def main(args):
"""main: Entry point."""
utils.prepare_dirs(args)
if args.num_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
if args.network_type == 'rnn':
pass
elif args.network_type == 'cnn':
dataset = data.image.Image(args, args.data_path)
else:
raise NotImplementedError(f"{args.dataset} is not supported")
trnr = trainer.Trainer(args, dataset)
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive()
else:
if not args.load_path:
raise Exception("[!] You should specify `load_path` to load a"
"pretrained model")
trnr.test()
def __init__(self, args):
self.args = args
self.device = args.device
self.start_iter = 1
self.train_iters = args.train_iters
# coeffs
self.lambda_A = args.lambda_A
self.lambda_B = args.lambda_B
self.lambda_idt = args.lambda_idt
self.dataloader_A, self.dataloader_B = get_dataloader(args)
self.D_B, self.G_AB = get_model(args)
self.D_A, self.G_BA = get_model(args)
self.criterion_GAN = GANLoss(use_lsgan=args.use_lsgan).to(args.device)
self.criterion_cycle = nn.L1Loss()
self.criterion_idt = nn.L1Loss()
self.optimizer_D = torch.optim.Adam(
itertools.chain(self.D_B.parameters(), self.D_A.parameters()),
lr=args.lr, betas=(args.beta1, args.beta2), weight_decay=args.weight_decay)
self.optimizer_G = torch.optim.Adam(
itertools.chain(self.G_AB.parameters(), self.G_BA.parameters()),
lr=args.lr, betas=(args.beta1, args.beta2), weight_decay=args.weight_decay)
self.logger = self.get_logger(args)
self.writer = SummaryWriter(args.log_dir)
save_args(args.log_dir, args)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', default=100, type=int, help='epoch number')
parser.add_argument('--start_epoch', default=0, type=int, help='start epoch number')
parser.add_argument('-b', '--batch_size', default=4, type=int, help='mini-batch size')
parser.add_argument('--lr', '--learning_rate', default=1e-4, type=float, help='initial learning rate')
parser.add_argument('--weight-decay', default=0.0, type=float, help='weight decay')
parser.add_argument('-c', '--continue', dest='continue_path', type=str, required=False)
parser.add_argument('--exp_name', default=config.exp_name, type=str, required=False)
parser.add_argument('--valid', action='store_true')
parser.add_argument('--style_loss', action='store_true')
args = parser.parse_args()
print(args)
config.exp_name = args.exp_name
config.make_dir()
save_args(args, config.log_dir)
net = network()
vgg = vgg_for_style_transfer()
net = torch.nn.DataParallel(net).cuda()
vgg = torch.nn.DataParallel(vgg).cuda()
sess = Session(config, net=net)
train_loader = get_dataloaders(os.path.join(config.data_dir, 'train.json'),
batch_size=args.batch_size, shuffle=True)
valid_loader = get_dataloaders(os.path.join(config.data_dir, 'val.json'),
batch_size=args.batch_size, shuffle=True)
if args.continue_path and os.path.exists(args.continue_path):
sess.load_checkpoint(args.continue_path)
clock = sess.clock
tb_writer = sess.tb_writer
criterion = nn.L1Loss().cuda()
optimizer = optim.Adam(sess.net.parameters(), args.lr, weight_decay=args.weight_decay)
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=10, verbose=True)
for e in range(args.epochs):
train_model(train_loader, sess.net, vgg,
criterion, optimizer, clock.epoch, tb_writer)
valid_out = valid_model(valid_loader, sess.net, vgg,
criterion, optimizer, clock.epoch, tb_writer)
tb_writer.add_scalar('train/learning_rate', optimizer.param_groups[-1]['lr'], clock.epoch)
scheduler.step(valid_out['epoch_loss'])
if valid_out['epoch_loss'] < sess.best_val_loss:
sess.best_val_loss = valid_out['epoch_loss']
sess.save_checkpoint('best_model.pth.tar')
if clock.epoch % 10 == 0:
sess.save_checkpoint('epoch{}.pth.tar'.format(clock.epoch))
sess.save_checkpoint('latest.pth.tar')
clock.tock()
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
utils.prepare_dirs(args)
torch.manual_seed(args.random_seed)
# Add this for the random seed
numpy.random.seed(args.random_seed)
random.seed(args.random_seed)
torch.backends.cudnn.deterministic = True
