def main():
parser = optparse.OptionParser()
parser.add_option("--use-tray-icon", dest="tray_icon",
action="store_true", default=False)
parser.add_option("--no-tray-icon", dest="tray_icon", action="store_false")
parser.add_option("--hide-main-window", action="store_true", default=False)
parser.add_option("-m", "--map", action="store", type="string", dest="filename",
help="Open a map from a given filename (from internal database)")
parser.add_option("-o", "--open", action="store", type="string",
dest="filepath", help="Open a map from a given filename (including path)")
(options, args) = parser.parse_args()
if not options.tray_icon:
options.hide_main_window=False
MapBrowser = Browser.Browser (
start_hidden = options.hide_main_window,
tray_icon = options.tray_icon
)
if options.filename != None:
MapBrowser.open_map_filename (utils.get_save_dir() + options.filename)
elif options.filepath != None:
MapBrowser.open_map_filename (options.filepath)
try:
gtk.main()
except:
print "Exception caught while running. Dying a death."
sys.exit(1)
def _download_module(self, module_url):
request = self.request
session = request.session
conn = sword2cnx.Connection(session['login'].service_document_url,
user_name=session['login'].username,
user_pass=session['login'].password,
always_authenticate=True,
download_service_document=False)
parts = urlparse.urlsplit(module_url)
path = parts.path.split('/')
path = path[:path.index('sword')]
module_url = '%s://%s%s' % (parts.scheme, parts.netloc, '/'.join(path))
# example: http://cnx.org/Members/user001/m17222/sword/editmedia
zip_file = conn.get_cnx_module(module_url = module_url,
packaging = 'zip')
save_dir = get_save_dir(request)
if save_dir is None:
request.session['upload_dir'], save_dir = create_save_dir(request)
extract_to_save_dir(zip_file, save_dir)
cnxml_file = open(os.path.join(save_dir, 'index.cnxml'), 'rb')
cnxml = cnxml_file.read()
cnxml_file.close()
conversionerror = None
try:
htmlpreview = cnxml_to_htmlpreview(cnxml)
save_and_backup_file(save_dir, 'index.html', htmlpreview)
files = get_files(save_dir)
save_zip(save_dir, cnxml, htmlpreview, files)
except libxml2.parserError:
conversionerror = traceback.format_exc()
raise ConversionError(conversionerror)
def __init__(self):
env_input_size: int = 6
super().__init__(env_input_size)
self.post_fn_remote = self.post_milp
self.get_nn_fn = self.get_nn
self.plot_fn = self.plot
self.template_2d: np.ndarray = np.array([[1, 0, 0, 0, 0, 0],
[1, -1, 0, 0, 0, 0]])
input_boundaries, input_template = self.get_template(0)
self.input_boundaries: List = input_boundaries
self.input_template: np.ndarray = input_template
_, template = self.get_template(1)
self.analysis_template: np.ndarray = template
collision_distance = 0
distance = [Experiment.e(6, 0) - Experiment.e(6, 1)]
# self.use_bfs = True
# self.n_workers = 1
self.rounding_value = 2**10
self.use_rounding = False
self.time_horizon = 40000
self.unsafe_zone: List[Tuple] = [(distance,
np.array([collision_distance]))]
self.input_epsilon = 0
# self.nn_path = os.path.join(utils.get_save_dir(),"tune_PPO_stopping_car/PPO_StoppingCar_acc24_00000_0_cost_fn=0,epsilon_input=0_2021-01-21_02-30-49/checkpoint_39/checkpoint-39")
self.nn_path = os.path.join(
utils.get_save_dir(),
"tune_PPO_stopping_car/PPO_StoppingCar_acc24_00001_1_cost_fn=0,epsilon_input=0_2021-01-21_02-30-49/checkpoint_58/checkpoint-58"
)
def main():
parser = optparse.OptionParser()
parser.add_option("--use-tray-icon", dest="tray_icon",
action="store_true", default=False)
parser.add_option("--no-tray-icon", dest="tray_icon", action="store_false")
parser.add_option("--hide-main-window", action="store_true", default=False)
parser.add_option("-m", "--map", action="store", type="string", dest="filename",
help="Open a map from a given filename (from internal database)")
parser.add_option("-o", "--open", action="store", type="string",
dest="filepath", help="Open a map from a given filename (including path)")
(options, args) = parser.parse_args()
if not options.tray_icon:
options.hide_main_window=False
MapBrowser = Browser.Browser (
start_hidden = options.hide_main_window,
tray_icon = options.tray_icon
)
if options.filename != None:
MapBrowser.open_map_filename (utils.get_save_dir() + options.filename)
elif options.filepath != None:
MapBrowser.open_map_filename (options.filepath)
try:
Gtk.main()
except:
print "Exception caught while running. Dying a death."
