def _update_evaluation(self):
"""Calculate evaluation function of system."""
now = time.time()
if not self._buffer['continue']:
func = self._get_evaluation_algorithm()
value = self._get_evaluation_value()
self._config['tracker_eval_enable'] = False
return ui.info('found optimum with: %s = %s' % (
func['name'], func['formater'](value)))
if ((now - self._buffer['eval_prev_time'])
> self._config['tracker_eval_time_interval']):
func = self._get_evaluation_algorithm()
value = self._get_evaluation_value()
progress = self._get_progress()
# update time of last evaluation
self._buffer['eval_prev_time'] = now
# add evaluation to array
if not isinstance(self._buffer['eval_values'],
numpy.ndarray):
self._buffer['eval_values'] = \
numpy.array([[progress, value]])
else:
self._buffer['eval_values'] = \
numpy.vstack((self._buffer['eval_values'], \
numpy.array([[progress, value]])))
return ui.info('finished %.1f%%: %s = %s' % (
progress * 100., func['name'], func['formater'](value)))
return False
def _install_pkg(self, pkg=None):
if not pkg:
ui.info("trying to install bioconductor base")
else:
ui.info("trying to install "
f"bioconductor package: '{pkg}'")
# try to evaluate the remote R script biocLite()
bioclite = "https://bioconductor.org/biocLite.R"
sysstout = sys.stdout
try:
sys.stdout = NullDevice()
from rpy2.robjects.packages import importr
base = importr('base')
base.source(bioclite)
base.require('biocLite')
sys.stdout = sysstout
except Exception as err:
sys.stdout = sysstout
raise ValueError(
f"could not evaluate remote R script: '{bioclite}'") from err
# try to install bioconductor packages with biocLite()
if not pkg:
return self._exec_rcmd("biocLite()")
return self._exec_rcmd("biocLite('%s')" % pkg)
def run_unittest() -> None:
"""Run unittest."""
ui.info(f"testing nemoa {env.get_var('version')}")
cur_level = ui.get_notification_level()
ui.set_notification_level('CRITICAL')
try:
nemoa.test.run(stream=sys.stderr)
finally:
ui.set_notification_level(cur_level)
def run_unittest() -> None:
"""Run unittest."""
ui.info(f"testing nemoa {env.get_var('version')}")
cur_level = ui.get_notification_level()
ui.set_notification_level('CRITICAL')
try:
nemoa.test.run(stream=sys.stderr)
finally:
ui.set_notification_level(cur_level)
def print_usage() -> None:
"""Print script usage to standard output."""
ui.info("Usage: nemoa [options]\n\n"
"Options:\n\n"
" -h --help "
" Print this\n"
" -s --shell "
" Start nemoa session in IPython interactive shell\n"
" -l --list "
" List workspaces\n"
" -w --workspace "
" List scripts in workspace\n"
" -r --run-script "
" Open workspace and execute script\n"
" -a --arguments "
" Arguments passed to script\n"
" -t --test "
" Run unittest on current installation\n"
" -v --version "
" Print version")
def print_usage() -> None:
"""Print script usage to standard output."""
ui.info("Usage: nemoa [options]\n\n"
"Options:\n\n"
" -h --help "
" Print this\n"
" -s --shell "
" Start nemoa session in IPython interactive shell\n"
" -l --list "
" List workspaces\n"
" -w --workspace "
" List scripts in workspace\n"
" -r --run-script "
" Open workspace and execute script\n"
" -a --arguments "
" Arguments passed to script\n"
" -t --test "
" Run unittest on current installation\n"
" -v --version "
" Print version")
def optimize(self, config = None, **kwds):
""" """
if not self._set_config(config, **kwds): return None
if not self._set_buffer_reset(): return None
# get name of optimization algorithm
name = self._config.get('algorithm', None)
if not name:
raise ValueError("""could not optimize '%s'
(%s): no optimization algorithm has been set."""
% (self.model.name, self.model.system.type)) or None
# get instance of optimization algorithm
algorithm = self._get_algorithm(name, category = 'optimization')
if not algorithm:
raise ValueError("""could not optimize '%s':
unsupported optimization algorithm '%s'."""
% (self.model.name, name)) or None
# start optimization
if algorithm.get('type', None) == 'algorithm':
ui.info("optimize '%s' (%s) using %s."