if args.num_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
if args.network_type == 'rnn':
dataset = data.text.Corpus(args.data_path)
trnr = trainer.Trainer(args, dataset)
elif 'cnn' in args.network_type:
dataset = data.image.Image(args)
trnr = trainer.CNNTrainer(args, dataset)
else:
raise NotImplementedError(f"{args.dataset} is not supported")
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive()
else:
if not args.load_path:
raise Exception("[!] You should specify `load_path` to load a "
"pretrained model")
trnr.test()
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
utils.prepare_dirs(args)
torch.manual_seed(args.random_seed)
if args.num_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
#args.network_type always 'cnn'
#args.dataset alwyas 'imagenet'
dataset = data.image.Image(args.data_path) #path of dataset
trnr = trainer.Trainer(args, dataset)
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive()
elif args.mode == 'test':
if not args.load_path:
raise Exception("[!] You should specify `load_path` to load a "
"pretrained model")
trnr.test()
elif args.mode == 'single':
if not args.dag_path:
raise Exception("[!] You should specify `dag_path` to load a dag")
utils.save_args(args)
trnr.train(single=True)
else:
raise Exception(f"[!] Mode not found: {args.mode}")
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
if args.mode == 'train':
logger = utils.get_logger(to_file=True)
else:
logger = utils.get_logger()
utils.prepare_dirs(args, logger)
torch.manual_seed(args.random_seed)
if args.num_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
if args.dataset != 'tumor':
raise NotImplementedError(f"{args.dataset} is not supported")
trnr = trainer.Trainer(args, logger)
if args.mode == 'train':
utils.save_args(args, logger)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive_final()
elif args.mode == 'single':
if not args.dag_path:
raise Exception("[!] You should specify `dag_path` to load a dag")
utils.save_args(args, logger)
trnr.train(single=True)
else:
raise Exception(f"[!] Mode not found: {args.mode}")
def test_load_save_args(self):
parser = argparse.ArgumentParser()
args = parser.parse_args(args=[])
args.__dict__ = {"name": "test", "foo": "bar"}
path = os.path.join(TMP, "args")
ensure_dir(path)
save_args(args, path)
args_loaded = load_args(path)
self.assertEqual(args, args_loaded)
def main():
input_dim = 6
spatial_dims = [0, 1, 2]
args = utils.read_args()
experiment_dir = utils.get_experiment_dir(args.name, args.run)
utils.initialize_experiment_if_needed(experiment_dir, args.evaluate)
# Logger will print to stdout and logfile
utils.initialize_logger(experiment_dir)
# Optionally restore arguments from previous training
# Useful if training is interrupted
if not args.evaluate:
try:
args = utils.load_args(experiment_dir)
except:
args.best_tpr = 0.0
args.nb_epochs_complete = 0 # Track in case training interrupted
utils.save_args(experiment_dir, args) # Save initial args
net = utils.create_or_restore_model(experiment_dir, args.nb_hidden,
args.nb_layer, input_dim, spatial_dims)
if torch.cuda.is_available():
net = net.cuda()
logging.warning("Training on GPU")
logging.info("GPU type:\n{}".format(torch.cuda.get_device_name(0)))
criterion = nn.functional.binary_cross_entropy
if not args.evaluate:
assert (args.train_file != None)
assert (args.val_file != None)
train_loader = construct_loader(args.train_file,
args.nb_train,
args.batch_size,
shuffle=True)
valid_loader = construct_loader(args.val_file, args.nb_val,
args.batch_size)
logging.info("Training on {} samples.".format(
len(train_loader) * args.batch_size))
logging.info("Validate on {} samples.".format(
len(valid_loader) * args.batch_size))
train(net, criterion, args, experiment_dir, train_loader, valid_loader)
# Perform evaluation over test set
try:
net = utils.load_best_model(experiment_dir)
logging.warning("\nBest model loaded for evaluation on test set.")
except:
logging.warning(
"\nCould not load best model for test set. Using current.")
