def train(opt):
torch.manual_seed(123)
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
mp = _mp.get_context("spawn")
env, num_states, num_actions = create_train_env(opt.world, opt.stage,
opt.action_type)
global_model = ActorCritic(num_states, num_actions)
if opt.use_gpu:
global_model.cuda()
global_model.share_memory()
if opt.load_from_previous_stage:
if opt.stage == 1:
previous_world = opt.world - 1
previous_stage = 4
else:
previous_world = opt.world
previous_stage = opt.stage - 1
file_ = "{}/a3c_super_mario_bros_{}_{}".format(opt.saved_path,
previous_world,
previous_stage)
if os.path.isfile(file_):
global_model.load_state_dict(torch.load(file_))
optimizer = GlobalAdam(global_model.parameters(), lr=opt.lr)
local_train(0, opt, global_model, optimizer, True)
def train(opt):
torch.manual_seed(123)
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
mp = _mp.get_context("spawn")
global_model = ActorCritic(num_inputs=3, num_actions=90)
global_icm = IntrinsicCuriosityModule(num_inputs=3, num_actions=90)
if opt.use_gpu:
global_model.cuda()
global_icm.cuda()
global_model.share_memory()
global_icm.share_memory()
optimizer = GlobalAdam(list(global_model.parameters()) + list(global_icm.parameters()), lr=opt.lr)
processes = []
for index in range(opt.num_processes):
if index == 0:
process = mp.Process(target=local_train, args=(index, opt, global_model, global_icm, optimizer, True))
else:
process = mp.Process(target=local_train, args=(index, opt, global_model, global_icm, optimizer))
process.start()
processes.append(process)
for process in processes:
process.join()
def train(opt):
torch.manual_seed(123)
opt.log_path = opt.log_path + "/" + opt.exp
opt.saved_path = opt.saved_path + "/" + opt.exp
opt.output_path = opt.output_path + "/" + opt.exp
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
mp = _mp.get_context("spawn")
global_model = ActorCritic(num_inputs=3, num_actions=opt.num_actions)
global_icm = IntrinsicCuriosityModule(num_inputs=3, num_actions=opt.num_actions)
if opt.resume_path:
print("Load model from checkpoint: {}".format(opt.resume_path))
global_model.load_state_dict(torch.load("{}/a3c".format(opt.resume_path)))
global_icm.load_state_dict(torch.load("{}/icm".format(opt.resume_path)))
if opt.use_gpu:
global_model.cuda()
global_icm.cuda()
global_model.share_memory()
global_icm.share_memory()
optimizer = GlobalAdam(list(global_model.parameters()) + list(global_icm.parameters()), lr=opt.lr)
processes = []
for index in range(opt.num_processes):
if index == 0:
process = mp.Process(target=local_train, args=(index, opt, global_model, global_icm, optimizer, True))
else:
process = mp.Process(target=local_train, args=(index, opt, global_model, global_icm, optimizer))
process.start()
processes.append(process)
for process in processes:
process.join()
def train(opt):
torch.manual_seed(SEED)
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
mp = _mp.get_context("spawn")
env, num_states, num_actions = create_train_env(opt.world, opt.stage,
opt.action_type)
global_model = ActorCritic(num_states, num_actions)
global_model.share_memory()
if opt.load_from_previous_stage:
if opt.stage == 1:
previous_world = opt.world - 1
previous_stage = 4
else:
previous_world = opt.world
previous_stage = opt.stage - 1
file_ = f"{opt.saved_path}/a3c_super_mario_bros_{previous_world}_{previous_stage}"
if os.path.isfile(file_):
global_model.load_state_dict(torch.load(file_))
optimizer = GlobalAdam(global_model.parameters(), lr=opt.lr)
processes = []
for index in range(opt.num_processes):
if index == 0:
process = mp.Process(target=local_train,
args=(index, opt, global_model, optimizer,
True))
else:
process = mp.Process(target=local_train,
args=(index, opt, global_model, optimizer))
process.start()
processes.append(process)
process = mp.Process(target=local_test,
args=(opt.num_processes, opt, global_model))
process.start()
processes.append(process)
for process in processes:
process.join()
def shared_learn(args):
os.environ['OMP_NUM_THREADS'] = '1'
torch.manual_seed(123)
# create path for logs
if os.path.isdir(args.sum_path):
shutil.rmtree(args.sum_path)
os.makedirs(args.sum_path)
if not os.path.isdir(args.trained_models_path):
os.makedirs(args.trained_models_path)
mp = _mp.get_context('spawn')
# create initial mario environment
env, num_states, num_actions = build_environment(args.world, args.stage)
print('Num of states: {}'.format(num_states)) #4
print('environment: {}'.format(env))
print('Num of actions: {}'.format(num_actions)) #12
# check if cuda is available else cpu
device = torch.device('cuda' if (
args.use_cuda and torch.cuda.is_available()) else 'cpu')
CAE_shared_model = Convolutional_AutoEncoder() #.to(device)
A3C_shared_model = ActorCritic(num_states, num_actions) #.to(device)
# if a new stage, then it picks up previous saved model
if args.new_stage:
A3C_shared_model.load_state_dict(
torch.load('{}/a3c_super_mario_bros_{}_{}_enc2'.format(
args.world, args.stage, args.trained_models_path)))
A3C_shared_model.eval()
# GPU check
if (args.use_cuda and torch.cuda.is_available()):
A3C_shared_model.cuda()
CAE_shared_model.cuda()
# shares memory with worker instances
CAE_shared_model.share_memory()
A3C_shared_model.share_memory()
print('A3C')
print(A3C_shared_model)
# intialize optimizer
optimizer_cae = CAE_shared_model.createLossAndOptimizer(
CAE_shared_model, 0.001)
optimizer_a3c = SharedAdam(A3C_shared_model.parameters(), lr=args.lr)
#optimizer.share_memory()
# processes
workers = []
# start train process (run for the set number of workers)
for rank in range(args.num_processes):
if rank == 0:
worker = mp.Process(target=train_a3c,
args=(rank, args, optimizer_a3c,
A3C_shared_model, CAE_shared_model,
optimizer_cae, True))
else:
worker = mp.Process(target=train_a3c,
args=(rank, args, optimizer_a3c,
A3C_shared_model, CAE_shared_model,
optimizer_cae, True))
worker.start()
worker.append(worker)
# test worker
worker = mp.Process(target=test_a3c,
args=(rank, args, A3C_shared_model, CAE_shared_model))
worker.start()
workers.append(worker)
# join all processes
for worker in workers:
worker.join()