def get_images(self):
# pdb.set_trace()
depth_image1 = self.depth_images[self.frame_id]
t = self.task[self.frame_id]
self.depth_time = utils.get_time_str(t)
self.thermal_time = utils.get_time_str(t)
return depth_image1
def test(args, model, test_loader, epoch, start_time, log_file,
train_epochs, train_losses, train_accuracy, valid_epochs, valid_losses, valid_accuracy, lr_change):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
counter = 0
for data, target in test_loader:
data, target = data.to(args.device), target.to(args.device)
output = model(data)
test_loss += nn.NLLLoss(reduction='sum')(output, target).item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
counter += len(pred)
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= counter
test_accuracy = correct/counter
valid_epochs.append(epoch)
valid_losses.append(test_loss)
valid_accuracy.append(test_accuracy)
# Get time elapsed
curr_time = time.time()
curr_time_str, elapsed_str = utils.get_time_str(start_time, curr_time)
log = '\n[{}] : Elapsed [{}] : Epoch {}:\tVALIDATION Loss: {:.4f}, Accuracy: {:.4f} ({}/{})\n'.format(
curr_time_str, elapsed_str, epoch,
test_loss, test_accuracy, correct, counter)
print(log)
log_file.write(log)
log_file.flush()
utils.make_plots(args.out_path, train_epochs, train_losses, train_accuracy, valid_epochs, valid_losses, valid_accuracy, lr_change)
def run_2_agents_easy(num=1):
"""one trained agent, one non-trained agent
opt['domain'] = domain
"""
from Domains import RCCarBarriers
# agent_paths = []
# for i in range(3):
# p = train_agent("params/Car/", i)
# agent_paths.append(p)
# print agent_paths
agent_paths = [
'./Results/Mixed_ActionsB/agent0/Aug16_03-49-898192',
'./Results/Mixed_ActionsB/agent2/Aug16_03-50-036324',
]
result_path = "./Results/CarMixed2/combine/" + get_time_str()
for i in range(1, 1 + num):
exp = generate_meta_experiment(i,
agent_paths[:1],
result_path,
expdomain=RCCarBarriers,
unique=False,
max_episode=2000)
exp.run(visualize_steps=0,
visualize_performance=0,
debug_on_sigurg=True)
# exp.plot()
exp.save()
print result_path
dom = exp.domain
return result_path
def run_2_agents_barriers_3(num=1):
from Domains import RCCarBarrier_2
# agent_paths = []
# for i in range(3):
# p = train_agent("params/Car/", i)
# agent_paths.append(p)
# print agent_paths
agent_paths = [
'./Results/Mixed_ActionsB/agent0/Aug16_03-49-898192',
'./Results/Mixed_ActionsB/agent1/Aug16_03-50-596823',
'./Results/Mixed_ActionsB/agent2/Aug16_03-50-036324',
'./Results/Mixed_ActionsB/agent2/Aug16_03-50-036324',
'./Results/Mixed_ActionsB/agent2/Aug16_03-50-036324',
'./Results/Mixed_ActionsB/agent2/Aug16_03-50-036324'
]
result_path = "./Results/CarMixed2/combine/" + get_time_str()
for i in range(1, 1 + num):
exp = generate_meta_experiment(i,
agent_paths,
result_path,
expdomain=RCCarBarrier_2,
unique=False)
exp.run(visualize_steps=0,
visualize_performance=0,
debug_on_sigurg=True)
# exp.plot()
exp.save()
print result_path
dom = exp.domain
return result_path
def run_4_agents(num=1):
from Domains import GridWorld4
# agent_paths = []
# for i in range(4):
# p = train_agent("params/gridworld/", i)
# agent_paths.append(p)
# print agent_paths
agent_paths = [
'./Results/Mixed_Actions3/agent0/Aug02_11-11-650788',
'./Results/Mixed_Actions3/agent1/Aug02_11-11-866788',
'./Results/Mixed_Actions3/agent2/Aug02_11-11-932882',
'./Results/Mixed_Actions3/agent3/Aug02_11-12-741840'
]
result_path = "./Results/Mixed4/combine/" + get_time_str()
for i in range(1, 1 + num):
exp = generate_meta_experiment(i,
agent_paths,
result_path,
expdomain=GridWorld4,
unique=False)
exp.run(visualize_steps=0, debug_on_sigurg=True)
exp.save()
print result_path
def run_2_slideturn_turn(num=1):
# agent_paths = []
# for i in range(2):
# p = train_agent("params/Car/", i)
# agent_paths.append(p)
# print agent_paths
agent_paths = [
