def __init__(self, load=None, **kwargs):
if load is None:
args = {}
else:
args = util.load_params(load, 'train')
util.update(args, mode=RL.Mode.TRAIN, **kwargs)
print(args)
Default.__init__(self, **args)
if self.init:
self.model.init()
self.model.save()
else:
self.model.restore()
context = zmq.Context.instance()
self.experience_socket = context.socket(zmq.PULL)
experience_addr = "tcp://%s:%d" % (
self.dump, util.port(self.model.name + "/experience"))
self.experience_socket.bind(experience_addr)
self.params_socket = context.socket(zmq.PUB)
params_addr = "tcp://%s:%d" % (self.dump,
util.port(self.model.name + "/params"))
print("Binding params socket to", params_addr)
self.params_socket.bind(params_addr)
self.sweep_size = self.batches * self.batch_size
print("Sweep size", self.sweep_size)
self.buffer = util.CircularQueue(self.sweep_size)
self.last_save = time.time()
def __init__(self, load=None, **kwargs):
if load is None:
args = {}
else:
args = util.load_params(load, 'train')
util.update(args,
mode=RL.Mode.TRAIN,
**kwargs
)
util.pp.pprint(args)
Default.__init__(self, **args)
if self.init:
self.rl.init()
self.rl.save()
else:
self.rl.restore()
print("Loading experiences from", self.data)
start_time = time.time()
self.experiences = hickle.load(self.data)
print("Loaded experiences in %d seconds." % (time.time() - start_time))
if 'initial' not in self.experiences:
self.experiences['initial'] = []
def run(**kwargs):
load = kwargs.get('load')
if load:
params = util.load_params(load, 'agent')
else:
params = {}
util.update(params, **kwargs)
print(params)
if params.get('gui'):
params['dolphin'] = True
if params.get('user') is None:
params['user'] = tempfile.mkdtemp() + '/'
print("Creating cpu.")
cpu = CPU(**params)
params['cpus'] = cpu.pids
if params.get('dolphin'):
dolphinRunner = DolphinRunner(**params)
# delay for a bit to let the cpu start up
time.sleep(2)
print("Running dolphin.")
dolphin = dolphinRunner()
else:
dolphin = None
print("Running cpu.")
cpu.run(dolphin_process=dolphin)
def get_pop_ids(path):
agent_params = util.load_params(path)
agent_pop_size = args.pop_size or agent_params.get('pop_size')
if agent_pop_size:
return list(range(agent_pop_size))
else:
return [-1]
def get_agents(path):
params = util.load_params(path)
pop_size = params.get('pop_size')
if pop_size:
pop_ids = list(range(pop_size))
else:
pop_ids = [-1]
return [(path, params, pop_id) for pop_id in pop_ids]
def get_agents(path):
params = util.load_params(path)
pop_size = params.get('pop_size')
if pop_size and args.pop_size:
pop_size = min(pop_size, args.pop_size)
if pop_size:
pop_ids = range(pop_size)
else:
pop_ids = [-1]
return [(path, params, pop_id) for pop_id in pop_ids]
def __init__(self, load=None, **kwargs):
if load is None:
args = {}
else:
args = util.load_params(load, 'train')
kwargs.update(experience_length=6000, )
util.update(args, mode=RL.Mode.TRAIN, **kwargs)
util.pp.pprint(args)
Default.__init__(self, **args)
if self.init:
self.rl.init()
self.rl.save()
else:
self.rl.restore()
if self.data is None:
self.data = os.path.join(self.rl.path, 'experience')
print("Loading experiences from", self.data)
files = os.listdir(self.data)
if self.file_limit:
files = files[:self.file_limit]
data_paths = [os.path.join(self.data, f) for f in files]
print("Loading %d experiences." % len(files))
self.experiences = []
parallel = True
if parallel:
for paths in util.chunk(data_paths, 100):
self.experiences.extend(
util.async_map(load_experience, paths)())
else:
for path in data_paths:
with open(path, 'rb') as f:
self.experiences.append(pickle.load(f))
