def _generate_browser_caps(self, capabilities):
caps = dict()
chrome_options = webdriver.ChromeOptions()
chrome_options.headless = os.environ.get(
"RX_ENDPOINT_HEADLESS") == "True"
if os.environ.get(
"RX_DEBUG") == "True"
chrome_options.add_argument("--auto-open-devtools-for-tabs")
chrome_options.add_experimental_option(
"prefs",
{"devtools.preferences.currentDockState": "\"undocked\""})
caps["chrome"] = recursive_update(chrome_options.to_capabilities(),
capabilities)
caps["opera"] = recursive_update(
{"operaOptions": {
"binary": os.environ.get("RX_OPERA_PATH")
}}, capabilities)
caps["safari"] = capabilities
caps["internet explorer"] = capabilities
caps["MicrosoftEdge"] = capabilities
firefox_options = webdriver.FirefoxOptions()
firefox_options.headless = os.environ.get(
"RX_ENDPOINT_HEADLESS") == "True"
caps["firefox"] = recursive_update(firefox_options.to_capabilities(),
capabilities)
return caps
def test_all(self):
self.assertEqual(
recursive_update({'a': {
'b': 2
}}, {
'a': {
'b': 3,
'd': 4
},
'e': 5
}), {
'a': {
'b': 3,
'd': 4
},
'e': 5
})
self.assertEqual(
recursive_update({'a': [1]}, {
'a': [2],
'c': {
'd': {
'c': 3
}
}
}), {
'a': [2],
'c': {
'd': {
'c': 3
}
}
})
self.assertEqual(
recursive_update({
'a': {
'c': 1,
'd': {}
},
'b': 4
}, {'b': 5}), {
'a': {
'c': 1,
'd': {}
},
'b': 5
})
self.assertEqual(
recursive_update({
'a': {
'c': 1,
'd': {}
},
'b': 4
}, {'a': 2}), {
'a': 2,
'b': 4
})
def main():
"""
Demo how to run an agent
"""
dirname = os.path.dirname(__file__)
if (len(sys.argv) > 1):
env_configName = str(sys.argv[1])
agent_configName = str(sys.argv[2])
else:
print("Default config ")
env_configName = "./configs/factory_floor_experiment.yaml"
env_filename = os.path.join(dirname, env_configName)
agent_configName = "./configs/agent_combined_config.yaml"
agent_filename = os.path.join(dirname, agent_configName)
env_parameters = getParameters(env_filename)
agent_parameters = getParameters(agent_filename)
# whao, you need to know exact contents of all files here..
recursive_update(
agent_parameters['subAgentList'][0]['parameters']['simulator'],
env_parameters['environment'])
print(env_parameters)
print(agent_parameters)
random.seed(env_parameters['seed'])
maxSteps = env_parameters['max_steps']
baseEnv = FactoryFloor(env_parameters['environment'])
packedActionSpace = PackedSpace(baseEnv.action_space,
{"robots": ["robot1", "robot2", "robot3"]})
env = ModifiedGymEnv(baseEnv, packedActionSpace)
logging.info("Starting example MCTS agent")
logoutput = io.StringIO("episode output log")
try:
logoutputpickle = open('./' + os.environ["SLURM_JOB_ID"] + '.pickle',
'wb')
except KeyError:
print("No SLURM_JOB_ID found")
logoutputpickle = io.BytesIO()
obs = env.reset()
complexAgent = createAgent(env.action_space, env.observation_space,
agent_parameters)
experiment = Experiment(complexAgent, env, maxSteps, render=True)
experiment.addListener(JsonLogger(logoutput))
experiment.addListener(PickleLogger(logoutputpickle))
stats, confidence_ints = experiment.run()
logoutputpickle.close()
print("json output:", logoutput.getvalue())
print("\n\nREWARD STATS: " + str(stats) + " \nCONFIDENCE INTERVALS " +
str(confidence_ints))
def __init__(self, config):
"""Create a driver with profile and repository configuration."""
config = recursive_update(self.default_config(), config)
self.check_config(config)
# The _config variables will never be changed
self._config = config
def publish(*args, context=None, **kwargs):
"""
Notify jinja to publish the template to the output_file location with all of the context provided.