assert (args.test_file != None)
test_loader = construct_loader(args.test_file, args.nb_test,
args.batch_size)
test_stats = evaluate(net, criterion, experiment_dir, args, test_loader,
TEST_NAME)
def main():
utils.display_args(args)
time.sleep(5)
# setting up reproducibility with selected seed
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# setting up the working directory and recording args
exp_tag = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
work_dir = os.path.join(args.work_dir, args.dataset, exp_tag)
os.makedirs("log", exist_ok=True)
log_path = os.path.join("log", exp_tag + ".log")
logger = get_logger(log_path)
res_list = []
config.DROPOUT = args.dropout
start = time.time()
for i in range(args.run):
logger.info("=" * 50 + f"Run {i+1}" + "=" * 50)
res = nas(logger)
res_list.append(res)
runnint_time = time.time() - start
result = pd.concat(res_list,
axis=0,
keys=[f"Run {i}" for i in range(len(res_list))])
if args.save is True:
os.makedirs(work_dir, exist_ok=False)
utils.save_args(args, os.path.join(work_dir, 'args.txt'))
file_name = os.path.join(work_dir, "result.csv")
result.to_csv(file_name, index=True)
logger.info(f"saving result to {file_name}")
fig_name = os.path.join(work_dir, "progress.png")
plt.figure()
avg_reward = result.groupby(
result.index.get_level_values(1))["reward"].mean()
plt.plot(avg_reward)
plt.title("Best reward in {} runs is {:.4f}".format(
args.run, result["reward"].max()))
logger.info(f"saving figure to {fig_name}")
plt.savefig(fig_name)
logger.info("Best reward in {} runs is {:.4f}".format(
args.run, result["reward"].max()))
device = 'cpu' if torch.cuda.is_available() is False else \
'cuda:{}'.format(args.gpu)
if args.evaluate is True:
test_result(args.dataset, res_list, device, logger, args.layers)
logger.info(f"Total running time {runnint_time}")
logger.info("*" * 50 + "End" + "*" * 50)
def train(
net,
criterion,
args,
experiment_dir,
train_loader,
valid_loader
):
optimizer = torch.optim.Adamax(net.parameters(), lr=args.lrate)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'max')
# Nb epochs completed tracked in case training interrupted
for i in range(args.nb_epochs_complete, args.nb_epoch):
# Update learning rate in optimizer
t0 = time.time()
logging.info("\nEpoch {}".format(i+1))
logging.info("Learning rate: {0:.3g}".format(args.lrate))
train_stats = train_one_epoch(net,
criterion,
optimizer,
args,
experiment_dir,
train_loader)
val_stats = evaluate(net, criterion, experiment_dir, args,
valid_loader, 'Valid')
utils.track_epoch_stats(i, args.lrate, 0, train_stats, val_stats, experiment_dir)
# Update learning rate, remaining nb epochs to train
scheduler.step(val_stats[0])
args.lrate = optimizer.param_groups[0]['lr']
args.nb_epochs_complete += 1
# Track best model performance
if (val_stats[0] > args.best_tpr):
logging.warning("Best performance on valid set.")
args.best_tpr = float(val_stats[0])
utils.update_best_plots(experiment_dir)
utils.save_best_model(experiment_dir, net)
utils.save_best_scores(i, val_stats[2], val_stats[0], val_stats[1], experiment_dir)
utils.save_epoch_model(experiment_dir, net)
utils.save_args(experiment_dir, args)
logging.info("Epoch took {} seconds.".format(int(time.time()-t0)))
if args.lrate < 10**-6:
logging.warning("Minimum learning rate reched.")
break
logging.warning("Training completed.")
def main(args):
utils.prepare_dirs(args)
utils.save_args(args)
dataset = get_dataset(args.dataset)
random_results = []
for step in range(args.controller_max_step):
actions = random_actions(args)
reward = get_reward(torch.LongTensor(actions), dataset)
random_results.append(reward)
top_scores = np.sort(list(set(random_results)))[-10:]
result_map = {'RandomSearch': top_scores}
df = pd.DataFrame(data=result_map)
path = os.path.join(args.model_dir, 'df.csv')
df.to_csv(path)
def main(args):
network_creator, env_creator = get_network_and_environment_creator(args)
utils.save_args(args, args.debugging_folder, file_name=ARGS_FILE)
logging.info('Saved args in the {0} folder'.format(args.debugging_folder))
logging.info(args_to_str(args))
batch_env = ConcurrentBatchEmulator(WorkerProcess, env_creator,
args.num_workers, args.num_envs)
set_exit_handler(concurrent_emulator_handler(batch_env))
try:
batch_env.start_workers()
learner = PAACLearner(network_creator, batch_env, args)
learner.set_eval_function(eval_network, learner.network, env_creator,
50, learner.use_rnn) # args to eval_network
learner.train()
finally:
batch_env.close()
def main(args):
env_creator = get_environment_creator(args)
network = create_network(args, env_creator.num_actions, env_creator.obs_shape)
utils.save_args(args, args.debugging_folder, file_name=ARGS_FILE)
logging.info('Saved args in the {0} folder'.format(args.debugging_folder))
logging.info(args_to_str(args))
#batch_env = SequentialBatchEmulator(env_creator, args.num_envs, init_env_id=1)
batch_env = ConcurrentBatchEmulator(WorkerProcess, env_creator, args.num_workers, args.num_envs)
set_exit_handler(concurrent_emulator_handler(batch_env))
try:
batch_env.start_workers()
learner = ParallelActorCritic(network, batch_env, args)
# evaluation results are saved as summaries of the training process:
learner.evaluate = lambda network: eval_network(network, env_creator, 10)
learner.train()
finally:
batch_env.close()
def main(args):
scores_rl = []
scores_randomforest = []
for _ in range(3):
utils.prepare_dirs(args)
utils.save_args(args)
