Skip to content

GitLab

Commit d36cce1c authored by Marius Isken's avatar Marius Isken
Browse files

adapting new state

parent fe362a65
Hide whitespace changes
Inline Side-by-side
Showing with 233 additions and 156 deletions
src/analysers/base.py
View file @ d36cce1c
import copy
import numpy as num
from pyrocko import util
from pyrocko.guts import Object, Int
from ..targets import TargetAnalysisResult
from ..meta import Forbidden
......@@ -12,8 +14,9 @@ class Analyser(object):
def __init__(self, niter):
self.niter = niter
self.pbar = util.progressbar('analysing problem', niter)
def analyse(self, problem, notifier):
def analyse(self, problem):
if self.niter == 0:
return
......@@ -36,10 +39,11 @@ class Analyser(object):
mss = num.zeros((self.niter, wproblem.ntargets))
rstate = num.random.RandomState(123)
notifier.emit('progress_start', 'analysing problem', self.niter)
isbad_mask = None
self.pbar.start()
for iiter in range(self.niter):
self.pbar.update(iiter)
while True:
x = []
for ipar in range(npar):
......@@ -61,9 +65,7 @@ class Analyser(object):
mss[iiter, :] = ms
isbad_mask = num.isnan(ms)
notifier.emit('progress_update', 'analysing problem', iiter)
notifier.emit('progress_finish', 'analysing problem')
self.pbar.finish()
mean_ms = num.mean(mss, axis=0)
weights = 1. / mean_ms
......
src/core.py
View file @ d36cce1c
......@@ -20,14 +20,27 @@ from .problems.base import ProblemConfig, Problem, \
from .optimizers.base import OptimizerConfig, BadProblem
from .targets.base import TargetGroup
from .analysers.base import AnalyserConfig
from .listeners import TerminalListener
from .meta import Path, HasPaths, expand_template, GrondError, Notifier, \
Forbidden
from .meta import Path, HasPaths, expand_template, GrondError, Forbidden
logger = logging.getLogger('grond.core')
guts_prefix = 'grond'
class RingBuffer(num.ndarray):
def __new__(cls, *args, **kwargs):
cls = num.ndarray.__new__(cls, *args, **kwargs)
cls.fill(0.)
return cls
def __init__(self, *args, **kwargs):
self.pos = 0
def put(self, value):
self[self.pos] = value
self.pos += 1
self.pos %= self.size
def mahalanobis_distance(xs, mx, cov):
imask = num.diag(cov) != 0.
icov = num.linalg.inv(cov[imask, :][:, imask])
......@@ -633,11 +646,8 @@ def process_event(ievent, g_data_id):
logger.info(
'start %i / %i' % (ievent+1, nevents))
notifier = Notifier()
notifier.add_listener(TerminalListener())
analyser = config.analyser_config.get_analyser()
analyser.analyse(problem, notifier=notifier)
analyser.analyse(problem)
basepath = config.get_basepath()
config.change_basepath(rundir)
......@@ -659,11 +669,19 @@ def process_event(ievent, g_data_id):
# update_every=10,
# movie_filename='grond_opt_time_magnitude.mp4')
def startThreads():
from .listeners import terminal
term = terminal.TerminalListener(rundir)
term.start()
return term
term = startThreads()
try:
optimizer = config.optimizer_config.get_optimizer()
if xs_inject is not None:
from .optimizers import highscore
if not isinstance(optimizer, highscore.HighScoreOptimizer()):
if not isinstance(optimizer, highscore.HighScoreOptimizer):
raise GrondError(
'optimizer does not support injections')
......@@ -678,6 +696,8 @@ def process_event(ievent, g_data_id):
except BadProblem as e:
logger.error(str(e))
finally:
term.join()
tstop = time.time()
logger.info(
......
src/listeners/base.py
View file @ d36cce1c
class Listener(object):
from threading import Thread
class Listener(Thread):
def __init__(self):
Thread.__init__(self)
def progress_start(self, name, niter):
raise NotImplementedError()
......
src/listeners/terminal.py
View file @ d36cce1c
import time
import logging
import numpy as num
from datetime import timedelta
from pyrocko import util
from grond.problems import ModelHistory
from .base import Listener
logger = logging.getLogger('TerminalListener')
class RingBuffer(num.ndarray):
def __new__(cls, *args, **kwargs):
cls = num.ndarray.__new__(cls, *args, **kwargs)
cls.fill(0.)