:param args: the arguments that will be used to find the template in the template configuration
:param context:
:param kwargs:
:return:
"""
# Build up the context argument, special kwarg context will be used to provide a starting dictionary
if not context:
context = {}
recursive_update(context, kwargs)
# Insert all of the site details into the context as well
site_configuration = get_configuration().site.toDict()
recursive_update(context.setdefault('site', {}), site_configuration)
# Find the template definition object
template_definition = _template_configuration.get('templates', {})
for template_path in args:
try:
template_definition = template_definition[template_path]
except KeyError:
print('Cannot find template definition: {} '
'in template definitions: {}'.format(
args, _template_configuration.get('templates', {})))
raise
# Load the template and render to the destination file defined in the template_definition
root_template = _environment.get_template(
str(template_definition['template']))
output_file = template_definition['destination'].format(**context)
output_content = root_template.render(context)
output_file = _output_path / output_file
# Verify that the path is possible and write out the file
output_file.parent.mkdir(exist_ok=True)
output_file.write_text(output_content)
return output_file.relative_to(_output_path)
def __init__(
self,
name: str,
period: str = None,
start: dt.datetime = None,
profile: Profile = None,
targets: Dict[str, str] = None,
description: str = "",
delay: str = None,
hooks: List[str] = None,
queue: str = "default",
configuration: Dict = None,
):
# Check required arguments
if period is None:
raise ValueError("`period` is required.")
if start is None:
raise ValueError("`start` is required.")
if profile is None:
raise ValueError("`profile` is required.")
if not targets:
raise ValueError("`targets` is required.")
# Set default values
if delay is None:
delay = period
if hooks is None:
hooks = []
if configuration is None:
configuration = {}
# Parse values
self.delay = dtutil.parse_timedelta(delay)
self.period = dtutil.parse_timedelta(period)
self.name = name
self.description = description
self.start = dtutil.normalize_datetime(start)
self.targets = targets
self.queue = queue
# Instantiate the driver
driver_config = recursive_update(profile.configuration, configuration)
for key, value in driver_config.items():
if isinstance(value, str):
if value.startswith("env:$"):
driver_config[key] = os.environ[value[5:]]
self.driver = profile.driver(driver_config)
def get_config():
global CONFIG_DIR
final_config = None
if CONFIG_DIR is not None:
env = os.environ[ENV_VARIABLE]
default_config_path = os.path.join(CONFIG_DIR, 'default.json')
specific_config_path = os.path.join(CONFIG_DIR, env + '.json')
default_config, specific_config = None, None
with open(default_config_path) as f:
default_config = json.load(f)
with open(specific_config_path) as f:
specific_config = json.load(f)
final_config = recursive_update(default_config, specific_config)
return final_config
def __init__(self, agentId, actionspace:Dict, observationspace, parameters:dict):
AtomicAgent.__init__(self, agentId, actionspace, observationspace, parameters)
self._parameters = copy.deepcopy(self.DEFAULT_PARAMETERS)
self._parameters = recursive_update(self._parameters, parameters)
self._prev_state = None
# TODO: change to self._step_output = dict({"obs": observation_space.sample(), "reward": None, "done": None, "prev_action": None})
self._step_output = None
self._action = [-1]
decoratedspace = DecoratedSpace.create(actionspace)
self._num_actions = decoratedspace.n
self._train_frequency = self._parameters['train_frequency']
self._save_frequency = self._parameters['save_frequency']
self._agentId = agentId
self._PPO = PPO(self._parameters, self._num_actions)
self._buffer = Buffer(self._parameters, self._num_actions)
self._cumulative_rewards = 0
self._episode_step = 0
self._episodes = 1
self._t = 0
self._stats = {"cumulative_rewards": [],
"episode_length": [],
"value": [],
"learning_rate": [],
"entropy": [],
"policy_loss": [],
"value_loss": []}
tf.reset_default_graph()
self._step = 0
summary_path = 'summaries/' + self._parameters['name'] + '_' + \
self._parameters['algorithm']
if not os.path.exists(summary_path):
os.makedirs(summary_path)
self._summary_writer = tf.summary.FileWriter(summary_path)
if self._parameters['influence']:
self._seq_len = self._parameters['inf_seq_len']
elif self._parameters['recurrent']:
self._seq_len = self._parameters['seq_len']
else:
self._seq_len = 1
def _collect(self):
merged = {}
for m in self._maps.values():
merged = recursive_update(merged, m)
return frozendict(merged)
def __init__(self, agentId, actionspace: Dict, observationspace,
parameters: dict):
"""
@param parameters dict that must contain keys 'otherAgents', 'treeAgent' and 'rolloutAgent'
'otherAgents' must map to a (possibly empty) list of dict objects for a call to createAgents
'treeAgent' and 'rolloutAgent' must map to a dict object for a call to createAgent.
