def _run_set_of_spectra(self, index_start: int, index_stop: int) -> None:
"""Internal function to run a chunk of spectra
Args:
index_start: spectral index to start execution at
index_stop: spectral index to stop execution at
"""
logging.basicConfig(format='%(levelname)s:%(message)s', level=self.loglevel, filename=self.logfile)
self._init_nonpicklable_objects()
io = IO(self.config, self.fm, self.iv, self.rows, self.cols)
for index in range(index_start, index_stop):
success, row, col, meas, geom = io.get_components_at_index(
index)
# Only run through the inversion if we got some data
if success:
if meas is not None and all(meas < -49.0):
# Bad data flags
self.states = []
else:
# The inversion returns a list of states, which are
# intepreted either as samples from the posterior (MCMC case)
# or as a gradient descent trajectory (standard case). For
# a trajectory, the last spectrum is the converged solution.
self.states = self.iv.invert(meas, geom)
# Write the spectra to disk
io.write_spectrum(row, col, self.states, meas,
geom, flush_immediately=True)
if (index - index_start) % 100 == 0:
logging.info(
'Core at start index {} completed inversion {}/{}'.format(index_start, index-index_start,
index_stop-index_start))
def __init__(self, config: configs.Config, loglevel: str, logfile: str, worker_id: int = None, total_workers: int = None):
"""
Worker class to help run a subset of spectra.
Args:
config: isofit configuration
loglevel: output logging level
logfile: output logging file
worker_id: worker ID for logging reference
total_workers: the total number of workers running, for logging reference
"""
logging.basicConfig(format='%(levelname)s:%(message)s', level=loglevel, filename=logfile)
self.config = config
self.fm = ForwardModel(self.config)
if self.config.implementation.mode == 'mcmc_inversion':
self.iv = MCMCInversion(self.config, self.fm)
elif self.config.implementation.mode in ['inversion', 'simulation']:
self.iv = Inversion(self.config, self.fm)
else:
# This should never be reached due to configuration checking
raise AttributeError('Config implementation mode node valid')
self.io = IO(self.config, self.fm)
self.approximate_total_spectra = None
if total_workers is not None:
self.approximate_total_spectra = self.io.n_cols * self.io.n_rows / total_workers
self.worker_id = worker_id
self.completed_spectra = 0
def __init__(self, config_file, row_column='', level='INFO', logfile=None):
# Explicitly set the number of threads to be 1, so we more effectively
#run in parallel
os.environ["MKL_NUM_THREADS"] = "1"
# Set logging level
self.loglevel = level
self.logfile = logfile
logging.basicConfig(format='%(levelname)s:%(message)s', level=self.loglevel, filename=self.logfile)
self.rows = None
self.cols = None
self.config = None
self.fm = None
self.iv = None
self.io = None
self.states = None
# Load configuration file
self.config = configs.create_new_config(config_file)
self.config.get_config_errors()
# Initialize ray for parallel execution
rayargs = {'address': self.config.implementation.ip_head,
'redis_password': self.config.implementation.redis_password,
'ignore_reinit_error':True,
'local_mode': self.config.implementation.n_cores == 1}
# only specify a temporary directory if we are not connecting to
# a ray cluster
if rayargs['local_mode']:
rayargs['temp_dir'] = self.config.implementation.ray_temp_dir
# Used to run on a VPN
ray.services.get_node_ip_address = lambda: '127.0.0.1'
# We can only set the num_cpus if running on a single-node
if self.config.implementation.ip_head is None and self.config.implementation.redis_password is None:
rayargs['num_cpus'] = self.config.implementation.n_cores
ray.init(**rayargs)
if len(row_column) > 0:
ranges = row_column.split(',')
if len(ranges) == 1:
self.rows, self.cols = [int(ranges[0])], None
if len(ranges) == 2:
row_start, row_end = ranges
self.rows, self.cols = range(
int(row_start), int(row_end)), None
elif len(ranges) == 4:
row_start, row_end, col_start, col_end = ranges
line_start, line_end, samp_start, samp_end = ranges
self.rows = range(int(row_start), int(row_end))
self.cols = range(int(col_start), int(col_end))
# Build the forward model and inversion objects
self._init_nonpicklable_objects()
self.io = IO(self.config, self.fm, self.iv, self.rows, self.cols)
def __init__(self, config: configs.Config, loglevel: str, logfile: str):
logging.basicConfig(format='%(levelname)s:%(message)s', level=loglevel, filename=logfile)
self.config = config
self.fm = ForwardModel(self.config)
if self.config.implementation.mode == 'mcmc_inversion':
self.iv = MCMCInversion(self.config, self.fm)
elif self.config.implementation.mode in ['inversion', 'simulation']:
self.iv = Inversion(self.config, self.fm)
else:
# This should never be reached due to configuration checking
raise AttributeError('Config implementation mode node valid')
self.io = IO(self.config, self.fm)
def run_forward():
"""Simulate the remote measurement of a spectrally uniform surface."""
# Configure the surface/atmosphere/instrument model
config = load_config('config_forward.json')
fm = ForwardModel(config['forward_model'])
iv = Inversion(config['inversion'], fm)
io = IO(config, fm, iv, [0], [0])
# Simulate a measurement and write result
for row, col, meas, geom, configs in io:
states = iv.invert(meas, geom)
io.write_spectrum(row, col, states, meas, geom)
assert True
return states[0]
def run_forward():
"""Simulate the remote measurement of a spectrally uniform surface."""