return cls
def __init__(self, *args, **kwargs):
self.pos = 0
def put(self, value):
self[self.pos] = value
self.pos += 1
self.pos %= self.size
class color:
PURPLE = '\033[95m'
CYAN = '\033[96m'
......@@ -21,58 +45,91 @@ class TerminalListener(Listener):
row_name = color.BOLD + '{:<{col_param_width}s}' + color.END
parameter_fmt = '{:{col_width}s}'
def __init__(self):
self.current_state = None
self.pbars = {}
def progress_start(self, name, niter):
pbar = util.progressbar('analysing problem', niter)
self.pbars[name] = pbar
pbar.start()
def __init__(self, rundir):
Listener.__init__(self)
self.rundir = rundir
self._iiter = 0
self.iter_per_second = 0
self._iter_buffer = RingBuffer(20)
def run(self):
logger.info('Waiting to follow %s' % self.rundir)
self.history = ModelHistory.follow(self.rundir)
self.problem = self.history.problem
self.niter = self.history.optimizer.niterations
self.starttime = time.time()
self.last_update = self.starttime
self.history.add_listener(self)
self.start_watch()
def start_watch(self):
while True:
self.history.update()
time.sleep(.1)
@property
def runtime(self):
return timedelta(seconds=time.time() - self.starttime)
@property
def iiter(self):
return self._iiter
@iiter.setter
def iiter(self, iiter):
dt = time.time() - self.last_update
self._iter_buffer.put(float((iiter - self.iiter) / dt))
self.iter_per_second = float(self._iter_buffer.mean())
self._iiter = iiter
self.last_update = time.time()
@property
def runtime_remaining(self):
if self.iter_per_second == 0.:
return timedelta()
return timedelta(seconds=(self.niter - self.iiter)
/ self.iter_per_second)
def extend(self, *args):
self.iiter = self.history.nmodels
problem = self.history.problem
def progress_update(self, name, iiter):
self.pbars[name].update(iiter)
def progress_finish(self, name):
self.pbars[name].finish()
del self.pbars[name]
def state(self, state):
lines = []
self.current_state = state
ladd = lines.append
def fmt(s):
return util.gform(s, significant_digits=(self.col_width-1-6)//2)
out_ln = self.row_name +\
''.join([self.parameter_fmt] * len(state.parameter_sets))
col_param_width = max([len(p) for p in state.parameter_names]) + 2
ladd('Problem name: {s.problem_name}'
'\t({s.runtime.seconds} - remaining {s.runtime_remaining}'
ladd('Problem name: {p.name}'
'\t({s.runtime} - remaining {s.runtime_remaining}'
' @ {s.iter_per_second:.1f} iter/s)'
.format(s=state))
.format(s=self, p=problem))
ladd('Iteration {s.iiter} / {s.niter}'
.format(s=state))
.format(s=self))
out_ln = self.row_name +\
''.join([self.parameter_fmt] * len(problem.parameter_names))
col_param_width = max([len(p) for p in problem.parameter_names]) + 2
ladd(out_ln.format(
*['Parameter'] + list(state.parameter_sets.keys()),
*['Parameter'] + list(problem.parameter_names),
col_param_width=col_param_width,
col_width=self.col_width,
type='s'))
for ip, parameter_name in enumerate(state.parameter_names):
ladd(out_ln.format(
parameter_name,
*[fmt(v[ip]) for v in state.parameter_sets.values()],
col_param_width=col_param_width,
col_width=self.col_width))
# for ip, parameter_name in enumerate(problem.parameter_names):
# ladd(out_ln.format(
# parameter_name,
# *[fmt(v[ip]) for v in problem.parameter_sets.values()],
# col_param_width=col_param_width,
# col_width=self.col_width))
ladd(state.extra_text.format(
col_param_width=col_param_width,
col_width=self.col_width,))
# ladd(problem.extra_text.format(
# col_param_width=col_param_width,
# col_width=self.col_width,))
lines[0:0] = ['\033[2J']
ladd('')
......