sys.exit(1)
def __init__(self):
env_input_size: int = 4
super().__init__(env_input_size)
self.post_fn_remote = self.post_milp
self.get_nn_fn = self.get_nn
self.plot_fn = self.plot
self.template_2d: np.ndarray = np.array([[0, 0, 1, 0], [0, 0, 0, 1]])
input_boundaries, input_template = self.get_template(0)
self.input_boundaries: List = input_boundaries
self.input_template: np.ndarray = input_template
_, template = self.get_template(1)
self.analysis_template: np.ndarray = template
safe_angle = 12 * 2 * math.pi / 360
theta = [self.e(4, 2)]
neg_theta = [-self.e(4, 2)]
self.unsafe_zone: List[Tuple] = [(theta, np.array([-safe_angle])),
(neg_theta, np.array([-safe_angle]))]
# self.use_rounding = False
self.rounding_value = 1024
self.time_horizon = 300
self.nn_path = os.path.join(
utils.get_save_dir(),
"tune_PPO_cartpole/PPO_CartPoleEnv_0205e_00001_1_cost_fn=1,tau=0.001_2021-01-16_20-25-43/checkpoint_3090/checkpoint-3090"
)
# self.tau = 0.001
self.tau = 0.02
def __init__(self):
super().__init__()
self.post_fn_remote = self.post_milp
self.before_start_fn = self.before_start
self.show_progress_plot = False
# self.nn_path = os.path.join(utils.get_save_dir(), "tune_PPO_bouncing_ball/PPO_BouncingBall_c7326_00000_0_2021-01-16_05-43-36/checkpoint_36/checkpoint-36")
self.nn_path = os.path.join(
utils.get_save_dir(),
"tune_PPO_bouncing_ball/PPO_BouncingBall_fb929_00003_3_2021-01-19_00-20-45/checkpoint_10/checkpoint-10"
)
def __init__(self):
super().__init__()
self.post_fn_remote = self.post_milp
self.before_start_fn = self.before_start
self.time_horizon = 300
self.use_rounding = False
# self.nn_path = os.path.join(utils.get_save_dir(), "tune_PPO_cartpole/PPO_CartPoleEnv_0205e_00001_1_cost_fn=1,tau=0.001_2021-01-16_20-25-43/checkpoint_3090/checkpoint-3090")
self.nn_path = os.path.join(
utils.get_save_dir(),
"tune_PPO_cartpole/PPO_CartPoleEnv_0205e_00000_0_cost_fn=0,tau=0.001_2021-01-16_20-25-43/checkpoint_193/checkpoint-193"
)
def import_clicked(self, button, other=None, *data):
chooser = gtk.FileChooserDialog(title=_("Open File"), action=gtk.FILE_CHOOSER_ACTION_OPEN, \
buttons=(gtk.STOCK_CANCEL, gtk.RESPONSE_CANCEL, gtk.STOCK_OPEN, gtk.RESPONSE_OK))
filtr = gtk.FileFilter ()
filtr.set_name(_('MAPZ Compressed Map (*.mapz)'))
filtr.add_pattern('*.mapz')
chooser.add_filter(filtr)
response = chooser.run()
if response == gtk.RESPONSE_OK:
filename = chooser.get_filename()
tf = tarfile.open(filename)
mapname = os.path.join (utils.get_save_dir (), tf.getnames()[0])
tf.extractall(utils.get_save_dir())
tf.close()
map = MapList.new_from_file(mapname)
map.filename = mapname
chooser.destroy()
def import_clicked(self, button, other=None, *data):
chooser = gtk.FileChooserDialog(title=_("Open File"), action=gtk.FILE_CHOOSER_ACTION_OPEN, \
buttons=(gtk.STOCK_CANCEL, gtk.RESPONSE_CANCEL, gtk.STOCK_OPEN, gtk.RESPONSE_OK))
filtr = gtk.FileFilter ()
filtr.set_name(_('MAPZ Compressed Map (*.mapz)'))
filtr.add_pattern('*.mapz')
chooser.add_filter(filtr)
response = chooser.run()
if response == gtk.RESPONSE_OK:
filename = chooser.get_filename()
tf = tarfile.open(filename)
mapname = os.path.join (utils.get_save_dir (), tf.getnames()[0])
tf.extractall(utils.get_save_dir())
tf.close()
map = MapList.new_from_file(mapname)
map.filename = mapname
chooser.destroy()
def import_clicked(self, button, other=None, *data):
chooser = Gtk.FileChooserDialog(title=_("Open File"), action=Gtk.FileChooserAction.OPEN, \
buttons=(Gtk.STOCK_CANCEL, Gtk.ResponseType.CANCEL, Gtk.STOCK_OPEN, Gtk.ResponseType.OK))
filtr = Gtk.FileFilter ()
filtr.set_name(_('MAPZ Compressed Map (*.mapz)'))
filtr.add_pattern('*.mapz')
chooser.add_filter(filtr)
response = chooser.run()
if response == Gtk.ResponseType.OK:
filename = chooser.get_filename()
tf = tarfile.open(filename)
mapname = utils.get_save_dir() + tf.getnames()[0]
tf.extractall(utils.get_save_dir())
tf.close()
map = MapList.new_from_file(mapname)
map.filename = mapname
chooser.destroy()
def import_clicked(self, button, other=None, *data):
chooser = Gtk.FileChooserDialog(title=_("Open File"), action=Gtk.FileChooserAction.OPEN, \
buttons=(Gtk.STOCK_CANCEL, Gtk.ResponeType.CANCEL, Gtk.STOCK_OPEN, Gtk.ResponseType.OK))
filtr = Gtk.FileFilter()
filtr.set_name(_('MAPZ Compressed Map (*.mapz)'))