'./Results/Mixed_ActionsB/agent0/Aug21_11-38-389943',
'./Results/Mixed_ActionsB/agent1/Aug21_11-43-003799'
]
result_path = "./Results/CarSlideTurn/combine/" + get_time_str()
for i in range(1, 1 + num):
exp = generate_meta_experiment(i,
agent_paths,
result_path,
max_episode=1000,
expdomain=RCCarRightTurn,
unique=False)
exp.run(visualize_steps=0,
visualize_performance=0,
debug_on_sigurg=True)
# exp.plot()
exp.save()
print result_path
def run_2_slideturn_mixed(num=1):
"""One agent trained on RightTurn, other on LeftSlide - comparing on stitched domain"""
from Domains import RCCarSlideTurn #, RCCarSlideInvert
# agent_paths = []
# for i in range(2):
# p = train_agent("params/Car/", i)
# agent_paths.append(p)
# print agent_paths
agent_paths = [
'./Results/Mixed_ActionsB/agent0/Aug21_11-38-389943',
'./Results/Mixed_ActionsB/agent1/Aug21_11-43-003799'
]
result_path = "./Results/CarSlideTurn/localrbf/increasedres/" + get_time_str(
)
for i in range(1, 1 + num):
exp = generate_meta_experiment(i,
agent_paths,
result_path,
max_episode=1000,
expdomain=RCCarSlideTurn,
unique=False)
exp.run(visualize_steps=0,
visualize_performance=0,
debug_on_sigurg=True)
# exp.plot()
exp.save()
print result_path
def get_server_stat(api_url):
r = requests.get(api_url)
stats = utils.load_json_str(r.text)
update_time_unix = utils.str_to_num(stats.get('updated'))
if not update_time_unix:
update_time = 'unknown'
else:
update_time = utils.get_time_str(update_time_unix) + ' UTC'
try:
stat = stats.get('servers')[4]
except TypeError:
stat_str = 'unknown'
else:
network_rx, rx_measure = utils.select_max_measure(
utils.str_to_num(stat.get('network_rx')))
network_tx, tx_measure = utils.select_max_measure(
utils.str_to_num(stat.get('network_tx')))
network_in = utils.str_to_num(stat.get('network_in')) // utils.GB
network_out = utils.str_to_num(stat.get('network_out')) // utils.GB
memory_total = utils.str_to_num(stat.get('memory_total')) // 1024
memory_used = utils.str_to_num(stat.get('memory_used')) // 1024
stat_str = 'mem: %dM / %dM\nnetwork: %.1f%s / %.1f%s\nbandwith: %dG / %dG\n' % (
memory_used, memory_total, network_tx, tx_measure, network_rx,
rx_measure, network_out, network_in)
return stat_str, update_time
def _evaluate(experiment, model_setup, session, eval_losses, losses_to_record):
loss_records_eval = {loss: [] for loss in losses_to_record}
print("\n%s ------- EVALUATING %s -------" % (utils.get_time_str(), model_setup.model_name))
print("\t\t" + "\t".join([l for l in losses_to_record]))
eval_iter = 0
# Loops until the iterator ends
while True:
# Record losses every x iterations
try:
cur_val_losses = session.run([eval_losses[l_name] for l_name in losses_to_record])
except tf.errors.OutOfRangeError:
break
to_print = [eval_iter] + cur_val_losses
print_form = "Iter%04d:" + "\t%.2f" * len(losses_to_record)
print(print_form % tuple(to_print))
# Store losses for later display
for i, loss_name in enumerate(losses_to_record):
loss_records_eval[loss_name].append(cur_val_losses[i])
eval_iter += 1
return loss_records_eval
def save_once(step, pring_log=True):
save_path = os.path.join(ckpt_dir, get_time_str())
saver_all.save(sess=sess,
save_path=save_path,
global_step=step,
write_meta_graph=False)
if pring_log:
print('save:', save_path)
return save_path
def generate_meta_experiment(exp_id,
agent_paths,
path,
unique=True,
expdomain=None):
opt = {}
if unique:
opt["path"] = os.path.join(path, get_time_str())
else:
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 10000
opt["num_policy_checks"] = 40
opt["checks_per_policy"] = 20
# start_at = np.array([4, 6])
agents = load_all_agents(agent_paths,
pretrained=True,
load_confidence=True)
for a in agents:
a.policy.epsilon = 0
if expdomain:
actual_domain = expdomain(mapname=mapname,
terrain_augmentation=False,
noise=0.1)
else:
actual_domain = GridWorldMixed(mapname=mapname,