src/meta.py
View file @ d36cce1c
......@@ -184,23 +184,3 @@ class HasPaths(Object):
extra(
op.normpath(xjoin(self._basepath, xjoin(path_prefix, p))))
for p in path]
class Notifier(object):
def __init__(self):
self._listeners = []
def add_listener(self, listener):
self._listeners.append(listener)
def remove_listener(self, listener):
self._listeners.remove(listener)
def emit(self, signal_name, *args, **kwargs):
for listener in self._listeners:
if not hasattr(listener, signal_name):
logger.warn(
'signal name \'%s\' not implemented in listener %s'
% (signal_name, type(listener)))
continue
getattr(listener, signal_name)(*args, **kwargs)
src/optimizers/base.py
View file @ d36cce1c
import logging
import time
import numpy as num
from datetime import timedelta
from collections import OrderedDict
from pyrocko.guts import Object
from ..meta import GrondError
from grond.meta import GrondError
guts_prefix = 'grond'
......@@ -17,75 +12,15 @@ class BadProblem(GrondError):
pass
class SimpleTimedelta(timedelta):
def __str__(self):
return timedelta.__str__(self).split('.')[0]
class RingBuffer(num.ndarray):
def __new__(cls, *args, **kwargs):
cls = num.ndarray.__new__(cls, *args, **kwargs)
cls.fill(0.)
return cls
def __init__(self, *args, **kwargs):
self.pos = 0
def put(self, value):
self[self.pos] = value
self.pos += 1
self.pos %= self.size
class SolverState(object):
problem_name = ''
parameter_sets = OrderedDict()
parameter_names = []
starttime = time.time()
niter = 0
iter_per_second = 0.
extra_text = ''
_iiter = 0
_iter_buffer = RingBuffer(25)
_last_update = time.time()
@property
def iiter(self):
return self._iiter
@iiter.setter
def iiter(self, value):
dt = time.time() - self._last_update
self._iter_buffer.put(float((value - self._iiter) / dt))
self.iter_per_second = float(self._iter_buffer.mean())
self._iiter = value
self._last_update = time.time()
@property
def runtime(self):
return timedelta(seconds=time.time() - self.starttime)
@property
def runtime_remaining(self):
if self.iter_per_second == 0.:
return timedelta()
return timedelta(seconds=(self.niter - self.iiter)
/ self.iter_per_second)
@property
def nparameters(self):
return len(self.parameter_names)
class Optimizer(Object):
def optimize(self, problem):
raise NotImplemented()
@property
def niterations(self):
raise NotImplementedError()
class OptimizerConfig(Object):
pass
......
src/optimizers/highscore/optimizer.py
View file @ d36cce1c
......@@ -373,6 +373,10 @@ class HighScoreOptimizer(Optimizer):
iiter += 1
@property
def niterations(self):
return sum([ph.niterations for ph in self.sampler_phases])
class HighScoreOptimizerConfig(OptimizerConfig):
......
src/problems/base.py
View file @ d36cce1c
......@@ -4,6 +4,7 @@ import copy
import logging
import os.path as op
import os
import time
from pyrocko import gf, util, guts
from pyrocko.guts import Object, String, Bool, List, Dict, Int
......@@ -407,11 +408,24 @@ class Problem(Object):
return results
class ModelHistory(object):
class InvalidRundir(Exception):
pass
class ModelHistory(object):
nmodels_capacity_min = 1024
def __init__(self, problem, path=None, mode='r'):
'''Model History lets you write, read and follow new models
:param problem: The Problem
:type problem: :class:`grond.Problem`
:param path: Rundir to use, defaults to None
:type path: str, optional
:param mode: Mode to use, defaults to 'r'.
'r': Read, 'w': Write
:type mode: str, optional
'''
self.problem = problem
self.path = path
self._models_buffer = None
......@@ -421,11 +435,43 @@ class ModelHistory(object):
self.nmodels_capacity = self.nmodels_capacity_min
self.listeners = []
self.mode = mode
self.optimizer = load_optimizer_info(path)
if mode == 'r':
self.verify_rundir(self.path)
models, misfits = load_problem_data(path, problem)
self.extend(models, misfits)
@staticmethod
def verify_rundir(rundir):
_rundir_files = ['config.yaml', 'optimizer.yaml', 'misfits', 'models']
if not op.exists(rundir):
raise OSError('Rundir %s does not exist!' % rundir)
for f in _rundir_files:
if not op.exists(op.join(rundir, f)):
raise InvalidRundir('File %s not found!' % f)
@classmethod
def follow(cls, path, wait=20.):
'''Start following a rundir
:param path: The path to follow, a grond rundir
:type path: str, optional
:param wait: Wait time until the folder become alive, defaults to 10.
