def _terminate_pool(cls,
taskqueue,
inqueue,
outqueue,
pool,
task_handler,
result_handler,
cache,
working_dirs,
):
info("terminating pool")
# this is guaranteed to only be called once
debug('finalizing pool')
TERMINATE = 2
task_handler._state = TERMINATE
for p in pool:
taskqueue.put(None) # sentinel
time.sleep(1)
debug('helping task handler/workers to finish')
cls._help_stuff_finish(inqueue, task_handler, len(pool))
assert result_handler.is_alive() or len(cache) == 0
result_handler._state = TERMINATE
outqueue.put(None) # sentinel
if pool and hasattr(pool[0], 'terminate'):
debug('terminating workers')
for p in pool:
p.terminate()
debug('joining task handler')
task_handler.join(1e100)
debug('joining result handler')
result_handler.join(1e100)
if pool and hasattr(pool[0], 'terminate'):
debug('joining pool workers')
for p in pool:
p.join()
# cleaning up directories
# TODO investigate whether the multiprocessing.util tempdirectory
# functionality can be used instead
for directory in working_dirs:
directory = os.path.dirname(directory)
debug("deleting "+str(directory))
shutil.rmtree(directory)
def oldcsv_load_results(file_name):
'''
load the specified bz2 file. the file is assumed to be saves
using save_results.
:param file_name: the path of the file
:raises: IOError if file not found
'''
outcomes = {}
with tarfile.open(file_name, 'r') as z:
# load experiments
experiments = z.extractfile('experiments.csv')
experiments = csv2rec(experiments)
# load metadata
metadata = z.extractfile('experiments metadata.csv').readlines()
metadata = [entry.strip() for entry in metadata]
metadata = [tuple(entry.split(",")) for entry in metadata]
metadata = np.dtype(metadata)
# cast experiments to dtype and name specified in metadata
temp_experiments = np.zeros((experiments.shape[0],), dtype=metadata)
for i, entry in enumerate(experiments.dtype.descr):
dtype = metadata[i]
name = metadata.descr[i][0]
temp_experiments[name][:] = experiments[entry[0]].astype(dtype)
experiments = temp_experiments
# load outcomes
fhs = z.getnames()
fhs.remove('experiments.csv')
fhs.remove('experiments metadata.csv')
for fh in fhs:
root = os.path.splitext(fh)[0]
data = z.extractfile(fh)
first_line = data.readline()
shape = first_line.split(":")[1].strip()[1:-1]
shape = shape.split((','))
shape = tuple([int(entry) for entry in shape if len(entry)>0])
data = np.loadtxt(data, delimiter=',')
data = data.reshape(shape)
outcomes[root] = data
info("results loaded succesfully from {}".format(file_name))
return experiments, outcomes
def __call__(self, case, policy, name, result):
'''
Method responsible for storing results. The implementation in this
class only keeps track of how many runs have been completed and
logging this.
:param case: the case to be stored
:param policy: the name of the policy being used
:param name: the name of the model being used
:param result: the result dict
'''
self.i+=1
debug(str(self.i)+" cases completed")
if self.i % self.reporting_interval == 0:
info(str(self.i)+" cases completed")
def __init__(self, msis, processes=None, kwargs=None):
"""
:param msis: an iterable of model structure interfaces
:param processes: nr. of processes to spawn, if none, it is
set to equal the nr. of cores
:param callback: callback function for handling the output
:param kwargs: kwargs to be pased to :meth:`model_init`
"""
self._setup_queues()
self._taskqueue = Queue(cpu_count() * 2)
self._cache = {}
self._state = RUN
if processes is None:
try:
processes = cpu_count()
except NotImplementedError:
processes = 1
info("nr of processes is " + str(processes))
self.log_queue = Queue()
h = NullHandler()
logging.getLogger(ema_logging.LOGGER_NAME).addHandler(h)