# max_step = range(100,1000,100)
dataset = get_dataset(args.dataset)
maxscore_randomforest = randomforest(dataset)
scores_randomforest.append(maxscore_randomforest)
# scores_autosklearn = auto_sklearn(dataset)
# scores_tpot = tpot_(dataset)
trnr = trainer.Trainer(dataset,
args.n_tranformers,
args.n_scalers,
args.n_constructers,
args.n_selecters,
args.n_models,
args.lstm_size,
args.temperature,
args.tanh_constant,
args.save_dir,
func_names=func_names,
model_dir=args.model_dir,
log_step=args.log_step,
controller_max_step=args.controller_max_step)
bset = trnr.train_controller()
scores_rl.append(bset)
print(bset)
print(scores_rl)
# method_names = ['RandomForest', 'AutoSklearn', 'TPOT', 'RL']
# result_map = {'RandomForest': scores_randomforest, 'AutoSklearn':scores_autosklearn,
# 'TPOT': scores_tpot, 'RL': scores_rl}
result_map = {'RandomForest': scores_randomforest, 'RL': scores_rl}
df = pd.DataFrame(data=result_map)
path = os.path.join(args.model_dir, 'df.csv')
df.to_csv(path)
def main():
# load arguments
args, framework_args = get_args()
# set seed
np.random.seed(args.seed)
# build problem, get initial samples
problem, true_pfront, X_init, Y_init = build_problem(
args.problem, args.n_var, args.n_obj, args.n_init_sample,
args.n_process)
args.n_var, args.n_obj = problem.n_var, problem.n_obj
# initialize optimizer
optimizer = get_algorithm(args.algo)(problem, args.n_iter, args.ref_point,
framework_args)
# save arguments & setup logger
save_args(args, framework_args)
logger = setup_logger(args)
print(problem, optimizer, sep='\n')
# initialize data exporter
exporter = DataExport(optimizer, X_init, Y_init, args)
# optimization
solution = optimizer.solve(X_init, Y_init)
# export true Pareto front to csv
if true_pfront is not None:
exporter.write_truefront_csv(true_pfront)
for _ in range(args.n_iter):
# get new design samples and corresponding performance
X_next, Y_next = next(solution)
# update & export current status to csv
exporter.update(X_next, Y_next)
exporter.write_csvs()
# close logger
if logger is not None:
logger.close()
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
utils.prepare_dirs(args) #data_dir="./data/ptb"
torch.manual_seed(args.random_seed)
if args.num_gpu > 0:
#Sets the seed for generating random numbers for the current GPU. It’s safe to call this function if CUDA is not available; in that case, it is silently ignored.
torch.cuda.manual_seed(args.random_seed)
if args.network_type == 'rnn':
dataset = data.text.Corpus(args.data_path) #将文本数据读入字典，生成词对应的序号的Tensor
elif args.dataset == 'cifar':
dataset = data.image.Image(args.data_path)
else:
raise NotImplementedError("{}is not supported".format(args.dataset))
trnr = trainer.Trainer(args, dataset)
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", (
"`--load_path` should be given in `derive` mode")
trnr.derive()
elif args.mode == 'test':
if not args.load_path:
raise Exception(
"[!] You should specify `load_path` to load a pretrained model"
)
trnr.test()
elif args.mode == 'single':
if not args.dag_path:
raise Exception("[!] You should specify `dag_path` to load a dag")
utils.save_args(args)
trnr.train(single=True)
else:
raise Exception("[!] Mode not found: {}".format(args.mode))
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
utils.prepare_dirs(args)
torch.manual_seed(args.random_seed)
if args.num_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
if args.network_type == 'rnn':
dataset = data.text.Corpus(args.data_path)
elif args.dataset == 'cifar':
#dataset = data.image.Image(args.data_path)
dataset = data.image.Image(args)
else:
raise NotImplementedError(f"{args.dataset} is not supported")
trnr = trainer.Trainer(args, dataset)
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive()
elif args.mode == 'test':
if not args.load_path:
raise Exception("[!] You should specify `load_path` to load a "
"pretrained model")
trnr.test()
elif args.mode == 'single':
if not args.dag_path:
raise Exception("[!] You should specify `dag_path` to load a dag")
utils.save_args(args)
trnr.train(single=True)
else:
raise Exception(f"[!] Mode not found: {args.mode}")
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
utils.prepare_dirs(args)
torch.manual_seed(args.random_seed)
trnr = trainer.Trainer(args)
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive()
elif args.mode == 'single':
if not args.dag_path:
raise Exception("[!] You should specify `dag_path` to load a dag")
utils.save_args(args)
trnr.train(single=True)
else:
raise Exception(f"[!] Mode not found: {args.mode}")
def main(args):
tt_logger = TestTubeLogger(save_dir=args.log_path,
name="",
description=args.description,
debug=False,
create_git_tag=args.git_tag)
tt_logger.experiment
log_dir = Path(tt_logger.save_dir) / f"version_{tt_logger.version}"
checkpoint_dir = log_dir / "checkpoints"
os.makedirs(checkpoint_dir, exist_ok=True)
chkpt_callback = ModelCheckpoint(
checkpoint_dir,
# monitor='Loss/val_loss',
save_last=True,
# mode='min',
# save_top_k=1,
period=5)
data_loader = CustomDataLoader.from_argparse_args(args, )
model = Pipeline.from_argparse_args(args, )
save_args(args, log_dir)
trainer = Trainer.from_argparse_args(
args,
logger=tt_logger,
checkpoint_callback=chkpt_callback,
# early_stop_callback=False,
weights_summary='full',
gpus=1,
profiler=True,
)
trainer.fit(model, data_loader)
def main(args): # pylint:disable=redefined-outer-name
"""main: Entry point."""
utils.prepare_dirs(args)
torch.manual_seed(args.random_seed)
if args.num_gpu > 0:
torch.cuda.manual_seed(args.random_seed)
if args.network_type == 'rnn':
assert args.network_type in ('rnn', 'cnn')
dataset = data.text.Corpus(args.data_path)
elif args.dataset == 'cifar':
args.network_type in ('rnn', 'cnn')
dataset = data.image.Image(args)
else:
args.network_type not in ('rnn', 'cnn')
dataset = data.tabular.Tabular(args)
if args.network_type in ('rnn', 'cnn'):
trnr = trainer.Trainer(args, dataset)
else:
trnr = trainer2.Trainer(args, dataset)
if args.mode == 'train':
utils.save_args(args)
trnr.train()
elif args.mode == 'derive':
assert args.load_path != "", ("`--load_path` should be given in "
"`derive` mode")
trnr.derive()
else:
if not args.load_path:
raise Exception("[!] You should specify `load_path` to load a "
"pretrained model")
trnr.test()
def main(args):
tt_logger = TestTubeLogger(save_dir=args.log_path,
name="",
description=args.description,
create_git_tag=args.git_tag,
debug=args.debug)
tt_logger.experiment
log_dir = Path(tt_logger.save_dir) / f"version_{tt_logger.version}"
checkpoint_dir = log_dir / "checkpoints"
os.makedirs(checkpoint_dir, exist_ok=True)
chkpt_callback = ModelCheckpoint(
checkpoint_dir,
monitor='Loss/val_loss/data_loader_idx_2',
save_last=True,
mode='min',
save_top_k=10,
)
data_module = ReIDDataModule.from_argparse_args(args)
model = ST_ReID(data_module.num_classes,
learning_rate=args.learning_rate,
criterion=args.criterion,
rerank=args.rerank)
save_args(args, log_dir)
trainer = Trainer.from_argparse_args(args,
logger=[tt_logger],
checkpoint_callback=chkpt_callback,
profiler=True)
trainer.fit(model, data_module)
trainer.test(model)
def experiment(config):
"""
A function that runs an experiment.
args
config (dict) hyperparameters and experiment setup
"""
with tf.Session() as sess:
seed = config.pop('seed')
if seed:
seed = int(seed)
random.seed(seed)
tf.set_random_seed(seed)
np.random.seed(seed)
env_id = config.pop('env_id')
LOGGER.info('using {} env'.format(env_id))
env = gym.make(env_id)
global_rewards = []
global_step, episode = 0, 0
config['env'] = env
config['env_repr'] = repr(env)
config['sess'] = sess
render = int(config.pop('render'))
agent = Agent(**config)
rl_writer = tf.summary.FileWriter('./results/rl')
save_args(config, 'results/args.txt')
while global_step < config['total_steps']:
episode += 1
done = False
rewards, actions = [], []
observation = env.reset()
while not done:
global_step += 1
# if episode % 1 == render:
env.render()
action = agent.act(observation)
next_observation, reward, done, info = env.step(action)
agent.remember(observation, action, reward, next_observation,
done)
train_info = agent.learn()
rewards.append(reward)
actions.append(action)
observation = next_observation
ep_rew = sum(rewards)
global_rewards.append(ep_rew)
avg_reward = sum(global_rewards[-100:]) / len(
global_rewards[-100:])
if episode % 10 == 0:
log_str = ' step {:.0f} ep {:.0f} reward {:.1f} avg {:.1f}'
logging.info(
log_str.format(global_step, episode, ep_rew, avg_reward))
summary = tf.Summary(value=[
tf.Summary.Value(tag='episode_reward', simple_value=ep_rew)
])
rl_writer.add_summary(summary, episode)
avg_sum = tf.Summary(value=[
tf.Summary.Value(tag='avg_last_100_ep',
simple_value=avg_reward)
])
rl_writer.add_summary(avg_sum, episode)
rl_writer.flush()
return config
help='weight decay')
parser.add_argument('--adam-beta1',
type=float,
default=0.9,
metavar='',
help='Beta 1 parameter for Adam')
parser.add_argument('--adam-beta2',
type=float,
default=0.999,
metavar='',
help='Beta 2 parameter for Adam')
args = parser.parse_args()
random.seed(args.manual_seed)
torch.manual_seed(args.manual_seed)
utils.save_args(args)
print('\n\n****** Creating {} model ******'.format(args.net_type))
setup = Model(args)
print("model created successfully!")
print('\n\n****** Preparing {} dataset *******'.format(args.dataset_train))
dataloader = Dataloader(args, setup.input_size)
loader_train, loader_test = dataloader.create()
print('data prepared successfully!')
# initialize model:
if args.resume is None:
model = setup.model
model.apply(utils.weights_init)
train = setup.train
test = setup.test
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# data
parser.add_argument('--imsize', type=int, default=64)
parser.add_argument('--n_workers', type=int, default=2)
# model
parser.add_argument('--nz', type=int, default=100)
parser.add_argument('--nc', type=int, default=3)
parser.add_argument('--ngf', type=int, default=64)
parser.add_argument('--ndf', type=int, default=64)
# training
parser.add_argument('--device', type=str, default='cuda:0')
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--epochs', type=int, default=100)
parser.add_argument('--lr', type=float, default=0.0002)
parser.add_argument('--beta1', type=float, default=0.5)
parser.add_argument('--beta2', type=float, default=0.999)
# log
parser.add_argument('--log_freq', type=int, default=100)
parser.add_argument('--output_dir', type=str, default='./results')
args, unknown_args = parser.parse_known_args()
args.output_dir = './results/exp1.0'
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
save_args(args.output_dir, args)
Experiment(args).run()
def main(args):
device = gpu_setup(args.gpu)
append_args = ["dataset", "model"]
checkpoint_dir = savedir_setup(args.savedir,
args=args,
append_args=append_args,
basedir=args.saveroot)
args.githash = check_githash()
save_args(checkpoint_dir, args)
dataloader = setup_dataloader(
name=args.dataset,
batch_size=args.batch,
num_workers=args.workers,
)
dataset_size = len(dataloader.dataset)
print("number of images (dataset size): ", dataset_size)
model = setup_model(args.model,
dataset_size=dataset_size,
resume=args.resume,
biggan_imagenet_pretrained_model_path=args.pretrained)
model.eval()
#this has to be eval() even if it's training time
#because we want to fix batchnorm running mean and var
#still tune batchnrom scale and bias that is generated by linear layer in biggan
optimizer, scheduler = setup_optimizer(
model,
lr_g_linear=args.lr_g_l,
lr_g_batch_stat=args.lr_g_batch_stat,
lr_bsa_linear=args.lr_bsa_l,
lr_embed=args.lr_embed,
lr_class_cond_embed=args.lr_c_embed,
step=args.step,
step_facter=args.step_facter,
)
criterion = AdaBIGGANLoss(
scale_per=args.loss_per,
scale_emd=args.loss_emd,
scale_reg=args.loss_re,
normalize_img=args.loss_norm_img,
normalize_per=args.loss_norm_per,
dist_per=args.loss_dist_per,
)
#start trainig loop
losses = AverageMeter()
print_freq = args.print_freq
eval_freq = args.eval_freq
save_freq = eval_freq
max_iteration = args.iters
log = {}
log["log"] = []
since = time.time()
iteration = 0
epoch = 0
#prepare model and loss into device
model = model.to(device)
criterion = criterion.to(device)
while (True):
# Iterate over dataset (one epoch).
for data in dataloader:
img = data[0].to(device)
indices = data[1].to(device)
scheduler.step()
#embeddings (i.e. z) + noise (i.e. epsilon)
embeddings = model.embeddings(indices)
embeddings_eps = torch.randn(embeddings.size(),
device=device) * 0.01
#see https://github.com/nogu-atsu/SmallGAN/blob/f604cd17516963d8eec292f3faddd70c227b609a/gen_models/ada_generator.py#L29
#forward
img_generated = model(embeddings + embeddings_eps)
loss = criterion(img_generated, img, embeddings,
model.linear.weight)
losses.update(loss.item(), img.size(0))
#compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
if iteration % print_freq == 0:
temp = "train loss: %0.5f " % loss.item()
temp += "| smoothed loss %0.5f " % losses.avg
log["log"].append({
"iteration": iteration,
"epoch": epoch,
"loss": losses.avg
})
print(iteration, temp)
losses = AverageMeter()
if iteration % eval_freq == 0 and iteration > 0:
out_path = os.path.join(checkpoint_dir,
"%d_recon.jpg" % iteration)
generate_samples(model, out_path, dataloader.batch_size)
if iteration % save_freq == 0 and iteration > 0:
save_checkpoint(checkpoint_dir,
device,
model,
iteration=iteration)
if iteration > max_iteration:
break
iteration += 1
if iteration > max_iteration:
break
epoch += 1
log_save_path = os.path.join(checkpoint_dir, "train-log.json")
save_json(log, log_save_path)
def main():
args = parse_args()
os.environ["HDF5_USE_FILE_LOCKING"] = "FALSE"
torch.manual_seed(args.seed)
if torch.cuda.is_available():
device = 'cuda'
torch.cuda.manual_seed(args.seed)
else:
device = 'cpu'
print(f"==> Using device: {device}")
if args.checkpoint is None:
time_stamp = str(datetime.datetime.now().strftime('-%Y%m%d%H%M%S'))
args.checkpoint = args.model + time_stamp
args.checkpoint = 'checkpoints/' + args.checkpoint
if not os.path.isdir(args.checkpoint):
mkdir_p(args.checkpoint)
save_args(args)
logger = Logger(os.path.join(args.checkpoint, 'log.txt'),
title="ModelNet" + args.model)
logger.set_names([
"Epoch-Num", 'Learning-Rate', 'Train-Loss', 'Train-acc-B',
'Train-acc', 'Valid-Loss', 'Valid-acc-B', 'Valid-acc'
])
print('==> Preparing data..')
train_loader = DataLoader(ModelNet40(partition='train',
num_points=args.num_points),
num_workers=8,
batch_size=args.batch_size,
shuffle=True,
drop_last=True)
test_loader = DataLoader(ModelNet40(partition='test',
num_points=args.num_points),
num_workers=8,
batch_size=args.batch_size,
shuffle=True,
drop_last=False)
# Model
print('==> Building model..')
net = models.__dict__[args.model]()
criterion = nn.CrossEntropyLoss().to(device)
net = net.to(device)
# criterion = criterion.to(device)
if device == 'cuda':
net = torch.nn.DataParallel(net)
cudnn.benchmark = True
optimizer = torch.optim.SGD(net.parameters(),
lr=args.learning_rate,
momentum=0.9,
weight_decay=args.weight_decay)
scheduler = CosineAnnealingLR(optimizer,
args.epoch,
eta_min=args.learning_rate / 100)
best_test_acc = 0. # best test accuracy
best_train_acc = 0.
best_test_acc_avg = 0.
best_train_acc_avg = 0.
best_test_loss = float("inf")
best_train_loss = float("inf")
start_epoch = 0 # start from epoch 0 or last checkpoint epoch
for epoch in range(start_epoch, args.epoch):
print('Epoch(%d/%s) Learning Rate %s:' %
(epoch + 1, args.epoch, optimizer.param_groups[0]['lr']))
train_out = train(net, train_loader, optimizer, criterion,
device) # {"loss", "acc", "acc_avg", "time"}
test_out = validate(net, test_loader, criterion, device)
scheduler.step()
if test_out["acc"] > best_test_acc:
best_test_acc = test_out["acc"]
is_best = True
else:
is_best = False
best_test_acc = test_out["acc"] if (
test_out["acc"] > best_test_acc) else best_test_acc
best_train_acc = train_out["acc"] if (
train_out["acc"] > best_train_acc) else best_train_acc
best_test_acc_avg = test_out["acc_avg"] if (
test_out["acc_avg"] > best_test_acc_avg) else best_test_acc_avg
best_train_acc_avg = train_out["acc_avg"] if (
train_out["acc_avg"] > best_train_acc_avg) else best_train_acc_avg
best_test_loss = test_out["loss"] if (
test_out["loss"] < best_test_loss) else best_test_loss
best_train_loss = train_out["loss"] if (
train_out["loss"] < best_train_loss) else best_train_loss
save_model(net,
epoch,
path=args.checkpoint,
acc=test_out["acc"],
is_best=is_best)
logger.append([
epoch, optimizer.param_groups[0]['lr'], train_out["loss"],
train_out["acc_avg"], train_out["acc"], test_out["loss"],
test_out["acc_avg"], test_out["acc"]
])
print(
f"Training loss:{train_out['loss']} acc_avg:{train_out['acc_avg']} acc:{train_out['acc']} time:{train_out['time']}s)"
)
print(
f"Testing loss:{test_out['loss']} acc_avg:{test_out['acc_avg']} acc:{test_out['acc']}% time:{test_out['time']}s) \n\n"
)
logger.close()
print(f"++++++++" * 2 + "Final results" + "++++++++" * 2)
print(
f"++ Last Train time: {train_out['time']} | Last Test time: {test_out['time']} ++"
)
print(
f"++ Best Train loss: {best_train_loss} | Best Test loss: {best_test_loss} ++"
)
print(
f"++ Best Train acc_B: {best_train_acc_avg} | Best Test acc_B: {best_test_acc_avg} ++"
)
print(
f"++ Best Train acc: {best_train_acc} | Best Test acc: {best_test_acc} ++"
)
print(f"++++++++" * 5)
parser.add_argument('-f', '--form_terms', nargs="+", default=['edges'], help='Formation network statistics.')
parser.add_argument('-d', '--diss_terms', nargs="+", default=['edges'], help='Dissolution network statistics.')
# parser.add_argument('--data_format', choices=['real', 'synthetic'], default='synthetic',
# help='The format of input data. Real data is a list of adj matrices while synthetic'
# 'data is a list with nw, terms, and extra elements.')
parser.set_defaults()
# Get commands from command line
args = parser.parse_args()
np.seterr(all='raise')
# Get the input and output filenames
output_dir = args.outdir + ('' if args.outdir.endswith('/') else '/')
args_dir = utils.make_directory(output_dir, 'args')
utils.save_args(args, args_dir + 'args.txt')
input_data = args.data_name + ('' if args.data_name.endswith('.rds') else '.rds')
H_outfile = output_dir + args.data_name + "_H.txt"
y_outfile = output_dir + args.data_name + "_y.txt"
# load the data
data = readRDS("../data/" + input_data)
data = np.array(data).astype(int)
t = len(data)
n = len(data[0])
p = len(args.form_terms) + len(args.diss_terms)
print(f"time series length: {t}")
print(f"network size: {n} x {n}")
print(f"statistics dimension: {p}")
experiment_dir = args.outdir + ('' if args.outdir.endswith('/') else '/')
with open(experiment_dir + 'args/args.txt', 'r') as f:
sim_args_str = f.read()
sim_args = ast.literal_eval(sim_args_str)
p = len(sim_args['form_terms']) + len(sim_args['diss_terms'])
H = np.loadtxt(experiment_dir + sim_args['data_name'] + '_H.txt')
y = np.loadtxt(experiment_dir + sim_args['data_name'] + '_y.txt')
t = H.shape[0]
H = H.reshape((t, -1, p)) # t x n^2(E) x p
n = np.sqrt(H.shape[1]).astype(int)
print(f"Data has dimension (t, n, p): ({t}, {n}, {p})")
# Get the output filenames
result_dir = utils.make_directory(experiment_dir, 'results')
args_dir = utils.make_directory(experiment_dir, 'args')
utils.save_args(args, args_dir + 'args_model.txt')
theta_outfile = result_dir + 'theta_' + sim_args['data_name'] + ".txt"
u_outfile = result_dir + 'u_' + sim_args['data_name'] + ".txt"
z_outfile = result_dir + 'z_' + sim_args['data_name'] + ".txt"
theta_plot_dir = result_dir + 'est_theta_diff.png'
print('Initialize STERGM model...')
model = mple_learn.STERGMGraph(
lam=args.lam,
admm_alpha=args.admm_alpha,
rel_tol=args.rel_tol,
max_steps=args.max_steps,
newton_max_steps=args.max_steps_newton,
converge_tol=args.conv_tol,
gd_lr=args.gd_lr,
type=float,
default=0.8,
help='probability bounding factor')
parser.add_argument('--lr_size',
type=int,
default=8,
help='Policy Network Image Size')
parser.add_argument('--test_interval',
type=int,
default=5,
help='At what epoch to test the model')
args = parser.parse_args()
if not os.path.exists(args.cv_dir):
os.system('mkdir ' + args.cv_dir)
utils.save_args(__file__, args)
def train(epoch):
agent.train()
rnet.train()
matches, rewards, rewards_baseline, policies = [], [], [], []
for batch_idx, (inputs, targets) in tqdm.tqdm(enumerate(trainloader),
total=len(trainloader)):
inputs, targets = Variable(inputs), Variable(targets).cuda(async=True)
if not args.parallel:
inputs = inputs.cuda()
# Get the low resolution agent images
#parser.add_argument('--wide', default = 2, type = int, help = 'using wider resnet18')
args = parser.parse_args()
DEVICE = torch.device('cuda:{}'.format(args.d))
if args.exp is None:
cur_dir = os.path.realpath('./')
args.exp = cur_dir.split(os.path.sep)[-1]
log_dir = os.path.join('../../logs', args.exp)
exp_dir = os.path.join('../../exps', args.exp)
train_res_path = os.path.join(exp_dir, 'train_results.txt')
val_res_path = os.path.join(exp_dir, 'val_results.txt')
final_res_path = os.path.join(exp_dir, 'final_results.txt')
if not os.path.exists(exp_dir):
os.mkdir(exp_dir)
save_args(args, exp_dir)
writer = SummaryWriter(log_dir)
clock = TrainClock()
learning_rate_policy = [[5, 0.01], [3, 0.001], [2, 0.0001]]
get_learing_rate = MultiStageLearningRatePolicy(learning_rate_policy)
def adjust_learning_rate(optimizer, epoch):
#global get_lea
lr = get_learing_rate(epoch)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