filtr.add_pattern('*.mapz')
chooser.add_filter(filtr)
response = chooser.run()
if response == Gtk.ResponseType.OK:
filename = chooser.get_filename()
tf = tarfile.open(filename)
mapname = utils.get_save_dir() + tf.getnames()[0]
tf.extractall(utils.get_save_dir())
tf.close()
map = MapList.new_from_file(mapname)
map.filename = mapname
chooser.destroy()
class List(object):
DB = utils.get_save_dir() + 'hosts.db'
def __init__(self):
self._hosts = anydbm.open(List.DB, 'c')
def __del__(self):
self._hosts.close()
def get(self, index):
return pickle.loads(self._hosts[index])
def add(self, host):
assert isinstance(host, Host)
if host.name == None:
raise ExceptHostNameInvalid("host.name is None")
if host.name in self._hosts:
raise ExceptHostNameExist("host.name exist")
self.save(host)
def save(self, host):
assert isinstance(host, Host)
self._hosts[host.name] = pickle.dumps(host)
def remove(self, host):
if host != None:
del self._hosts[host.name]
def iteritems(self):
items = []
type(self._hosts.keys())
for k in self._hosts.keys():
items.append(self.get(k))
return items
def create_host(self):
i = ""
name = _("Unnamed").encode()
while name + ((i != "" or "") and "-" + str(i)) in self._hosts:
if i == "":
i = 1
else:
i += 1
if i != "":
name += "-" + str(i)
return Host(name, Type.create_type(Type.MYSQL_TCP_IP))
def doc_save_cb (self, widget, doc, top_element):
save_string = self.serialize_to_xml(doc, top_element)
if not self.save_file:
hsh = hashlib.sha256 (save_string)
save_loc = utils.get_save_dir ()
self.save_file = save_loc+hsh.hexdigest()+".map"
counter = 1
while os.path.exists(self.save_file):
print "Warning: Duplicate File. Saving to alternative"
self.save_file = save_loc + "Dup"+str(counter)+hsh.hexdigest()+".map"
counter += 1
self.save_map(self.save_file, save_string)
self.emit ('file_saved', self.save_file, self)
def doc_save_cb (self, widget, doc, top_element):
save_string = self.serialize_to_xml(doc, top_element)
if not self.save_file:
sham = sha.new (save_string)
save_loc = utils.get_save_dir ()
self.save_file = save_loc+sham.hexdigest()+".map"
counter = 1
while os.path.exists(self.save_file):
print "Warning: Duplicate File. Saving to alternative"
self.save_file = save_loc + "Dup"+str(counter)+sham.hexdigest()+".map"
counter += 1
self.save_map(self.save_file, save_string)
self.emit ('file_saved', self.save_file, self)
def doc_save_cb(self, widget, doc, top_element):
save_string = self.serialize_to_xml(doc, top_element)
if not self.save_file:
hsh = hashlib.sha256(save_string)
save_loc = utils.get_save_dir()
self.save_file = os.path.join(save_loc, hsh.hexdigest() + ".map")
counter = 1
while os.path.exists(self.save_file):
print "Warning: Duplicate File. Saving to alternative"
alt_name = "Dup" + str(counter) + hsh.hexdigest() + ".map"
self.save_file = save_loc + os.path.join(save_loc, alt_name)
counter += 1
with open(self.save_file, 'w') as f:
f.write(save_string)
self.emit('file_saved', self.save_file, self)
def doc_save_cb (self, widget, doc, top_element):
save_string = self.serialize_to_xml(doc, top_element)
if not self.save_file:
hsh = hashlib.sha256 (save_string)
save_loc = utils.get_save_dir ()
self.save_file = os.path.join (save_loc, hsh.hexdigest() + ".map")
counter = 1
while os.path.exists(self.save_file):
print "Warning: Duplicate File. Saving to alternative"
alt_name = "Dup" + str(counter) + hsh.hexdigest() + ".map"
self.save_file = save_loc + os.path.join (save_loc, alt_name)
counter += 1
with open(self.save_file, 'w') as f:
f.write(save_string)
self.emit ('file_saved', self.save_file, self)
def __init__(self, arg, torch_device):
self.torch_device = torch_device
self.model_type = arg.model
self.z_idx = 0 if arg.in_channel == 1 else arg.in_channel // 2 # middle of 2.5d slices
self.epoch = arg.epoch
self.start_epoch = 0
self.batch_size = arg.batch_size
self.save_path = utils.get_save_dir(arg)
self.log_file_path = self.save_path + "/log.txt"
self.logger = Logger(arg, self.save_path)
def __init__(self, arg, torch_device):
self.torch_device = torch_device
self.model_type = arg.model
self.z_idx = 0
self.epoch = arg.epoch
self.start_epoch = 0
self.batch_size = arg.batch_size
self.save_path = utils.get_save_dir(arg)
self.log_file_path = self.save_path+"/log.txt"