The dict must also contain a 'simulator' key containing a copy of the env parameters,
so that the agent can create a duplicate environment. The simulator dict must contain
a key 'fullname' containing the full name of the environment
for the class loader (se EnvironmentsFactory).
"""
super().__init__(agentId, actionspace, observationspace, parameters)
if not ('treeAgent' in parameters and 'rolloutAgent' in parameters):
raise "parameters does not contain 'treeAgent', 'rolloutAgent':" + str(
parameters)
self._parameters = copy.deepcopy(self.DEFAULT_PARAMETERS)
self._parameters = recursive_update(self._parameters, parameters)
if 'timeLimit' in self._parameters:
if 'iterationLimit' in self._parameters:
raise ValueError(
"Cannot have both a time limit and an iteration limit")
self._limitType = 'time'
else:
if 'iterationLimit' not in self._parameters:
DEFAULT_LIMIT = 1000
logging.error(
"Must have either a time limit or an iteration limit. Using default iteration limit: "
+ str(DEFAULT_LIMIT))
self._parameters['iterationLimit'] = DEFAULT_LIMIT
# number of iterations of the search
if self._parameters['iterationLimit'] < 1:
raise ValueError("Iteration limit must be greater than one")
self._limitType = 'iterations'
# start the simulator environment
envparams = self._parameters['simulator']
e = EnvironmentFactory.createEnvironment(envparams['fullname'],
envparams)
self._simulator = ModifiedGymEnv(
e, DecoratedSpace.create(copy.deepcopy(e.action_space)))
# diyBonus logic: to refactor -- include in a simulator factory / only for FactoryFloor env
diyBonus = self._parameters.get("diyBonus")
if diyBonus is not None:
self._simulator = DiyFactoryFloorAdapter(self._simulator, diyBonus,
self.agentId)
self._treeAgent = createAgent(self._simulator.action_space,
self._simulator.observation_space,
parameters['treeAgent'])
if 'otherAgents' in parameters:
rolloutAgentDict = copy.deepcopy(parameters['otherAgents'])
rolloutAgentList = rolloutAgentDict['subAgentList']
rolloutAgentList.append(parameters['rolloutAgent'])
rolloutAgentDict['subAgentList'] = rolloutAgentList
self._rolloutAgent = createAgent(self._simulator.action_space,
self._simulator.observation_space,
rolloutAgentDict)
self._otherAgents = createAgent(self._simulator.action_space,
self._simulator.observation_space,
parameters['otherAgents'])
else:
self._otherAgents = None
self._rolloutAgent = createAgent(self._simulator.action_space,
self._simulator.observation_space,
parameters['rolloutAgent'])
def setUpClass(cls):
logger.log(
60,
'#################### Testplan %s - START ####################' %
cls.__name__)
logging.getLogger('beehive.test.run')\
.log(60, '#################### Testplan %s - START ####################' % cls.__name__)
self = cls
# ssl
path = os.path.dirname(__file__).replace('beehive/common',
'beehive/tests')
pos = path.find('tests')
path = path[:pos + 6]
keyfile = None
certfile = None
# load configs
try:
home = os.path.expanduser('~')
if self.main_config_file is None:
config_file = '%s/beehive.yml' % home
self.main_config_file = config_file
else:
config_file = self.main_config_file
config = self.load_file(config_file)
logger.info('Get beehive test configuration: %s' % config_file)
except Exception as ex:
raise Exception(
'Error loading config file. Search in user home. %s' % ex)
# load configs fernet key
try:
home = os.path.expanduser('~')
if self.main_config_file is None:
config_file = '%s/beehive.fernet' % home
self.main_fernet_file = config_file
else:
config_file = self.main_config_file.replace('yml', 'fernet')
fernet = self.load_file(config_file)
logger.info('Get beehive test fernet key: %s' % config_file)
except Exception as ex:
raise Exception(
'Error loading fernet key file. Search in user home. %s' % ex)
# load specific configs for a set of test
try:
if self.spec_config_file is not None:
config2 = self.load_file(self.spec_config_file)
recursive_update(config, config2)
logger.info('Get beehive test specific configuration: %s' %
self.spec_config_file)
except Exception as ex:
raise Exception(
'Error loading config file. Search in user home. %s' % ex)
logger.info('Validation active: %s' % cls.validation_active)
cfg = config
self.test_config = config.get('configs', {})
if self.test_config.get('resource', None) is not None:
for key in self.test_config.get('resource').keys():
if 'configs' in cfg.keys() and 'resource' in cfg.get(
'configs').keys():
self.test_config.get('resource').get(key).update(
cfg.get('configs').get('resource').get(key, {}))
if 'configs' in cfg.keys() and 'container' in cfg.get(
'configs').keys():
self.test_config.get('container').update(
cfg.get('configs').get('container'))
self.fernet = fernet
# endpoints
self.endpoints = cfg.get('endpoints')
self.swagger_endpoints = cfg.get('swagger')
logger.info('Endpoints: %s' % self.endpoints)
# redis connection
if cfg.get('redis') is not None:
self.redis_uri = cfg.get('redis').get('uri')
if self.redis_uri is not None and self.redis_uri != '':
rhost, rport, db = self.redis_uri.split(';')
self.redis = redis.StrictRedis(host=rhost,
port=int(rport),
db=int(db))
# celery broker
self.worker = cfg.get('worker')
# mysql connection
self.db_uris = cfg.get('db-uris')
# get users
self.users = cfg.get('users')
# create auth client
self.auth_client = BeehiveApiClient([], 'keyauth', None, '', None)
# create api endpoint
self.api = {}
self.schema = {}
for subsystem, endpoint in self.endpoints.items():
self.api[subsystem] = RemoteClient(endpoint,
keyfile=keyfile,
certfile=certfile)
self.load_result()
self.custom_headers = {}
self.endpoit_service = 'auth'
def make_message(self, d):
self.payload = {}
[
recursive_update(self.payload, self.dot_expand(key, d[key]))
for key in d
]
def run_experiment(run_name, params=None):
exp = NIC_Experiment('stroke_mri_segmentation',
run_name,
device_name='cuda')
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
print("Using GPU {}".format(os.environ['CUDA_VISIBLE_DEVICES']))
# --------------------------------------------------------------
# CONFIGURATION
# --------------------------------------------------------------
cfg = {
# Dataset
'dataset': None,
'architecture': None,
# Patches
'npatches': 10000,
'patch_shape': (24, 24, 16),
'ratio_lesion': 0.5,
# Training
'do_train': True,
'batch_size': 16,
'train_loss': NIC_binary_focal_loss(gamma=2.0, alpha=0.25),
'early_stop': {
'patience': 8,
'metric': 'l1_er'
},
# Evaluation
'num_folds': 4,
'crossval_folds': None,
# Prediction
'patch_shape_pred': (24, 24, 16),
'pred_extraction_step': (4, 4, 1),
'uncertainty': {
'runs': 0,
'rate': 0.1,
'type': 'alpha'
}
}
if params is not None:
print("Setting parameters in configuration queue")
recursive_update(
cfg, params
) # Overwrite cfg with keys present in paramas (others left untouched)
# Other parameters
thresholds = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
lesion_sizes = [1, 10, 50, 100, 200, 300, 500, 1000]
# --------------------------------------------------------------
# Experiment
# --------------------------------------------------------------
# 1. Dataset
assert isinstance(cfg['dataset'], NIC_Dataset)
dataset = copy.copy(cfg['dataset'])
dataset.load()
print("Loaded {} dataset with {} training and {} testing".format(
dataset.__class__.__name__, len(dataset.train), len(dataset.test)))
# 2. Model
model_def = copy.deepcopy(cfg['architecture'])
model_parameters = filter(lambda p: p.requires_grad,
model_def.parameters())
nparams = sum([np.prod(p.size()) for p in model_parameters])
print("Setting model instance from {} architecture with {} parameters".