terrain_augmentation=False,
noise=0.1)
domain = PolicyMixer(actual_domain, agents)
representation = Tabular(domain)
policy = eGreedy(representation) # , tau=.1)
opt['agent'] = Q_Learning(policy,
representation,
discount_factor=0.9,
initial_learn_rate=0.8,
lambda_=0.5,
learn_rate_decay_mode='boyan',
boyan_N0=2380)
opt['domain'] = domain
experiment = Experiment(**opt)
path_join = lambda s: os.path.join(opt["path"], s)
if not os.path.exists(opt["path"]):
os.makedirs(opt["path"])
shutil.copy(inspect.getsourcefile(inspect.currentframe()),
path_join("experiment.py"))
return experiment
def crawl_vtvprograms(start_date, end_date):
url_fmt = "https://vtv.vn/lich-phat-song-ngay-{}-thang-{}-nam-{}.htm"
title_prefix = "Xem truyền hình trực tiếp kênh"
data = {}
for date in utils.generate_datetime_objs(start_date, end_date):
data_by_date = {}
# Crawl single page
day, month, year = utils.get_specific_fmt_time(date)
# print("(day={}, month={}, year={})".format(day, month, year))
url = url_fmt.format(day, month, year)
root = html.document_fromstring(requests.get(url).content)
# Find channels
titles = root.xpath("//ul[@class = 'list-channel']//a/@title")
channels = [title[len(title_prefix) + 1:] for title in titles]
# print(channels)
# Find programs
program_elms = root.xpath("//div[@id = 'wrapper']/ul[@class = 'programs']")
for program_elm, channel in zip(program_elms, channels):
# print("Channel = ", channel)
li_elms = program_elm.xpath("./li")
program_list = []
for li in li_elms:
duration = li.xpath("@duration")[0] or ''
start_time = li.cssselect("span.time")[0].text or ''
program_name = li.cssselect("span.title")[0].text or ''
program_genre = li.cssselect("a.genre")
if program_genre is None or len(program_genre) == 0:
program_genre = ''
else:
# print(html.tostring(program_genre[0], encoding='utf-8'))
program_genre = program_genre[0].text
# print(program_genre)
# print("Start time = {}, duration = {}, program = {}".format(start_time, duration, program))
program_list.append(DEFAULT_DELIMITER.join([start_time, program_name, program_genre]))
data_by_date.update({channel: program_list})
date_str = utils.get_time_str(date)
data.update({date_str: data_by_date})
print("Crawl broadcast schedule of date {} done".format(date_str))
# break
return data
def train(args, model, train_loader, optimizer, epoch, start_time, log_file,
train_epochs, train_losses, train_accuracy, valid_epochs, valid_losses, valid_accuracy, lr_change):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
# Get data
data, target = data.to(args.device), target.to(args.device)
# Get model output
optimizer.zero_grad()
output = model(data)
# Calc loss
loss = nn.NLLLoss()(output, target)
# Backprop
loss.backward()
optimizer.step()
# Log, Plot
if (epoch*len(train_loader) + batch_idx) % args.log_interval == 0:
# Check loss, accuracy
train_epochs.append(epoch + batch_idx/len(train_loader))
train_losses.append(loss.item())
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
train_accuracy.append(pred.eq(target.view_as(pred)).sum().item()/len(pred))
# Get time elapsed
curr_time = time.time()
curr_time_str, elapsed_str = utils.get_time_str(start_time, curr_time)
# Log
log = '[{}] : Elapsed [{}]: Epoch: {} [{}/{} ({:.0f}%)]\tTRAIN Loss: {:.6f}\tAccuracy: {:.4f}\n'.format(
curr_time_str, elapsed_str, epoch, batch_idx, len(train_loader), 100.*batch_idx/len(train_loader),
train_losses[-1], train_accuracy[-1])
print(log)
log_file.write(log)
log_file.flush()
utils.mem_check()
utils.make_plots(args.out_path, train_epochs, train_losses, train_accuracy, valid_epochs, valid_losses, valid_accuracy, lr_change)
# Save models
if (epoch*len(train_loader) + batch_idx) % args.model_save_interval == 0:
model_name = os.path.join(args.out_path, 'model_epoch_{:04d}_batch_{:05d}_of_{:05d}.pth'.format(epoch, batch_idx, len(train_loader)))