# This thread will read from the subprocesses and write to the
# main log's handlers.
log_queue_reader = LogQueueReader(self.log_queue)
log_queue_reader.start()
self._pool = []
working_dirs = []
# msis = [copy.deepcopy(msi) for msi in msis]
debug("generating workers")
workerRoot = None
for i in range(processes):
debug("generating worker " + str(i))
workerName = "PoolWorker" + str(i)
def ignore_function(path, names):
if path.find(".svn") != -1:
return names
else:
return []
# setup working directories for parallelEMA
for msi in msis:
if msi.workingDirectory != None:
if workerRoot == None:
workerRoot = os.path.dirname(os.path.abspath(msis[0].workingDirectory))
workingDirectory = os.path.join(workerRoot, workerName, msi.name)
working_dirs.append(workingDirectory)
shutil.copytree(msi.workingDirectory, workingDirectory, ignore=ignore_function)
msi.set_working_directory(workingDirectory)
w = LoggingProcess(
self.log_queue,
level=logging.getLogger(ema_logging.LOGGER_NAME).getEffectiveLevel(),
target=worker,
args=(self._inqueue, self._outqueue, msis, kwargs),
)
self._pool.append(w)
w.name = w.name.replace("Process", workerName)
w.daemon = True
w.start()
debug(" worker " + str(i) + " generated")
# thread for handling tasks
self._task_handler = threading.Thread(
target=CalculatorPool._handle_tasks,
name="task handler",
args=(self._taskqueue, self._quick_put, self._outqueue, self._pool),
)
self._task_handler.daemon = True
self._task_handler._state = RUN
self._task_handler.start()
# thread for handling results
self._result_handler = threading.Thread(
target=CalculatorPool._handle_results,
name="result handler",
args=(self._outqueue, self._quick_get, self._cache, self.log_queue),
)
self._result_handler.daemon = True
self._result_handler._state = RUN
self._result_handler.start()
# function for cleaning up when finalizing object
self._terminate = Finalize(
self,
self._terminate_pool,
args=(
self._taskqueue,
self._inqueue,
self._outqueue,
self._pool,
self._task_handler,
self._result_handler,
self._cache,
working_dirs,
),
exitpriority=15,
)
info("pool has been set up")
def merge_results(results1, results2, downsample=None):
'''
convenience function for merging the return from
:meth:`~modelEnsemble.ModelEnsemble.perform_experiments`.
The function merges results2 with results1. For the experiments,
it generates an empty array equal to the size of the sum of the
experiments. As dtype is uses the dtype from the experiments in results1.
The function assumes that the ordering of dtypes and names is identical in
both results.
A typical use case for this function is in combination with
:func:`~util.experiments_to_cases`. Using :func:`~util.experiments_to_cases`
one extracts the cases from a first set of experiments. One then
performs these cases on a different model or policy, and then one wants to
merge these new results with the old result for further analysis.
:param results1: first results to be merged
:param results2: second results to be merged
:param downsample: should be an integer, will be used in slicing the results
in order to avoid memory problems.
:return: the merged results
'''
#start of merging
old_exp, old_res = results1
new_exp, new_res = results2
#merge experiments
dtypes = old_exp.dtype
merged_exp = np.empty((old_exp.shape[0]+new_exp.shape[0],),dtype= dtypes)
merged_exp[0:old_exp.shape[0]] = old_exp
merged_exp[old_exp.shape[0]::] = new_exp
#only merge the results that are in both
keys = old_res.keys()
[keys.append(key) for key in new_res.keys()]
keys = set(keys)
info("intersection of keys: %s" % keys)
#merging results
merged_res = {}
for key in keys:
info("merge "+key)
old_value = old_res.get(key)
new_value = new_res.get(key)
i = old_value.shape[0]+new_value.shape[0]
j = old_value.shape[1]
slice_value = 1
if downsample:
j = int(math.ceil(j/downsample))
slice_value = downsample
merged_value = np.empty((i,j))
debug("merged shape: %s" % merged_value.shape)
merged_value[0:old_value.shape[0], :] = old_value[:, ::slice_value]
merged_value[old_value.shape[0]::, :] = new_value[:, ::slice_value]
merged_res[key] = merged_value
mr = (merged_exp, merged_res)
return mr
def load_results(file_name):
'''
load the specified bz2 file. the file is assumed to be saves
using save_results.