# Configure the surface/atmosphere/instrument model
testdir, fname = os.path.split(os.path.abspath(__file__))
datadir = os.path.join(testdir, 'data')
config = create_new_config(os.path.join(datadir, 'config_forward.json'))
fm = ForwardModel(config)
iv = Inversion(config, fm)
io = IO(config, fm, iv, [0], [0])
# Simulate a measurement and write result
for row, col, meas, geom, configs in io:
states = iv.invert(meas, geom)
io.write_spectrum(row, col, states, meas, geom)
assert True
return states[0]
def run_inverse():
"""Invert the remote measurement."""
# Configure the surface/atmosphere/instrument model
config = load_config('config_inversion.json')
fm = ForwardModel(config['forward_model'])
iv = Inversion(config['inversion'], fm)
io = IO(config, fm, iv, [0], [0])
geom = None
# Get our measurement from the simulation results, and invert.
# Calculate uncertainties at the solution state, write result
for row, col, meas, geom, configs in io:
states = iv.invert(meas, geom)
io.write_spectrum(row, col, states, meas, geom)
assert True
return states[-1]
def run_inverse():
"""Invert the remote measurement."""
# Configure the surface/atmosphere/instrument model
testdir, fname = os.path.split(os.path.abspath(__file__))
datadir = os.path.join(testdir, 'data')
config = create_new_config(os.path.join(datadir, 'config_forward.json'))
fm = ForwardModel(config)
iv = Inversion(config, fm)
io = IO(config, fm, iv, [0], [0])
geom = None
# Get our measurement from the simulation results, and invert.
# Calculate uncertainties at the solution state, write result
for row, col, meas, geom, configs in io:
states = iv.invert(meas, geom)
io.write_spectrum(row, col, states, meas, geom)
assert True
return states[-1]
def run_forward():
"""Simulate the remote measurement of a spectrally uniform surface."""
# Configure the surface/atmosphere/instrument model
testdir, fname = os.path.split(os.path.abspath(__file__))
datadir = os.path.join(testdir, 'data')
config = create_new_config(os.path.join(datadir, 'config_forward.json'))
fm = ForwardModel(config)
iv = Inversion(config, fm)
io = IO(config, fm)
# Simulate a measurement and write result
for row in range(io.n_rows):
for col in range(io.n_cols):
id = io.get_components_at_index(row, col)
if id is not None:
states = iv.invert(id.meas, id.geom)
io.write_spectrum(row, col, states, fm, iv)
assert True
return states[0]
def run_inverse():
"""Invert the remote measurement."""
# Configure the surface/atmosphere/instrument model
testdir, fname = os.path.split(os.path.abspath(__file__))
datadir = os.path.join(testdir, 'data')
config = create_new_config(os.path.join(datadir, 'config_forward.json'))
fm = ForwardModel(config)
iv = Inversion(config, fm)
io = IO(config, fm)
# Get our measurement from the simulation results, and invert.
# Calculate uncertainties at the solution state, write result
for row in range(io.n_rows):
for col in range(io.n_cols):
id = io.get_components_at_index(row, col)
if id is not None:
states = iv.invert(id.meas, id.geom)
io.write_spectrum(row, col, states, fm, iv)
assert True
return states[-1]
def run(self, row_column = None):
"""
Iterate over spectra, reading and writing through the IO
object to handle formatting, buffering, and deferred write-to-file.
Attempts to avoid reading the entire file into memory, or hitting
the physical disk too often.
row_column: The user can specify
* a single number, in which case it is interpreted as a row
* a comma-separated pair, in which case it is interpreted as a
row/column tuple (i.e. a single spectrum)
* a comma-separated quartet, in which case it is interpreted as
a row, column range in the order (line_start, line_end, sample_start,
sample_end) all values are inclusive.
If none of the above, the whole cube will be analyzed.
"""
logging.info("Building first forward model, will generate any necessary LUTs")
fm = ForwardModel(self.config)
if row_column is not None:
ranges = row_column.split(',')
if len(ranges) == 1:
self.rows, self.cols = [int(ranges[0])], None
if len(ranges) == 2:
row_start, row_end = ranges
self.rows, self.cols = range(
int(row_start), int(row_end)), None
elif len(ranges) == 4:
row_start, row_end, col_start, col_end = ranges
self.rows = range(int(row_start), int(row_end) + 1)
self.cols = range(int(col_start), int(col_end) + 1)
else:
io = IO(self.config, fm)
self.rows = range(io.n_rows)
self.cols = range(io.n_cols)
del io, fm
index_pairs = np.vstack([x.flatten(order='f') for x in np.meshgrid(self.rows, self.cols)]).T
n_iter = index_pairs.shape[0]
if self.config.implementation.n_cores is None:
n_workers = multiprocessing.cpu_count()
else:
n_workers = self.config.implementation.n_cores
# Max out the number of workers based on the number of tasks
n_workers = min(n_workers, n_iter)
if self.workers is None:
remote_worker = ray.remote(Worker)
self.workers = ray.util.ActorPool([remote_worker.remote(self.config, self.loglevel, self.logfile, n, n_workers)
for n in range(n_workers)])
start_time = time.time()
n_tasks = min(n_workers * self.config.implementation.task_inflation_factor, n_iter)
logging.info(f'Beginning {n_iter} inversions in {n_tasks} chunks using {n_workers} cores')
# Divide up spectra to run into chunks
index_sets = np.linspace(0, n_iter, num=n_tasks, dtype=int)
if len(index_sets) == 1:
indices_to_run = [index_pairs[0:1,:]]
else:
indices_to_run = [index_pairs[index_sets[l]:index_sets[l + 1], :]
for l in range(len(index_sets) - 1)]
res = list(self.workers.map_unordered(lambda a, b: a.run_set_of_spectra.remote(b),
indices_to_run))
total_time = time.time() - start_time
logging.info(f'Inversions complete. {round(total_time,2)}s total, {round(n_iter/total_time,4)} spectra/s, '
f'{round(n_iter/total_time/n_workers,4)} spectra/s/core')