print("Saving model", model_name)
torch.save(model.state_dict(), model_name)
def report_label_distribution(name: str, labels: List[Dict[int, int]],
label_name_map: Dict[int, str],
report_count: int) -> None:
logname = os.path.join('reports', utils.get_time_str(),
'distr_{}'.format(name))
report = '{} set label distribution:\n\ntotal: {}\n'.format(
name, sum(map(len, labels)))
counter = Counter[int](utils.flatten(l.values() for l in labels))
most_common = counter.most_common(report_count)
table_entries = [(count, label_name_map[key])
for (key, count) in most_common]
report += tabulate(table_entries, headers=['Count', 'Label']) + '\n'
utils.write(report, logname)
def send(self):
while True:
msg = self.msg_queue.get()
time_str = get_time_str()
print('[send email]', time_str, msg)
body = f"<html><body><h1>{time_str}</h1><p>{msg}</p></body></html>"
# -------------------------------------------------------------------------------------------------
msg = MIMEText(body, 'html', 'utf-8')
msg['From'] = self._format_addr('TraderMonitor <%s>' % self.from_addr)
msg['To'] = self._format_addr('<%s>' % self.to_addr)
msg['Subject'] = Header('sfl_trader', 'utf-8').encode()
server = smtplib.SMTP_SSL(self.smtp_server, 465)
# server.set_debuglevel(1)
server.login(self.from_addr, self.password)
server.sendmail(self.from_addr, [self.to_addr], msg.as_string())
server.quit()
def init_logger(self, log_dir=None, level=logging.INFO):
"""Init the logger.
Args:
log_dir(str, optional): Log file directory. If not specified, no
log file will be used.
level (int or str): See the built-in python logging module.
Returns:
:obj:`~logging.Logger`: Python logger.
"""
logger = get_logger(level, name=__name__)
if log_dir:
filename = '{}.log'.format(get_time_str())
log_file = osp.join(log_dir, filename)
logger_file_handler(logger, log_file, level=level)
return logger
def generate_multinomial_experiment(exp_id,
agent_paths,
path,
unique=True,
expdomain=None):
opt = {}
if unique:
opt["path"] = os.path.join(path, get_time_str())
else:
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 8000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 1
agents = load_all_agents(agent_paths,
pretrained=True,
load_confidence=True)
for a in agents:
assert type(a.policy).__name__ == "DynamicsConfidenceGreedy"
a.policy.epsilon = 0
if expdomain:
domain = expdomain(mapname=mapname, noise=0.1)
else:
domain = GridWorldMixed(mapname=mapname, noise=0.1)
representation = IncrementalTabular(domain)
policy = MAgentMultinomial(representation, agents) # , tau=.1)
print "$" * 10
print "You are currently running {}".format(policy.__class__)
opt['agent'] = NoopAgent(representation=representation, policy=policy)
opt['domain'] = domain
experiment = Experiment(**opt)
path_join = lambda s: os.path.join(opt["path"], s)
if not os.path.exists(opt["path"]):
os.makedirs(opt["path"])
shutil.copy(inspect.getsourcefile(inspect.currentframe()),
path_join("experiment.py"))
return experiment
def send(self):
while True:
jsondata = self.msg_queue.get()
msg = jsondata['msg']
to_addr = jsondata['to']
time_str = get_time_str()
print('[send email]', time_str, msg)
body = f"<html><body>{msg}</body></html>"
# -------------------------------------------------------------------------------------------------
msg = MIMEText(body, 'html', 'utf-8')
msg['From'] = self._format_addr("{name} {address}".format(name=self.from_name, address=self.from_addr))
msg['To'] = ','.join(to_addr)
msg['Subject'] = Header(self.subject, 'utf-8').encode()
msg.add_header('reply-to', self.reply_to)
server = smtplib.SMTP_SSL(self.smtp_server, self.smtp_port)
server.set_debuglevel(1)
server.login(self.from_addr, self.password)
server.sendmail(self.from_addr, to_addr, msg.as_string())
server.quit()
def init_logger(self, log_dir=None, level=logging.INFO):
"""Init the logger.