self.logger = Logger(arg, self.save_path)
def _iter():
for problem in ["cartpole"]: # "bouncing_ball", "stopping_car", "cartpole"
for method in ["standard"]: # , "ora"
if problem == "cartpole":
for tau in [0.001
]: # "tau": tune.grid_search([0.001, 0.02, 0.005]
for template in [1]: # [1,0,2]
for nn_index in range(1, min(100,
len(folder_cartpole))):
for checkpoint, check_folder in find_checkpoints(
os.path.join(utils.get_save_dir(),
folder_cartpole[nn_index])):
yield problem, method, {
"tau": tau,
"template": template,
"agent": nn_index,
"checkpoint": checkpoint,
"folder": check_folder
}
def __init__(self):
env_input_size: int = 2
super().__init__(env_input_size)
self.post_fn_remote = self.post_milp
self.get_nn_fn = self.get_nn
self.plot_fn = self.plot
self.template_2d: np.ndarray = np.array([[0, 1], [1, 0]])
input_boundaries, input_template = self.get_template(0)
self.input_boundaries: List = input_boundaries
self.input_template: np.ndarray = input_template
_, template = self.get_template(0)
self.analysis_template: np.ndarray = template
self.time_horizon = 500
self.rounding_value = 2**8
p = Experiment.e(self.env_input_size, 0)
v = Experiment.e(self.env_input_size, 1)
self.unsafe_zone: List[Tuple] = [([p, -v, v], np.array([0, 1, 0]))]
self.nn_path = os.path.join(
utils.get_save_dir(),
"tune_PPO_bouncing_ball/PPO_BouncingBall_c7326_00000_0_2021-01-16_05-43-36/checkpoint_36/checkpoint-36"
)
def __init__(self, arg, torch_device):
self.torch_device = torch_device
self.model_type = arg.model
self.z_idx = 0
self.epoch = arg.epoch
self.start_epoch = 0
self.batch_size = arg.batch_size
self.save_path = utils.get_save_dir(arg)
self.log_file_path = self.save_path + "/fold%s/log.txt" % (arg.fold)
if os.path.exists(self.save_path) is False:
os.mkdir(self.save_path)
if os.path.exists(self.save_path + "/fold%s" % (arg.fold)) is False:
os.mkdir(self.save_path + "/fold%s" % (arg.fold))
self.logger = Logger(arg, self.save_path + "/fold%s" % (arg.fold))
def _download_module(self, module_url):
request = self.request
session = request.session
conn = sword2cnx.Connection(session['login'].service_document_url,
user_name=session['login'].username,
user_pass=session['login'].password,
always_authenticate=True,
download_service_document=False)
parts = urlparse.urlsplit(module_url)
path = parts.path.split('/')
path = path[:path.index('sword')]
module_url = '%s://%s%s' % (parts.scheme, parts.netloc, '/'.join(path))
# example: http://cnx.org/Members/user001/m17222/sword/editmedia
zip_file = conn.get_cnx_module(module_url=module_url, packaging='zip')
save_dir = get_save_dir(request)
if save_dir is None:
request.session['upload_dir'], save_dir = create_save_dir(request)
extract_to_save_dir(zip_file, save_dir)
cnxml_file = open(os.path.join(save_dir, 'index.cnxml'), 'rb')
cnxml = cnxml_file.read()
cnxml_file.close()
conversionerror = None
try:
structuredhtml = cnxml_to_structuredhtml(cnxml)
save_and_backup_file(save_dir, 'index.structured.html',
structuredhtml)
htmlpreview = structuredhtml_to_htmlpreview(structuredhtml)
save_and_backup_file(save_dir, 'index.html', htmlpreview)
files = get_files(save_dir)
save_zip(save_dir, cnxml, structuredhtml, files)
except libxml2.parserError:
conversionerror = traceback.format_exc()
raise ConversionError(conversionerror)
action="store_true",
default=False,
help="don't use instructions in the model")
parser.add_argument("--no-mem",
action="store_true",
default=False,
help="don't use memory in the model")
args = parser.parse_args()
# Define run dir
suffix = datetime.datetime.now().strftime("%y-%m-%d-%H-%M-%S")
default_model_name = "{}_{}_seed{}_{}".format(args.env, args.algo, args.seed,
suffix)
model_name = args.model or default_model_name
save_dir = utils.get_save_dir(model_name)
# Define logger, CSV writer and Tensorboard writer
logger = utils.get_logger(save_dir)
csv_writer = utils.get_csv_writer(save_dir)
if args.tb:
from tensorboardX import SummaryWriter
tb_writer = SummaryWriter(save_dir)
# Log command and all script arguments
logger.info("{}\n".format(" ".join(sys.argv)))
logger.info("{}\n".format(args))