format(model_def.__class__.__name__, nparams))
# 3. Training and validation sets
train_gen = PatchGeneratorBuilder(
batch_size=cfg['batch_size'],
zeropad_shape=cfg['patch_shape'],
instruction_generator=PatchInstructionGenerator(
in_shape=cfg['patch_shape'],
out_shape=cfg['patch_shape'],
sampler=StrokeLesionSampling(in_shape=cfg['patch_shape'],
num_patches=cfg['npatches'],
ratio_lesion=cfg['ratio_lesion'],
augment_factor=2),
augment_to=cfg['npatches']),
num_workers=4,
shuffle=True)
val_gen = PatchGeneratorBuilder(
batch_size=cfg['batch_size'],
zeropad_shape=cfg['patch_shape'],
instruction_generator=PatchInstructionGenerator(
in_shape=cfg['patch_shape'],
out_shape=cfg['patch_shape'],
sampler=StrokeLesionSampling(in_shape=cfg['patch_shape'],
num_patches=cfg['npatches'],
ratio_lesion=cfg['ratio_lesion'],
augment_factor=2),
augment_to=cfg['npatches']),
num_workers=4,
shuffle=True)
trainer = EarlyStoppingTrain(
do_train=cfg['do_train'],
device=exp.get_device(),
max_epochs=200,
batch_size=cfg['batch_size'],
loss_func=cfg['train_loss'],
optimizer=TorchOptimizer(torch.optim.Adadelta, opts={'rho': 0.95}),
train_metrics={'acc': nic_binary_accuracy},
eval_metrics={
'acc': nic_binary_accuracy,
'l1_er': nic_binary_l1_er
},
early_stopping_metric=cfg['early_stop']['metric'],
early_stopping_patience=cfg['early_stop']['patience'])
predictor = PatchPredictor(
device=exp.get_device(),
num_classes=2,
lesion_class=1,
uncertainty_passes=cfg['uncertainty']['runs'],
uncertainty_dropout=cfg['uncertainty']['rate'],
uncertainty_dotype=cfg['uncertainty']['type'],
zeropad_shape=cfg['patch_shape_pred'],
instruction_generator=PatchGeneratorBuilder(
batch_size=64,
shuffle=False,
zeropad_shape=None,
instruction_generator=PatchInstructionGenerator(
in_shape=cfg['patch_shape_pred'],
out_shape=cfg['patch_shape_pred'],
sampler=UniformSampling(
in_shape=cfg['patch_shape_pred'],
extraction_step=cfg['pred_extraction_step']),
augment_to=None)))
crossval = SimpleCrossvalidation(
model_definition=model_def,
images=dataset.train,
num_folds=cfg['num_folds'],
model_trainer=trainer,
train_instr_gen=train_gen,
val_instr_gen=val_gen,
checkpoint_pathfile=os.path.join(
exp.checkpoints_dir,
'{}.pt'.format(run_name)), # '{}.pt'.format(run_name)
log_pathfile=exp.get_log_filename(),
test_predictor=predictor,
results_path=os.path.join(exp.results_dir, '{}/'.format(run_name)),
)
# EXECUTION
# Run validation
crossval.run_crossval()
# Load probability maps
results = load_result_set(os.path.join(exp.results_dir,
'{}/'.format(run_name)),
dataset.train,
result_type='pred',
file_format='nii.gz')
# Grid search of post-processing options
metrics_list, metrics_names = thresh_size_search(results, dataset.train,
thresholds, lesion_sizes)
print_metrics_list(metrics_list, metrics_names)
save_metrics_list(
os.path.join(exp.metrics_dir,
'{}_{}.csv'.format(get_formatted_timedate(), run_name)),
metrics_list, metrics_names)