Args:
log_dir(str, optional): Log file directory. If not specified, no
log file will be used.
level (int or str): See the built-in python logging module.
Returns:
:obj:`~logging.Logger`: Python logger.
"""
logging.basicConfig(format='%(asctime)s - %(levelname)s - %(message)s',
level=level)
logger = logging.getLogger(__name__)
if log_dir and self.rank == 0:
filename = '{}.log'.format(get_time_str())
log_file = osp.join(log_dir, filename)
self._add_file_handler(logger, log_file, level=level)
return logger
def login(username, passwd):
print("Trying to login... " + utils.get_time_str())
data = dict(username=username, password=passwd)
s = requests.Session()
response = s.post(const.login_url,
data=data,
headers=const.headers,
timeout=5)
reply = response.json()
reply_code = reply["reply_code"]
if reply_code == 1:
print("User {} logged in successfully!".format(
reply["userinfo"]["username"]))
elif reply_code == 6:
print("User {} has logged in!".format(reply["userinfo"]["username"]))
elif reply_code == 3:
print("Authentication Failed!")
else:
print(reply)
def make_experiment(agent_paths=["./"],
sublearning=False,
exp_id=3,
path="./Results/Confidence2/",
temp=0.10517212721767522,
discount_factor=0.7,
lambda_=0.0,
init_state=None):
opt = {}
opt["path"] = os.path.join(path, get_time_str())
opt["exp_id"] = exp_id
opt["max_steps"] = 40000
opt["num_policy_checks"] = 10
opt["checks_per_policy"] = 20
# start_at = np.array([4, 6])
agents = load_all_agents(agent_paths,
pretrained=True,
load_confidence=True)
for i, a in enumerate(agents):
# import ipdb; ipdb.set_trace()
a.policy.epsilon = i * 0.02 + 0.1
# a.learn_rate_decay_mode = 'boyan'
# a.learn_rate = a.initial_learn_rate = 0.9
# a.boyan_N0 = 3000
a.learn_rate_decay_mode = 'dabney'
domain = RCCarModified(noise=0.1, init_state=(-2, 0.8, 0, 2.5))
representation = IncrementalTabular(domain)
policy = MultiAgentConfidence(representation, agents, tau=.1)
print "$" * 10
print "You are currently running {}".format(policy.__class__)
opt['agent'] = MetaAgent(representation=representation,
policy=policy,
learn_rate_decay_mode="const")
opt['domain'] = domain
experiment = Experiment(**opt)
return experiment
def report_rates_on_epoch(self, label: str, epno: int,
batch_results: BatchResult,
report_params: ReportParameters) -> None:
report_str = 'Total #preds: {}\n'.format(batch_results.total_weight)
true_pred = batch_results.weighted_true_preds
false_miss = batch_results.weighted_n_labs - batch_results.weighted_true_preds
false_pred = batch_results.weighted_n_preds - batch_results.weighted_true_preds
# true_miss = (batch_results.total_weight - batch_results.weighted_n_labs) - false_pred
report_for_i = lambda i: list(
map(lambda x: x[i], (report_params.label_name_map, true_pred,
false_miss, false_pred)))
report_str += tabulate(
list(map(report_for_i, range(report_params.top_k))),
headers=['Label', '#Correct', '#Missed', '#Falsely Predicted'])
report_str += '\n'
utils.write(
report_str,
os.path.join('reports', utils.get_time_str(),
'epoch_{}_{}'.format(epno, label)))
def inference(model, data_loader, dataset_name, device='cuda', output_folder=None,
expected_results=(), expected_results_sigma_tol=4):
device = torch.device(device)
num_devices = get_world_size()
logger = logging.getLogger("RetinaNet.inference")
dataset = data_loader.dataset