# Set seed for all randomness sources
# env.state = np.array([-0.57,0.21]) #guaranteed fail
# env.state = np.array([-0.57,0.21]) #guaranteed fail
# env.state = np.array([-0.22,0.05]) #success right
# env.state = np.array([0.25, -1]) # success left
# env.state = np.array([0.39,0.9]) # fail left
# env.state = np.array([0.834,0.497]) # fail
# start_state = np.array([-0.12, 0.33]) #100% success
# start_state = np.array([0.35,-0.83]) # should be 83% but appears 100%
# start_state = np.array([-0.48, -0.44]) #100% fail
start_state = np.array([0.37, 0.43]) # min 33% max 100%
state_size = 2
agent = Agent(state_size, 2)
agent.load(
os.path.join(
utils.get_save_dir(),
"Pendulum_Apr07_12-17-45_alpha=0.6, min_eps=0.01, eps_decay=0.2/checkpoint_final.pth"
))
action = None
render = False
n_trials = 10000
n_fail = 0
widgets = [
progressbar.Percentage(),
progressbar.Bar(),
progressbar.Variable('fails'),
', ',
progressbar.Variable('trials'),
]
with progressbar.ProgressBar(max_value=n_trials, widgets=widgets) as bar:
return options
if __name__ == '__main__':
# Book-keeping & paths
args = get_args()
dir_img = 'data/train/'
dir_mask = 'data/train_masks/'
dir_checkpoint = 'save/'
splitfile = "data/trainval.json"
runname = args.name
save_path = os.path.join(dir_checkpoint, runname)
save_dir = get_save_dir(save_path, runname,
training=False) # unique save dir
log = get_logger(save_dir, runname) # logger
log.info('Args: {}'.format(
json.dumps(
{
"batch_size": args.batch_size,
"taylor_batches": args.taylor_batches,
"prune_channels": args.prune_channels,
"gpu": args.gpu,
"load": args.load,
"channel_txt": args.channel_txt,
"scale": args.scale,
"lr": args.lr,
"iters": args.iters,
"epochs": args.epochs
},
def main(args):
#denoiser = VQ_CVAE(128, k=512, num_channels=3)
#denoiser.load_state_dict(torch.load("/mnt/home2/dlongo/eegML/VQ-VAE-master/vq_vae/saved_models/train.pt"))
#denoiser = torch.no_grad(denoiser)
#denoiser.cuda()
#denoiser = nn.DataParallel(denoiser)
# Get device
device, args.gpu_ids = utils.get_available_devices()
args.train_batch_size *= max(1, len(args.gpu_ids))
args.test_batch_size *= max(1, len(args.gpu_ids))
# Set random seed
utils.seed_torch(seed=SEED)
# Get save directories
train_save_dir = utils.get_save_dir(args.save_dir, training=True)
args.train_save_dir = train_save_dir
# Save args
args_file = os.path.join(train_save_dir, ARGS_FILE_NAME)
with open(args_file, 'w') as f:
json.dump(vars(args), f, indent=4, sort_keys=True)
# Set up logging and devices
log = utils.get_logger(train_save_dir, 'train_denoised')
tbx = SummaryWriter(train_save_dir)
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
if args.cross_val:
# Loop over folds
for fold_idx in range(args.num_folds):
log.info('Starting fold {}...'.format(fold_idx))
# Train
fold_save_dir = os.path.join(train_save_dir,
'fold_' + str(fold_idx))
if not os.path.exists(fold_save_dir):
os.makedirs(fold_save_dir)
# Training on current fold...
train_fold(args,
device,
fold_save_dir,
log,
tbx,
cross_val=True,
fold_idx=fold_idx)
best_path = os.path.join(fold_save_dir, 'best.pth.tar')
# Predict on current fold with best model..
if args.model_name == 'SeizureNet':
model = SeizureNet(args)
model = nn.DataParallel(model, args.gpu_ids)
model, _ = utils.load_model(model, best_path, args.gpu_ids)
model.to(device)
results = evaluate_fold(model,
args,
fold_save_dir,
device,
cross_val=True,
fold_idx=fold_idx,
is_test=True,
write_outputs=True)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
print('Fold {} test results: {}'.format(fold_idx, results_str))
log.info('Finished fold {}...'.format(fold_idx))
else:
# no cross-validation
# Train
train_fold(args, device, train_save_dir, log, tbx, cross_val=False)
best_path = os.path.join(train_save_dir, 'best.pth.tar')
if args.model_name == 'SeizureNet':
model = SeizureNet(args)
model = nn.DataParallel(model, args.gpu_ids)
model, _ = utils.load_model(model, best_path, args.gpu_ids)
model.to(device)
results = evaluate_fold(model,
args,
train_save_dir,
device,
cross_val=False,
fold_idx=None,
is_test=True,
write_outputs=True)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
print('Test set prediction results: {}'.format(results_str))
def main(args):
# set up logging and device
args.save_dir = utils.get_save_dir(args.save_dir, args.name, training=True)
logger = utils.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory () already exists and is not empty.")