% (self.model.name, self.model.system.type, name))
# start key events
if not self._buffer['key_events_started']:
ui.info("press 'h' for help or 'q' to quit.")
self._buffer['key_events_started'] = True
nemoa.set('shell', 'buffmode', 'key')
# 2Do retval, try / except etc.
transformation = algorithm.get('reference', None)
if not transformation: return None
retval = transformation()
retval &= self.model.network.initialize(self.model.system)
return retval
def print_workspaces() -> None:
"""Print list of workspaces to standard output."""
nemoa.set('mode', 'silent')
workspaces = nemoa.list('workspaces', base='user')
ui.info('Workspaces:\n')
for workspace in workspaces:
ui.info(' %s' % (workspace))
ui.info('')
def print_workspaces() -> None:
"""Print list of workspaces to standard output."""
nemoa.set('mode', 'silent')
workspaces = nemoa.list('workspaces', base='user')
ui.info('Workspaces:\n')
for workspace in workspaces:
ui.info(' %s' % (workspace))
ui.info('')
def print_scripts(workspace: str) -> None:
"""Print list of scripts to standard output."""
nemoa.set('mode', 'silent')
if nemoa.open(workspace):
ui.info('Scripts in workspace %s:\n' % (nemoa.get('workspace')))
for script in nemoa.list('scripts'):
ui.info(' %s' % (script))
ui.info('')
def print_scripts(workspace: str) -> None:
"""Print list of scripts to standard output."""
nemoa.set('mode', 'silent')
if nemoa.open(workspace):
ui.info('Scripts in workspace %s:\n' % (nemoa.get('workspace')))
for script in nemoa.list('scripts'):
ui.info(' %s' % (script))
ui.info('')
def _update_keypress(self):
"""Check Keyboard."""
char = nemoa.get('shell', 'inkey')
if not char: return True
if char == 'e':
pass
elif char == 'h':
ui.info(
"Keyboard Shortcuts\n"
"'e' -- calculate evaluation function\n"
"'h' -- show this\n"
"'q' -- quit optimization\n"
"'t' -- estimate finishing time")
elif char == 'q':
ui.info('aborting optimization')
self._buffer['continue'] = False
elif char == 't':
ftime = self._get_estimatetime()
ui.info('estimated finishing time %s' % ftime)
return True
def _dbn_pretraining(self):
"""Deep belief network pretraining.
Deep belief network pretraining is a meta algorithm that wraps
unittype specific optimization schedules, intended to perform
system local optimization from outer layers to inner layers.
The default optimization schedules uses restricted boltzmann
machines and contrastive divergency optimization.
"""
system = self.model.system
config = self._config
if 'units' not in system._params:
raise ValueError("""could not configure subsystems:
no layers have been defined!""") or None
# create backup of dataset (before transformation)
dataset = self.model.dataset
dataset_backup = dataset.get('copy')
# create layerwise subsystems for RBM pretraining
cid = int((len(system._units) - 1) / 2)
rbmparams = {'units': [], 'links': []}
for lid in range(cid):
src = system._params['units'][lid]
srcnodes = src['id'] + system._params['units'][-1]['id'] \
if src['visible'] else src['id']
tgt = system._params['units'][lid + 1]
tgtnodes = tgt['id']
cpy = system._params['units'][-(lid + 1)]
links = system._params['links'][(lid, lid + 1)]
linkclass = (src['class'], tgt['class'])
name = '%s <-> %s <-> %s' % (src['layer'], tgt['layer'],
cpy['layer'])
systype = {
('gauss', 'sigmoid'): 'rbm.GRBM',
('sigmoid', 'sigmoid'): 'rbm.RBM'
}.get(linkclass, None)
if not systype:
raise ValueError("""could not create
rbm: unsupported pair of unit classes '%s <-> %s'""" %
linkclass) or None
# create subsystem
subsystem = nemoa.system.new(
config={
'name': name,
'type': systype,
'init': {
'ignore_units': ['visible'] if lid else []
}
})
# create subnetwork and configure subsystem with network
network = nemoa.network.create('factor',
name=name,
visible_nodes=srcnodes,
visible_type=src['class'],
hidden_nodes=tgtnodes,
hidden_type=tgt['class'])
subsystem.configure(network)
# transform dataset with previous system and initialize
# subsystem with dataset
if lid:
vlayer = prevsys._params['units'][0]['layer']
hlayer = prevsys._params['units'][1]['layer']
dataset._initialize_transform_system(system=prevsys,
mapping=(vlayer, hlayer),
func='expect')
dataset.set('colfilter', visible=srcnodes)
# create model
model = nemoa.model.new(config={
'type': 'base.Model',
'name': name
},
dataset=dataset,
network=network,
system=subsystem)
# copy parameters from perantal subsystems hidden units
# to current subsystems visible units
if lid:
dsrc = rbmparams['units'][-1]
dtgt = model.system._params['units'][0]
lkeep = ['id', 'layer', 'layer_id', 'visible', 'class']