:param file_name: the path of the file
:raises: IOError if file not found
'''
outcomes = {}
with tarfile.open(file_name, 'r') as z:
# load experiments
experiments = z.extractfile('experiments.csv')
experiments = csv2rec(experiments)
# load experiment metadata
metadata = z.extractfile('experiments metadata.csv').readlines()
metadata = [entry.strip() for entry in metadata]
metadata = [tuple(entry.split(",")) for entry in metadata]
metadata = np.dtype(metadata)
# cast experiments to dtype and name specified in metadata
temp_experiments = np.zeros((experiments.shape[0],), dtype=metadata)
for i, entry in enumerate(experiments.dtype.descr):
dtype = metadata[i]
name = metadata.descr[i][0]
temp_experiments[name][:] = experiments[entry[0]].astype(dtype)
experiments = temp_experiments
# load outcome metadata
metadata = z.extractfile('outcomes metadata.csv').readlines()
metadata = [entry.strip() for entry in metadata]
metadata = [tuple(entry.split(",")) for entry in metadata]
metadata = {entry[0]: entry[1:] for entry in metadata}
# load outcomes
for outcome, shape in metadata.iteritems():
shape = list(shape)
shape[0] = shape[0][1:]
shape[-1] = shape[-1][0:-1]
temp_shape = []
for entry in shape:
if entry:
temp_shape.append(int(entry))
shape = tuple(temp_shape)
if len(shape)>2:
nr_files = shape[-1]
data = np.empty(shape)
for i in range(nr_files):
values = z.extractfile("{}_{}.csv".format(outcome, i))
values = read_csv(values, index_col=False, header=None).values
data[:,:,i] = values
else:
data = z.extractfile("{}.csv".format(outcome))
data = read_csv(data, index_col=False, header=None).values
data = np.reshape(data, shape)
outcomes[outcome] = data
info("results loaded succesfully from {}".format(file_name))
return experiments, outcomes
def save_results(results, file_name):
'''
save the results to the specified tar.gz file. The results are stored as
csv files. There is an experiments.csv, and a csv for each outcome. In
addition, there is a metadata csv which contains the datatype information
for each of the columns in the experiments array.
:param results: the return of run_experiments
:param file_name: the path of the file
:raises: IOError if file not found
'''
def add_file(tararchive, string_to_add, filename):
tarinfo = tarfile.TarInfo(filename)
tarinfo.size = len(string_to_add)
z.addfile(tarinfo, StringIO.StringIO(string_to_add))
def save_numpy_array(fh, data):
data = pd.DataFrame(data)
data.to_csv(fh, header=False, index=False)
experiments, outcomes = results
with tarfile.open(file_name, 'w:gz') as z:
# write the experiments to the zipfile
experiments_file = StringIO.StringIO()
rec2csv(experiments, experiments_file, withheader=True)
add_file(z, experiments_file.getvalue(), 'experiments.csv')
# write experiment metadata
dtype = experiments.dtype.descr
dtype = ["{},{}".format(*entry) for entry in dtype]
dtype = "\n".join(dtype)
add_file(z, dtype, 'experiments metadata.csv')
# write outcome metadata
outcome_names = outcomes.keys()
outcome_meta = ["{},{}".format(outcome, outcomes[outcome].shape)
for outcome in outcome_names]
outcome_meta = "\n".join(outcome_meta)
add_file(z, outcome_meta, "outcomes metadata.csv")
# outcomes
for key, value in outcomes.iteritems():
fh = StringIO.StringIO()
nr_dim = len(value.shape)
if nr_dim==3:
for i in range(value.shape[2]):
data = value[:,:,i]
save_numpy_array(fh, data)
fh = fh.getvalue()
fn = '{}_{}.csv'.format(key, i)
add_file(z, fh, fn)
fh = StringIO.StringIO()
else:
save_numpy_array(fh, value)
fh = fh.getvalue()
add_file(z, fh, '{}.csv'.format(key))
info("results saved successfully to {}".format(file_name))
def __init__(self,
msis,
processes=None,
kwargs=None):
'''
:param msis: an iterable of model structure interfaces
:param processes: nr. of processes to spawn, if none, it is
set to equal the nr. of cores
:param callback: callback function for handling the output
:param kwargs: kwargs to be pased to :meth:`model_init`
'''
self._setup_queues()
self._taskqueue = Queue.Queue(cpu_count()*2)
self._cache = {}
self._state = RUN
if processes is None:
try:
processes = cpu_count()
except NotImplementedError:
processes = 1
info("nr of processes is "+str(processes))
self.log_queue = multiprocessing.Queue()
h = NullHandler()
logging.getLogger(ema_logging.LOGGER_NAME).addHandler(h)