os.makedirs(args.output_dir, exist_ok=True)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
# Generating the dictionaries
dep_dict, pos_dict, ent_dict, total_features = generate_dictionary(
args.train_ling_features_file, args.eval_ling_features_file,
args.test_ling_features_file)
# from IPython import embed; embed()
# Generating total_dictionary
total_dict = convert_string_features_to_array(total_features, dep_dict,
pos_dict, ent_dict)
# from IPython import embed; embed()
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_squad_examples(
input_file=args.train_file,
is_training=True,
version_2_with_negative=args.version_2_with_negative,
total_dictionary=total_dict)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
model = BertForQuestionAnsweringLing.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizer import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
# load training features
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length))
train_features = None
print(cached_train_features_file)
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
# load eval features
eval_examples = read_squad_examples(
input_file=args.predict_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative,
total_dictionary=total_dict)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
test_examples = read_squad_examples(
input_file=args.test_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative,
total_dictionary=total_dict)
test_features = convert_examples_to_features(
examples=test_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
# from IPython import embed; embed()
all_ling_features = torch.tensor(
[f.ling_features for f in train_features], dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_ling_features,
all_start_positions, all_end_positions)
steps_till_eval = args.eval_steps
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
best_F1 = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
if n_gpu == 1:
batch = tuple(
t.to(device)
for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, ling_features, start_positions, end_positions = batch
# from IPython import embed; embed()
loss = model(input_ids, segment_ids, input_mask, ling_features,
start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used and handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# add to tensorboard
loss_val = loss.item()
tbx.add_scalar('train/NLL', loss_val, global_step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
global_step)
steps_till_eval -= args.train_batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
logger.info('Evaluating at step {}...'.format(step))
# ema.assign(model)
results, _ = evaluate(model, eval_examples, eval_features,
device, args, logger,
args.version_2_with_negative,
args.dev_eval_file)
# saver.save(step, model, results[args.metric_name], device)
# ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
logger.info('Dev {}'.format(results_str))
# Log to TensorBoard
logger.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, global_step)
"""
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
"""
if results['F1'] > best_F1:
best_F1 = results['F1']
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
output_model_file = os.path.join(
args.output_dir, "pytorch_model_best.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
#model.to(device)
# Save a trained model
"""
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
model_state_dict = torch.load(output_model_file)
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model, state_dict=model_state_dict)
else:
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model)
model.to(device)
"""
# load the best trained model and eval on the eval set and test set
best_model_file = os.path.join(args.output_dir, "pytorch_model_best.bin")
model_state_dict = torch.load(best_model_file)
model = BertForQuestionAnsweringLing.from_pretrained(
args.bert_model, state_dict=model_state_dict)
model.to(device)
if args.do_predict and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
logger.info('Evaluating at the best model')
results, all_results = evaluate(model, eval_examples, eval_features,
device, args, logger,
args.version_2_with_negative,
args.dev_eval_file)
logger.info('Write the best eval results')
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file,
args.verbose_logging, args.version_2_with_negative,
args.null_score_diff_threshold, 'dev')
logger.info('Test set at the best model')
results, all_results = evaluate(model, test_examples, test_features,
device, args, logger,
args.version_2_with_negative,
args.test_eval_file)
logger.info('Write the best test set results')
output_prediction_file = os.path.join(args.output_dir,
"predictions_test.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions_test.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds_test.json")
write_predictions(test_examples, test_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file,
args.verbose_logging, args.version_2_with_negative,
args.null_score_diff_threshold, 'test')
"""
def main(args):
# Set up logging and devices
args.save_dir = utils.get_save_dir(args.save_dir, args.name, training=True)
log = utils.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = utils.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get model
log.info('Building model...')