self.valid_size = self.valid_batches * self.batch_size
def __init__(self, load=None, **kwargs):
if load is None:
args = {}
else:
args = util.load_params(load, 'train')
util.update(args, mode=RL.Mode.TRAIN, **kwargs)
util.pp.pprint(args)
Default.__init__(self, **args)
addresses = netifaces.ifaddresses(self.dump)
address = addresses[netifaces.AF_INET][0]['addr']
with open(os.path.join(self.model.path, 'ip'), 'w') as f:
f.write(address)
context = zmq.Context.instance()
self.experience_socket = context.socket(zmq.PULL)
experience_addr = "tcp://%s:%d" % (
address, util.port(self.model.name + "/experience"))
self.experience_socket.bind(experience_addr)
if self.send:
import nnpy
self.params_socket = nnpy.Socket(nnpy.AF_SP, nnpy.PUB)
params_addr = "tcp://%s:%d" % (
address, util.port(self.model.name + "/params"))
print("Binding params socket to", params_addr)
self.params_socket.bind(params_addr)
self.sweep_size = self.batches * self.batch_size
print("Sweep size", self.sweep_size)
if self.init:
self.model.init()
self.model.save()
else:
self.model.restore()
self.last_save = time.time()
def run(**kwargs):
load = kwargs.get('load')
if load:
params = util.load_params(load, 'agent')
else:
params = {}
util.update(params, **kwargs)
pp.pprint(params)
if params.get('gui'):
params['dolphin'] = True
if params.get('user') is None:
params['user'] = tempfile.mkdtemp() + '/'
if params.get('random_swap'):
task_id = os.environ.get('SLURM_ARRAY_TASK_ID')
if task_id is not None:
params['swap'] = int(task_id) % 2
else:
import random
params['swap'] = random.getrandbits(1)
print("Creating cpu.")
cpu = CPU(**params)
params['cpus'] = cpu.pids
if params.get('dolphin'):
dolphinRunner = DolphinRunner(**params)
# delay for a bit to let the cpu start up
time.sleep(2)
print("Running dolphin.")
dolphin = dolphinRunner()
else:
dolphin = None
print("Running cpu.")
cpu.run(dolphin_process=dolphin)
def __init__(self, load=None, **kwargs):
if load is None:
args = {}
else:
args = util.load_params(load, 'train')
util.update(args, **kwargs)
util.pp.pprint(args)
Default.__init__(self, **args)
addresses = netifaces.ifaddresses(self.dump)
address = addresses[netifaces.AF_INET][0]['addr']
util.makedirs(self.learner.path)
with open(os.path.join(self.learner.path, 'ip'), 'w') as f:
f.write(address)
self.experience_socket = nnpy.Socket(nnpy.AF_SP, nnpy.PULL)
experience_addr = "tcp://%s:%d" % (
address, util.port(self.learner.path + "/experience"))
self.experience_socket.bind(experience_addr)
if self.send:
self.params_socket = nnpy.Socket(nnpy.AF_SP, nnpy.PUB)
params_addr = "tcp://%s:%d" % (
address, util.port(self.learner.path + "/params"))
print("Binding params socket to", params_addr)
self.params_socket.bind(params_addr)
self.sweep_size = self.batch_size
print("Sweep size", self.sweep_size)
if self.init:
self.learner.init()
self.learner.save()
else:
self.learner.restore()
self.last_save = time.time()
parser.add_argument('--profile',
action='store_true',
help='heap profile trainer')
parser.add_argument('--disk',
action='store_true',
help='run agents and dump experiences to disk')
parser.add_argument('--tenenbaum',
action='store_true',
help='run trainer on higher priority')
parser.add_argument('--use_everything',
action='store_true',
help='run agents on lower priority')
args = parser.parse_args()
params = util.load_params(args.path)
run_trainer = True
run_agents = True
if args.local:
agent_dump = "localhost"
trainer_dump = "127.0.0.1"
else: # running on openmind
agent_dump = None
if args.disk:
run_trainer = False
elif args.trainer:
agent_dump = "172.16.24.%s" % args.trainer
run_trainer = False
else:
