, dtype=np.float32)
v_const[:, :] = 1.5
dm_const = (1 / v)**2 - (1 / v_const)**2
from PyModel import Model from checkpoint import DevitoCheckpoint, CheckpointOperator from pyrevolve import Revolver import matplotlib.pyplot as plt from JAcoustic_codegen import forward_modeling, adjoint_born # Model shape = (101, 101) spacing = (10., 10.) origin = (0., 0.) v = np.empty(shape, dtype=np.float32) v[:, :51] = 1.5 v[:, 51:] = 2.5 v0 = np.empty(shape, dtype=np.float32) v0[:, :] = 1.5 model = Model(shape=shape, origin=origin, spacing=spacing, vp=v) model0 = Model(shape=shape, origin=origin, spacing=spacing, vp=v0) # Time axis t0 = 0. tn = 1000. dt = model.critical_dt nt = int(1 + (tn - t0) / dt) time = np.linspace(t0, tn, nt) # Source f0 = 0.010 src = RickerSource(name='src', grid=model.grid, f0=f0, time=time) src.coordinates.data[0, :] = np.array(model.domain_size) * 0.5 src.coordinates.data[0, -1] = 20.
shape = (120, 120, 120)
spacing = (10, 10, 10)
origin = (0., 0., 0.)
nrec = 101
# Velocity
v = np.empty(shape, dtype=np.float32)
v[:, :, :55] = 1.5
v[:, :, 55:] = 3.0
v0 = ndimage.gaussian_filter(v, sigma=5)
m = (1. / v)**2
m0 = (1. / v0)**2
dm = m - m0
# Set up model structures
model = Model(shape=shape, origin=origin, spacing=spacing, vp=v, nbpml=30)
model0 = Model(shape=shape,
origin=origin,
spacing=spacing,
vp=v0,
dm=dm,
nbpml=30)
# Time axis
t0 = 0.
tn = 1200.
num_wavefields = int(tn / 4)
mem = ((model.shape[0] + 2 * model.nbpml) *
(model.shape[1] + 2 * model.nbpml) *
(model.shape[2] + 2 * model.nbpml) * num_wavefields /
subsampling_factor * 8) / 1024**3
# Velocity v = np.empty(shape, dtype=np.float32) v[:, :51] = 1.5 v[:, 51:] = 1.5 v0 = np.empty(shape, dtype=np.float32) v0[:, :] = 1.5 # Density rho = np.empty(shape, dtype=np.float32) rho[:, :51] = 1.0 rho[:, 51:] = 2.0 rho0 = np.empty(shape, dtype=np.float32) rho0[:, :] = 1.0 # Set up model structures model = Model(shape=shape, origin=origin, spacing=spacing, vp=v, rho=rho) model0 = Model(shape=shape, origin=origin, spacing=spacing, vp=v0, rho=rho0) # Time axis t0 = 0. tn = 1000. dt = model.critical_dt nt = int(1 + (tn - t0) / dt) time = np.linspace(t0, tn, nt) # Source f0 = 0.010 src = RickerSource(name='src', grid=model.grid, f0=f0, time=time) src.coordinates.data[0, :] = np.array(model.domain_size) * 0.5 src.coordinates.data[0, -1] = 20.
# Model
shape = (301, 301)
spacing = (10., 10.)
origin = (0., 0.)
v1 = np.empty(shape, dtype=np.float32)
v1[:, :51] = 1.5
v1[:, 51:] = 3.5
# Density
rho = np.empty(shape, dtype=np.float32)
rho[:, :51] = 1.0
rho[:, 51:] = 2.0
# Set up model structures
model1 = Model(shape=shape, origin=origin, spacing=spacing, vp=v1, rho=rho)
def smooth10(vel, shape):
if np.isscalar(vel):
return .9 * vel * np.ones(shape, dtype=np.float32)
out = np.copy(vel)
nz = shape[-1]
for a in range(5, nz - 6):
if len(shape) == 2:
out[:, a] = np.sum(vel[:, a - 5:a + 5], axis=1) / 10
else:
out[:, :, a] = np.sum(vel[:, :, a - 5:a + 5], axis=2) / 10
return out