model = get_model(args)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = utils.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
# Get saver
saver = utils.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Set optimizer and scheduler
optimizer_grouped_params = [{
'params':
[p for n, p in model.named_parameters() if 'classifier' not in n]
}, {
'params':
[p for n, p in model.named_parameters() if 'classifier' in n],
'lr':
args.lr_1
}]
optimizer = optim.AdamW(optimizer_grouped_params,
args.lr_2,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
train_dataset = NEWS(args.train_record_file, transform=transform)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = NEWS(args.dev_record_file, transform=transform)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), tqdm(
total=len(train_loader.dataset)) as progress_bar:
# for input_idxs, atten_masks, images, ids, y in train_loader:
for input_idxs, atten_masks, y in train_loader:
# print(y)
# Setup for forward
input_idxs = input_idxs.to(device)
atten_masks = atten_masks.to(device)
y = y.to(device)
batch_size = input_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p = model(input_idxs, atten_masks)
loss = torch.nn.functional.binary_cross_entropy(
log_p,
y,
weight=None,
size_average=None,
reduce=None,
reduction='mean')
loss_val = loss.item()
# Backward
loss.backward()
optimizer.step()
scheduler.step(step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
results, pred_dict = evaluate(model, dev_loader, device)
saver.save(step, model, results[args.metric_name], device)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}'
for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
help='Path to pre-trained model')
parser.add_argument(
'--freeze_layers',
type=int,
default=0,
help=
'Number of initial layers to freeze when fine-tuning | Choose from 1, 2, 3'
)
args = parser.parse_args()
logs_dir = 'logs'
if not os.path.exists(logs_dir):
os.makedirs(logs_dir)
args.save_dir = utils.get_save_dir(logs_dir, args.name)
# Training
if not torch.cuda.is_available() and args.cuda:
print(
'--cuda is passed but torch.cuda.is_available() returned False. Will use CPU instead.'
)
env = utils.wrap_deepmind(utils.make_atari(
args.env,
max_episode_steps=args.episode_length,
frameskip=args.frameskip),
frame_stack=True,
stacks=args.agent_history_length)
agent = Agent(env, args)
if not args.do_not_eval_after_epoch:
eval_results = evaluator.evaluate(model, global_step)
if args.local_rank in [-1, 0]:
saver.save(model, global_step, eval_results)
if __name__ == '__main__':
args = get_args()
if args.local_rank != -1:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl')
if args.local_rank in [-1, 0]:
args.save_dir = get_save_dir(args.save_dir, args.name)
logger = get_logger(args.save_dir, args.name, log_file=f'log_0.txt')
logger.info(f'Results will be saved to {args.save_dir}.')
tb_writer = SummaryWriter(args.save_dir)
else:
torch.distributed.barrier()
args.save_dir = get_save_dir(args.save_dir, args.name, use_existing_dir=True)
logger = get_logger(args.save_dir, args.name, verbose=False, log_file=f'log_{args.local_rank}.txt')
tb_writer = None
if args.local_rank == 0:
torch.distributed.barrier()
try:
train(args, logger, tb_writer)
except:
},
"env_config": {"cost_fn": tune.grid_search([0, 1, 2]),
"tau": tune.grid_search([0.001, 0.02, 0.005])}
}
return config
if __name__ == "__main__":
seed = 1234
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
ray.init(local_mode=False, include_dashboard=True, log_to_driver=False)
config = get_PPO_config(use_gpu=0.5, seed=seed)
datetime_str = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
tune.run(
"PPO",
stop={"info/num_steps_trained": 2e7, "episode_reward_mean": 7950},
config=config,
name=f"tune_PPO_cartpole",
checkpoint_freq=10,
checkpoint_at_end=True,
log_to_file=True,
local_dir=utils.get_save_dir(),
callbacks=[MyCallback()],
# resume="PROMPT",
verbose=1,
)
ray.shutdown()
# to see the progress run
# tensorboard --logdir=./save/tune_PPO_cartpole
@classmethod
def get_iter_by_col_id(cls, col_id):
found = False
iter = cls.tree_view_model.get_iter_first()
while iter:
(num, ) = cls.tree_view_model.get(iter, MapList.COL_ID)
if col_id == num:
found = True
break
iter = cls.tree_view_model.iter_next(iter)
if not found:
iter = None
return iter
@classmethod
def next_col_id(cls):
next_col_id = -1
iter = cls.tree_view_model.get_iter_first()
while iter:
(num, ) = cls.tree_view_model.get(iter, MapList.COL_ID)
if next_col_id < num:
next_col_id = num
iter = cls.tree_view_model.iter_next(iter)
return next_col_id + 1
MapList.load_all_from_dir(utils.get_save_dir())
# Set seed for all randomness sources
utils.seed(args.seed)
# Generate environment
envs = []
for i in range(args.procs):
env = gym.make(args.env)
env.seed(args.seed + 10000 * i)
envs.append(env)
env = ParallelEnv(envs)
# Define agent
save_dir = utils.get_save_dir(args.model)
agent = utils.Agent(save_dir, env.observation_space, args.argmax, args.procs)
print("CUDA available: {}\n".format(torch.cuda.is_available()))
# Initialize logs
logs = {"num_frames_per_episode": [], "return_per_episode": []}
# Run the agent
start_time = time.time()
obss = env.reset()
log_done_counter = 0