logger.info("Start evaluation on {} dataset({} images).".format(dataset_name, len(dataset)))
total_timer = Timer()
inference_timer = Timer()
total_timer.tic()
predictions = compute_on_dataset(model, data_loader, device, inference_timer)
# wait for all processes to complete before measuring the time
synchronize()
total_time = total_timer.toc()
total_time_str = get_time_str(total_time)
logger.info(
"Total run time: {} ({} s / img per device, on {} devices)".format(
total_time_str, total_time * num_devices / len(dataset), num_devices
)
)
predictions = accumulate_predictions_from_multiple_gpus(predictions)
if not is_main_process():
return
if output_folder:
torch.save(predictions, os.path.join(output_folder, "predictions.pth"))
extra_args = dict(
expected_results=expected_results,
expected_results_sigma_tol=expected_results_sigma_tol,
)
return evaluate(dataset=dataset,
predictions=predictions,
output_folder=output_folder,
**extra_args)
def __init__(self,
final_results_path="./Results/StartTimeExperimentMaha/",
subagent_path_root='./GeneratedAgents/RCCarConfidenceMaha/',
seed=1):
self.seed = seed
self.logger = logging.getLogger("StartingExperiment")
self.all_agent_paths = []
self.delay = 10000 # max_steps after beginning meta experiments
self.max_start_time = 10000
self.num_policy_checks = 25 # how many start times we evaluate
self.meta_iter_per_check = 200 # for the meta agents
self.results = []
self.final_results_path = os.path.join(final_results_path,
get_time_str())
try:
os.makedirs(self.final_results_path)
except Exception:
print "[ERR]: Path not created"
pass
self.subagent_path_root = subagent_path_root
steps = self.max_start_time / self.num_policy_checks
self.start_times = np.r_[steps:self.max_start_time + 1:steps]
def generate_meta_experiment(exp_id,
agent_paths,
path,
unique=True,
expdomain=None,
max_episode=5000):
opt = {}
if unique:
opt["path"] = os.path.join(path, get_time_str())
else:
opt["path"] = path
opt["exp_id"] = exp_id
# opt["max_steps"] = 50000
opt["max_episode"] = max_episode
opt["num_policy_checks"] = 50
opt["checks_per_policy"] = 1
# start_at = np.array([4, 6])
agents = load_all_agents(agent_paths,
pretrained=True,
load_confidence=False)
for a in agents:
a.policy.epsilon = 0
a.policy.turnOffExploration()
if expdomain:
actual_domain = expdomain(noise=0.1)
else:
actual_domain = RCCarModified(noise=0.1)
domain = RCPolicyMixer(actual_domain, agents)
# representation = MahaRBF(domain,
# num_rbfs=3000,
# # state_dimensions=np.array([0,1,3]),
# const_feature=False,
# resolution_min=21,
# resolution_max=21,
# include_border=True,
# normalize=True,
# seed=exp_id)
# representation = RandomLocalBases(domain, gaussian_kernel,
# num=100,
# normalization=True,
# resolution_max=20,
# seed=exp_id)
representation = NonparametricLocalBases(domain,
gaussian_kernel,
normalization=True)
policy = eGreedy(representation, 0.05) #, epsilon=0.1) # , tau=.1)
# policy = GibbsPolicy(representation)
opt['agent'] = Q_Learning(policy,
representation,
discount_factor=0.8,
initial_learn_rate=.8,
lambda_=0.1,
learn_rate_decay_mode='boyan',
boyan_N0=500)
opt['domain'] = domain
experiment = ExperimentMod(**opt)
path_join = lambda s: os.path.join(opt["path"], s)
if not os.path.exists(opt["path"]):
os.makedirs(opt["path"])
shutil.copy(inspect.getsourcefile(inspect.currentframe()),