lcopy = [key for key in list(dsrc.keys()) if not key in lkeep]
for key in lcopy:
dtgt[key] = dsrc[key]
# reference parameters of current subsystem
# in first layer reference visible, links and hidden
# in other layers only reference links and hidden
links['init'] = model.system._params['links'][(0, 1)]
if lid == 0:
src['init'] = model.system._units['visible'].params
tgt['init'] = model.system._units['hidden'].params
# optimize model
schedule = self._get_schedule(
self._config.get('schedule_%s' % systype.lower(), 'default'))
if systype in schedule:
model.optimize(schedule[systype])
else:
model.optimize()
if not lid:
rbmparams['units'].append(model.system.get('layer', 'visible'))
rbmparams['links'].append(model.system._params['links'][(0, 1)])
rbmparams['units'].append(model.system.get('layer', 'hidden'))
prevsys = model.system
# reset data to initial state (before transformation)
dataset.set('copy', **dataset_backup)
# keep original inputs and outputs
mapping = system._get_mapping()
inputs = system._units[mapping[0]].params['id']
outputs = system._units[mapping[-1]].params['id']
# initialize ann with rbm optimized parameters
units = system._params['units']
links = system._params['links']
# initialize units and links until central unit layer
cid = int((len(units) - 1) / 2)
for id in range(cid):
# copy unit parameters
for attrib in list(units[id]['init'].keys()):
# keep name and visibility of layers
if attrib in ['layer', 'layer_id', 'visible', 'class']:
continue
# keep labels of hidden layers
if attrib == 'id' and not units[id]['visible']:
continue
units[id][attrib] = units[id]['init'][attrib]
units[-(id + 1)][attrib] = units[id][attrib]
del units[id]['init']
# copy link parameters and transpose numpy arrays
for attrib in list(links[(id, id + 1)]['init'].keys()):
if attrib in ['source', 'target']:
continue
links[(id, id + 1)][attrib] = \
links[(id, id + 1)]['init'][attrib]
links[(len(units) - id - 2,
len(units) - id - 1)][attrib] = \
links[(id, id + 1)]['init'][attrib].T
del links[(id, id + 1)]['init']
# initialize central unit layer
for attrib in list(units[cid]['init'].keys()):
# keep name and visibility of layers
if attrib in ['id', 'layer', 'layer_id', 'visible', 'class']:
continue
units[cid][attrib] = \
units[cid]['init'][attrib]
del units[cid]['init']
# remove output units from input layer, and vice versa
ui.info('cleanup unit and linkage parameter arrays.')
system._remove_units(mapping[0], outputs)
system._remove_units(mapping[-1], inputs)
return True
def _cdiv(self):
"""Contrastive Divergency parameter optimization."""
system = self.model.system
config = self._config
# set enable flags for restriction extensions
config['con_klpt_enable'] = False
if config['con_module']:
found = False
if config['con_module'] == 'klpt':
config['con_klpt_enable'] = True
about = """Kullback-Leibler penalty (expectation
value %.2f)""" % config['con_klpt_expect']
found = True
if found:
ui.info('using restriction: %s' % about)
# set enable flags for denoising extensions
if config['denoising']:
found = False
if config['denoising'].lower() == 'noise':
config['noise_enable'] = True
about = """data corruption (noise model '%s',
factor %.2f)""" % (config['noise_type'],
config['noise_factor'])
found = True
if found:
ui.info('using denoising: %s' % (about))
# set enable flags for acceleration extensions
config['acc_vmra_enable'] = False
if config['acc_module']:
found = False
if config['acc_module'].lower() == 'vmra':
config['acc_vmra_enable'] = True
about = """variance maximizing rate adaption (tail
length %i)""" % config['acc_vmra_length']
found = True
if found:
ui.info('using acceleration: %s' % about)
# set enable flags for globalization extensions
config['gen_rasa_enable'] = False
if config['gen_module']:
found = False
if config['gen_module'].lower() == 'rasa':
config['gen_rasa_enable'] = True
about = """rate adaptive annealing (temperature %.1f,
annealing %.1f)""" % (
config['gen_rasa_init_temperature'],
config['gen_rasa_annealing_factor'])
found = True
if found:
ui.info('using generalization: %s' % (about))
# init rasa
self.write('sa', init_rate=config['update_rate'])
while self.update():
# get training data (sample from stratified minibatches)
data = self._get_data_training()[0]
# update parameters
self._cdiv_update(data)
return True
def print_version() -> None:
"""Print nemoa version to standard output."""
version = env.get_var('version') or ''
ui.info('nemoa ' + version)
def print_version() -> None:
"""Print nemoa version to standard output."""
version = env.get_var('version') or ''
ui.info('nemoa ' + version)