log_episode_return = torch.zeros(args.procs, device=agent.device)
problem, other_config = config["main_params"]
add_string = ""
if problem == "cartpole":
add_string = f"phi: {other_config.get('phi', 0)} template: {other_config.get('template', 0)} agent:{other_config['nn_path']}"
if problem == "bouncing_ball":
add_string = f"phi: {other_config.get('phi', 0)} template: {other_config.get('template', 0)} initial_state: {other_config.get('initial_state', 0)} agent:{other_config['nn_path']}"
if problem == "stopping_car":
add_string = f"phi: {other_config.get('phi', 0)} template: {other_config.get('template', 0)} agent:{other_config['nn_path']}"
return os.path.join(self.name, f"{problem} {add_string}")
if __name__ == '__main__':
ray.init(local_mode=os.environ.get("local_mode", False), log_to_driver=False, include_dashboard=True)
cpu = 8
trials = list(_iter())
n_trials = len(trials) - 1
print(f"Total n of trials: {n_trials}")
start_from = 0
name_group = NameGroup()
for i, (problem, other_config) in enumerate(trials):
if i < start_from:
continue
print(f"Starting trial: {i + 1}/{n_trials + 1}")
experiment_config = {
"main_params": (problem, other_config),
"n_workers": cpu
}
trial_dir = os.path.join(utils.get_save_dir(), "experiment_collection_NFM", name_group.trial_str_creator(experiment_config))
run_parameterised_experiment(config=experiment_config, trial_dir=trial_dir)
print(f"Finished trial: {i + 1}/{n_trials + 1}")
@classmethod def get_iter_by_col_id(cls, col_id): found = False iter = cls.tree_view_model.get_iter_first () while iter: (num,) = cls.tree_view_model.get (iter, MapList.COL_ID) if col_id == num: found = True break iter = cls.tree_view_model.iter_next (iter) if not found: iter = None return iter @classmethod def next_col_id(cls): next_col_id = -1 iter = cls.tree_view_model.get_iter_first () while iter: (num,) = cls.tree_view_model.get (iter, MapList.COL_ID) if next_col_id < num: next_col_id = num iter = cls.tree_view_model.iter_next (iter) return next_col_id + 1 MapList.load_all_from_dir(utils.get_save_dir ())
def main(args):
# set up logging and device
args.output = utils.get_save_dir(args.output, args.name, training=True)
logger = utils.get_logger(args.output, args.name)
tbx = SummaryWriter(args.output)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError("At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train:
raise ValueError("Output directory () already exists and is not empty.")
os.makedirs(args.output_dir, exist_ok=True)
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
# Generating the dictionaries
dep_dict, pos_dict, ent_dict, total_features = generate_dictionary(args.train_ling_features_file, args.eval_ling_features_file, args.test_ling_features_file)
# from IPython import embed; embed()
# Generating total_dictionary
total_dict = convert_string_features_to_array(total_features, dep_dict, pos_dict, ent_dict)
# Prepare model
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
global_step = 0
# load eval features
eval_examples = read_squad_examples(
input_file=args.predict_file, is_training=False, version_2_with_negative=args.version_2_with_negative, total_dictionary=total_dict)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
test_examples = read_squad_examples(
input_file=args.test_file, is_training=False, version_2_with_negative=args.version_2_with_negative, total_dictionary=total_dict)
test_features = convert_examples_to_features(
examples=test_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
# Save a trained model
"""
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
model_state_dict = torch.load(output_model_file)
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model, state_dict=model_state_dict)
else:
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model)
model.to(device)
"""
# load the best trained model and eval on the eval set and test set
models = []
base_path = '/home/zhangyue/results'
model_names = ['bert_base_linear', 'bert_base_linear_best', 'bert_large_1', 'bert_large_relu_1e-5']
best_model_file = os.path.join(base_path, model_name[0], "pytorch_model_best.bin")
model_state_dict = torch.load(best_model_file)
model = BertForQuestionAnsweringLing.from_pretrained('bert-base-cased', state_dict=model_state_dict)
model.to(device)
env = StoppingCar(config)
# env = CartPoleEnv() # gym.make("CartPole-v0")
env.seed(seed)
np.random.seed(seed)
state_size = 2
action_size = 2
STARTING_BETA = 0.6 # the higher the more it decreases the influence of high TD transitions
ALPHA = 0.6 # the higher the more aggressive the sampling towards high TD transitions
EPS_DECAY = 0.2
MIN_EPS = 0.01
agent = InvariantAgent(state_size=state_size,
action_size=action_size,
alpha=ALPHA)
agent.load(os.path.join(
utils.get_save_dir(),
"Aug05_16-14-33_alpha=0.6, min_eps=0.01, eps_decay=0.2/checkpoint_3000.pth"
),
invariant=False)
agent2 = InvariantAgent(state_size=state_size,
action_size=action_size,
alpha=ALPHA)
# agent2.load("/home/edoardo/Development/SafeDRL/runs/Aug20_09-16-25_invariant/checkpoint_1000.pth")
agent2.load(
os.path.join(utils.get_save_dir(),
"Aug20_11-58-57_invariant/checkpoint_1500.pth"))
agent_model = agent.qnetwork_local
invariant_model = agent2.inetwork_local
agent_model.cpu()