# This thread will read from the subprocesses and write to the
# main log's handlers.
log_queue_reader = LogQueueReader(self.log_queue)
log_queue_reader.start()
self._pool = []
working_dirs = []
debug('generating workers')
worker_root = None
for i in range(processes):
debug('generating worker '+str(i))
# generate a random string helps in running repeatedly with
# crashes
choice_set = string.ascii_uppercase + string.digits + string.ascii_lowercase
random_string = ''.join(random.choice(choice_set) for _ in range(5))
workername = 'tpm_{}_PoolWorker_{}'.format(random_string, i)
#setup working directories for parallel_ema
for msi in msis:
if msi.working_directory != None:
if worker_root == None:
worker_root = os.path.dirname(os.path.abspath(msis[0].working_directory))
working_directory = os.path.join(worker_root, workername)
# working_directory = tempfile.mkdtemp(suffix=workername,
# prefix='tmp_',
# dir=worker_root)
working_dirs.append(working_directory)
shutil.copytree(msi.working_directory,
working_directory,
)
msi.set_working_directory(working_directory)
w = LoggingProcess(
self.log_queue,
level = logging.getLogger(ema_logging.LOGGER_NAME)\
.getEffectiveLevel(),
target=worker,
args=(self._inqueue,
self._outqueue,
msis,
kwargs
)
)
self._pool.append(w)
w.name = w.name.replace('Process', workername)
w.daemon = True
w.start()
debug(' worker '+str(i) + ' generated')
# thread for handling tasks
self._task_handler = threading.Thread(
target=CalculatorPool._handle_tasks,
name='task handler',
args=(self._taskqueue,
self._quick_put,
self._outqueue,
self._pool
)
)
self._task_handler.daemon = True
self._task_handler._state = RUN
self._task_handler.start()
# thread for handling results
self._result_handler = threading.Thread(
target=CalculatorPool._handle_results,
name='result handler',
args=(self._outqueue,
self._quick_get,
self._cache,
self.log_queue)
)
self._result_handler.daemon = True
self._result_handler._state = RUN
self._result_handler.start()
# function for cleaning up when finalizing object
self._terminate = Finalize(self,
self._terminate_pool,
args=(self._taskqueue,
self._inqueue,
self._outqueue,
self._pool,
self._task_handler,
self._result_handler,
self._cache,
working_dirs,
),
exitpriority=15
)
info("pool has been set up")
def __init__(self,
msis,
processes=None,
kwargs=None):
'''
:param msis: an iterable of model structure interfaces
:param processes: nr. of processes to spawn, if none, it is
set to equal the nr. of cores
:param callback: callback function for handling the output
:param kwargs: kwargs to be pased to :meth:`model_init`
'''
if processes is None:
try:
processes = cpu_count()
except NotImplementedError:
processes = 1
info("nr of processes is "+str(processes))
# setup queues etc.
self._setup_queues()
self._taskqueue = Queue.Queue(processes*2)
self._cache = {}
self._state = RUN
# handling of logging
self.log_queue = multiprocessing.Queue()
h = NullHandler()
logging.getLogger(ema_logging.LOGGER_NAME).addHandler(h)
log_queue_reader = LogQueueReader(self.log_queue)
log_queue_reader.start()
# setup of the actual pool
self._pool = []
working_dirs = []
debug('generating workers')
worker_root = None
for i in range(processes):
debug('generating worker '+str(i))
workername = self._get_worker_name(i)
#setup working directories for parallel_ema
for msi in msis:
if msi.working_directory != None:
if worker_root == None:
wd = msis[0].working_directory
abs_wd = os.path.abspath(wd)
worker_root = os.path.dirname(abs_wd)
wd_name = workername + msi.name
working_directory = os.path.join(worker_root, wd_name)
# working_directory = tempfile.mkdtemp(suffix=workername,
# prefix='tmp_',
# dir=worker_root)
working_dirs.append(working_directory)
shutil.copytree(msi.working_directory,
working_directory,
)
msi.set_working_directory(working_directory)
w = LoggingProcess(
self.log_queue,
level = logging.getLogger(ema_logging.LOGGER_NAME)\
.getEffectiveLevel(),
target=worker,
args=(self._inqueue,
self._outqueue,
msis,
kwargs
)
)
self._pool.append(w)
w.name = w.name.replace('Process', workername)
w.daemon = True
w.start()
debug(' worker '+str(i) + ' generated')
# thread for handling tasks
self._task_handler = threading.Thread(
target=CalculatorPool._handle_tasks,
name='task handler',
args=(self._taskqueue,
self._quick_put,
self._outqueue,
self._pool
)
)
self._task_handler.daemon = True
self._task_handler._state = RUN
self._task_handler.start()
# thread for handling results
self._result_handler = threading.Thread(
target=CalculatorPool._handle_results,
name='result handler',
args=(self._outqueue,
self._quick_get,
self._cache,
self.log_queue)
)
self._result_handler.daemon = True
self._result_handler._state = RUN
self._result_handler.start()
# function for cleaning up when finalizing object
self._terminate = Finalize(self,
self._terminate_pool,
args=(self._taskqueue,
self._inqueue,
self._outqueue,
self._pool,
self._task_handler,
self._result_handler,
self._cache,
working_dirs,
),
exitpriority=15
)
info("pool has been set up")