path_join("experiment.py"))
return experiment
def save_index_maps(index_maps: graph_construction.IndexMaps) -> None:
path = os.path.join(utils.DIRNAME, 'index_maps')
os.makedirs(path, exist_ok=True)
with open(os.path.join(path, utils.get_time_str()), 'wb') as f:
pickle.dump(index_maps, f)
def make_experiment(exp_id=1,
path="./Results/Temp/MetaHP/",
unique=False,
max_episode=300,
num_rbfs=4000,
initial_learn_rate=0.9,
lambda_=0.7,
resolution=20):
opt = {}
if unique:
opt["path"] = os.path.join(path, get_time_str())
else:
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 30000
# opt["max_episode"] = max_episode
opt["num_policy_checks"] = 10
opt["checks_per_policy"] = 1
# start_at = np.array([4, 6])
from Domains import RCCarSlideTurn
expdomain = Domains.RCCarSlideTurn
agent_paths = [
'/home/jarvis/work/clipper/models/rl/Results/Mixed_ActionsB/agent0/Aug21_11-38-389943',
'/home/jarvis/work/clipper/models/rl/Results/Mixed_ActionsB/agent1/Aug21_11-43-003799'
]
agents = load_all_agents(agent_paths,
pretrained=True,
load_confidence=False)
for a in agents:
a.policy.epsilon = 0
# a.policy.turnOffExploration()
if expdomain:
actual_domain = expdomain(noise=0.)
else:
actual_domain = RCCarModified(noise=0.1)
domain = RCPolicyMixer(actual_domain, agents)
representation = MahaRBF(
domain,
num_rbfs=int(num_rbfs),
# state_dimensions=np.array([0,1,3]),
const_feature=False,
resolution_min=resolution,
resolution_max=resolution,
include_border=True,
normalize=True,
seed=exp_id)
policy = eGreedy(representation) #, epsilon=0.1) # , tau=.1)
opt['agent'] = Q_Learning(policy,
representation,
discount_factor=0.8,
initial_learn_rate=initial_learn_rate,
lambda_=lambda_,
learn_rate_decay_mode='const')
opt['domain'] = domain
experiment = Experiment(**opt)
# path_join = lambda s: os.path.join(opt["path"], s)
# if not os.path.exists(opt["path"]):
# os.makedirs(opt["path"])
# shutil.copy(inspect.getsourcefile(inspect.currentframe()), path_join("experiment.py"))
return experiment
if __name__ == '__main__':
# exp0 = make_experiment("params/gridworld/", yaml_file="agent0.yml",
# result_path="./Results/Mixed/agent0", save=True) #TODO
# assert mapname in exp0.domain.mapname, "Not using correct map!"
# exp0.run(visualize_performance=0)
# representation = exp0.agent.representation
# representation.dump_to_directory(exp0.full_path) # TODO
# result_path1 = "./Results/Mixed/agent1"
# exp1 = make_experiment("params/gridworld/", yaml_file="agent1.yml", result_path=result_path1, save=True)
# assert mapname in exp1.domain.mapname, "Not using correct map!"
# exp1.run(visualize_performance=0)
# representation = exp1.agent.representation
# representation.dump_to_directory(exp1.full_path)
# agent_paths = [exp0.full_path, exp1.full_path]
# print agent_paths
# agent_paths = ['./Results/Mixed/agent0/Jul26_05-40-200563', './Results/Mixed/agent1/Jul26_05-40-636595'] # 11x11 1
# agent_paths = ['./Results/Mixed/agent0/Jul29_02-39-566096', './Results/Mixed/agent1/Jul29_02-40-516981'] # 11x11 2
agent_paths = [
"./Results/Mixed/agent0/Jul29_02-48-334415/",
"./Results/Mixed/agent1/Jul29_02-48-430882/"
]
path = run_exp_trials(generate_meta_experiment, agent_paths,
"./Results/Meta/Harder11x11maze/" + get_time_str())
print path
# exp.run(visualize_performance=0, debug_on_sigurg=True)
# exp.save()
# exp.plot()
# import joblib