:type wait: number in seconds, optional
:returns: A ModelHistory instance
:rtype: :class:`grond.core.ModelHistory`
'''
start_watch = time.time()
while (time.time() - start_watch) < wait:
try:
cls.verify_rundir(path)
problem = load_problem_info(path)
return cls(problem, path, mode='r')
except InvalidRundir:
time.sleep(.25)
@property
def nmodels(self):
if self.models is None:
......@@ -473,7 +519,6 @@ class ModelHistory(object):
self.nmodels_capacity = self.nmodels_capacity_min
def extend(self, models, misfits):
nmodels = self.nmodels
n = models.shape[0]
......@@ -491,14 +536,13 @@ class ModelHistory(object):
self.misfits = self._misfits_buffer[:nmodels+n, :, :]
if self.path and self.mode == 'w':
for i in xrange(n):
for i in range(n):
self.problem.dump_problem_data(
self.path, models[i, :], misfits[i, :, :])
self.emit('extend', nmodels, n, models, misfits)
def append(self, model, misfits):
nmodels = self.nmodels
nmodels_capacity_want = max(
......@@ -521,18 +565,41 @@ class ModelHistory(object):
'extend', nmodels, 1,
model[num.newaxis, :], misfits[num.newaxis, :, :])
def update(self):
''' Update history from path '''
nmodels_available = get_nmodels(self.path, self.problem)
if self.nmodels == nmodels_available:
return
new_models, new_misfits = load_problem_data(
self.path, self.problem, skip_models=self.nmodels)
self.extend(new_models, new_misfits)
def add_listener(self, listener):
self.listeners.append(listener)
def emit(self, event_name, *args, **kwargs):
for listener in self.listeners:
getattr(listener, event_name)(*args, **kwargs)
slot = getattr(listener, event_name, None)
if callable(slot):
slot(*args, **kwargs)
def get_nmodels(dirname, problem):
fn = op.join(dirname, 'models')
with open(fn, 'r') as f:
return os.fstat(f.fileno()).st_size // (problem.nparameters * 8)
def load_problem_info_and_data(dirname, subset=None):
problem = load_problem_info(dirname)
xs, misfits = load_problem_data(xjoin(dirname, subset), problem)
return problem, xs, misfits
models, misfits = load_problem_data(xjoin(dirname, subset), problem)
return problem, models, misfits
def load_optimizer_info(dirname):
fn = op.join(dirname, 'optimizer.yaml')
return guts.load(filename=fn)
def load_problem_info(dirname):
......@@ -540,29 +607,35 @@ def load_problem_info(dirname):
return guts.load(filename=fn)
def load_problem_data(dirname, problem, skip_models=0):
def load_problem_data(dirname, problem, skip_models=None):
if skip_models is None:
skip_models = 0
nmodels = get_nmodels(dirname, problem)
fn = op.join(dirname, 'models')
with open(fn, 'r') as f:
nmodels = os.fstat(f.fileno()).st_size // (problem.nparameters * 8)
nmodels -= skip_models
f.seek(skip_models * problem.nparameters * 8)
data1 = num.fromfile(
f, dtype='<f8',
count=nmodels * problem.nparameters)\
models = num.fromfile(
f, dtype='<f8',
count=nmodels * problem.nparameters)\
.astype(num.float)
nmodels = data1.size // problem.nparameters - skip_models
xs = data1.reshape((nmodels, problem.nparameters))
nmodels = models.size // problem.nparameters
models = models.reshape((nmodels, problem.nparameters))
fn = op.join(dirname, 'misfits')
with open(fn, 'r') as f:
f.seek(skip_models * problem.ntargets * 2 * 8)
misfits = num.fromfile(
f, dtype='<f8', count=nmodels*problem.ntargets*2).astype(num.float)
f, dtype='<f8',
count=nmodels*problem.ntargets*2)\
.astype(num.float)
misfits = misfits.reshape((nmodels, problem.ntargets, 2))
return xs, misfits
return models, misfits
__all__ = '''
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment