def train(self):
args = self.args
torch.manual_seed(args.seed)
env = env = grid2op.make(args.env_name,
test=args.for_test,
reward_class=L2RPNReward)
shared_model = ActorCritic(env.observation_space.size(),
self.action_space, args.hidden_size)
shared_model.share_memory()
if args.no_shared:
optimizer = None
else:
optimizer = my_optim.SharedAdam(shared_model.parameters(),
lr=args.lr)
optimizer.share_memory()
processes = []
counter = mp.Value('i', 0)
lock = mp.Lock()
p = mp.Process(target=self.do_test,
args=(args.num_processes, args, shared_model, counter))
p.start()
processes.append(p)
for rank in range(0, args.num_processes):
p = mp.Process(target=self.do_train,
args=(rank, args, shared_model, counter, lock,
optimizer))
p.start()
processes.append(p)
for p in processes:
p.join()
def __init__(self, args_, logger_):
self.args = args_
self.logger = logger_
self.env = AtariEnv(gym.make(self.args.game),
args_.frame_seq,
args_.frame_skip,
render=True)
self.shared_model = A3CLSTMNet(self.env.state_shape,
self.env.action_dim)
self.shared_model.share_memory()
self.optim = my_optim.SharedAdam(self.shared_model.parameters(),
lr=self.args.lr)
self.optim.share_memory()
# visdom
self.vis = visdom.Visdom()
self.main_update_step = Value('d', 0)
# load model
if self.args.load_weight != 0:
self.load_model(self.args.load_weight)
self.jobs = []
if self.args.t_flag:
for process_id in xrange(self.args.jobs):
job = A3CSingleProcess(process_id, self, logger_)
self.jobs.append(job)
self.test_win = None
def main():
#env
args = config()
mp.set_start_method("spawn")
os.environ['OMP_NUM_THREADS'] = '1'
os.environ['CUDA_VISIBLE_DEVICES'] = ""
env = create_atari_env(args.env_name)
shared_model = AcotrCritic(env.observation_space.shape[0],
env.action_space)
shared_model.share_memory()
optimizer = my_optim.SharedAdam(shared_model.parameters(), lr=args.lr)
optimizer.share_memory()
processes = []
counter = mp.Value('i', 0)
lock = mp.Lock()
p = mp.Process(target=test,
args=(args.num_processes, args, shared_model, counter,
"./log/"))
p.start()
processes.append(p)
for rank in range(0, args.num_processes):
p = mp.Process(target=train,
args=(rank, args, shared_model, counter, lock,
optimizer))
p.start()
processes.append(p)
for p in processes:
p.join()
class Params():
def __init__(self):
self.lr = 0.0001
self.gamma = 0.99
self.tau = 1.
self.seed = 1
self.num_processes = 16
self.num_steps = 20
self.max_episode_length = 10000
self.env_name = 'Breakout-v0'
# Main run
os.environ['OMP_NUM_THREADS'] = '1'
params = Params()
torch.manual_seed(params.seed)
env = create_atari_env(params.env_name)
shared_model = ActorCritic(env.observation_space.shape[0], env.action_space)
shared_model.share_memory()
optimizer = my_optim.SharedAdam(shared_model.parameters(), lr=params.lr)
optimizer.share_memory()
processes = []
p = mp.Process(target=test, args=(params.num_processes, params, shared_model))
p.start()
processes.append(p)
for rank in range(0, params.num_processes):
p = mp.Process(target=train, args=(rank, params, shared_model, optimizer))
p.start()
processes.append(p)
for p in processes:
p.join()
loader = DataLoader(opt) # not used in training procedure, just used to set vocab_size and seq_length
opt.vocab_size = loader.vocab_size
opt.seq_length = loader.seq_length
model = models.setup(opt)
model.train()
num_parameter = get_num_params(model)
print('number of parameters: ' + str(num_parameter))
if opt.async_opt:
if opt.use_cuda:
model.cuda()
model.share_memory()
optimizer = my_optim.SharedAdam(model.parameters(),
lr=opt.optim_lr,
betas=(opt.optim_adam_beta1, opt.optim_adam_beta2),
weight_decay=opt.optim_weight_decay)
optimizer.share_memory()
processes = []
for rank in range(opt.num_processes):
p = mp.Process(target=train, args=(rank, model, opt, optimizer))
p.start()
processes.append(p)
for p in processes:
p.join()
else:
if opt.use_cuda:
model.cuda()
rank = 0
optimizer = None
shared_model.share_memory()
if not args.no_curiosity:
# <---ICM---
shared_curiosity = IntrinsicCuriosityModule(
# env.observation_space.shape[0], env.action_space)
args.num_stack,
env.action_space)
shared_curiosity.share_memory()
# ---ICM--->
if args.no_shared:
optimizer = None
else:
if args.no_curiosity:
optimizer = my_optim.SharedAdam(shared_model.parameters(),
lr=args.lr)
elif not args.no_curiosity:
if not args.curiosity_only:
optimizer = my_optim.SharedAdam( # ICM
chain(shared_model.parameters(),
shared_curiosity.parameters()),
lr=args.lr)
elif args.curiosity_only:
optimizer = my_optim.SharedAdam(shared_curiosity.parameters(),
lr=args.lr)
optimizer.share_memory()
if (args.model_file is not None) and (args.optimizer_file is not None):
logging.info("Start with a pretrained model")
shared_model.load_state_dict(torch.load(args.model_file))
optimizer.load_state_dict(torch.load(args.optimizer_file))
def main(method):
args = built_parser(method=method)
env = gym.make(args.env_name)
state_dim = env.observation_space.shape
action_dim = env.action_space.shape[0]
args.state_dim = state_dim
args.action_dim = action_dim
action_high = env.action_space.high
action_low = env.action_space.low
args.action_high = action_high.tolist()
args.action_low = action_low.tolist()
args.seed = np.random.randint(0, 30)
args.init_time = time.time()
if args.alpha == 'auto' and args.target_entropy == 'auto':
delta_a = np.array(args.action_high, dtype=np.float32) - np.array(
args.action_low, dtype=np.float32)
args.target_entropy = -1 * args.action_dim #+ sum(np.log(delta_a/2))
Q_net1 = QNet(args)
Q_net1.train()
Q_net1.share_memory()
Q_net1_target = QNet(args)
Q_net1_target.train()
Q_net1_target.share_memory()
Q_net2 = QNet(args)
Q_net2.train()
Q_net2.share_memory()
Q_net2_target = QNet(args)
Q_net2_target.train()
Q_net2_target.share_memory()
actor1 = PolicyNet(args)
actor1.train()
actor1.share_memory()
actor1_target = PolicyNet(args)
actor1_target.train()
actor1_target.share_memory()
actor2 = PolicyNet(args)
actor2.train()
actor2.share_memory()
actor2_target = PolicyNet(args)
actor2_target.train()
actor2_target.share_memory()
Q_net1_target.load_state_dict(Q_net1.state_dict())
Q_net2_target.load_state_dict(Q_net2.state_dict())
actor1_target.load_state_dict(actor1.state_dict())
actor2_target.load_state_dict(actor2.state_dict())
Q_net1_optimizer = my_optim.SharedAdam(Q_net1.parameters(),
lr=args.critic_lr)
Q_net1_optimizer.share_memory()
Q_net2_optimizer = my_optim.SharedAdam(Q_net2.parameters(),
lr=args.critic_lr)
Q_net2_optimizer.share_memory()
actor1_optimizer = my_optim.SharedAdam(actor1.parameters(),
lr=args.actor_lr)
actor1_optimizer.share_memory()
actor2_optimizer = my_optim.SharedAdam(actor2.parameters(),
lr=args.actor_lr)
actor2_optimizer.share_memory()
log_alpha = torch.zeros(1, dtype=torch.float32, requires_grad=True)
log_alpha.share_memory_()
alpha_optimizer = my_optim.SharedAdam([log_alpha], lr=args.alpha_lr)
alpha_optimizer.share_memory()
share_net = [
Q_net1, Q_net1_target, Q_net2, Q_net2_target, actor1, actor1_target,
actor2, actor2_target, log_alpha
]
share_optimizer = [
Q_net1_optimizer, Q_net2_optimizer, actor1_optimizer, actor2_optimizer,
alpha_optimizer
]
experience_in_queue = []
experience_out_queue = []
for i in range(args.num_buffers):
experience_in_queue.append(Queue(maxsize=10))
experience_out_queue.append(Queue(maxsize=10))
shared_queue = [experience_in_queue, experience_out_queue]
step_counter = mp.Value('i', 0)
stop_sign = mp.Value('i', 0)
iteration_counter = mp.Value('i', 0)
shared_value = [step_counter, stop_sign, iteration_counter]
lock = mp.Lock()
procs = []
if args.code_model == "train":
for i in range(args.num_actors):
procs.append(
Process(target=actor_agent,
args=(args, shared_queue, shared_value,
[actor1, Q_net1], lock, i)))
for i in range(args.num_buffers):
procs.append(
Process(target=buffer,
args=(args, shared_queue, shared_value, i)))
procs.append(
Process(target=evaluate_agent,
args=(args, shared_value, share_net)))
for i in range(args.num_learners):
#device = torch.device("cuda")
device = torch.device("cpu")
procs.append(
Process(target=leaner_agent,
args=(args, shared_queue, shared_value, share_net,
share_optimizer, device, lock, i)))
elif args.code_model == "simu":
procs.append(Process(target=simu_agent, args=(args, shared_value)))
for p in procs:
p.start()
for p in procs:
p.join()
if args.planning:
d_module = load_d_module(env.action_space.shape[0], args)
shared_model = R_Module(env.action_space.shape[0],
args.dim,
discrete=args.discrete,
baseline=args.baseline,
state_space=env.observation_space.shape[0])
# shared reward module for everyone
shared_model.share_memory()
if args.no_shared:
optimizer = None
else:
optimizer = my_optim.SharedAdam(shared_model.parameters(),
lr=args.lr)
optimizer.share_memory()
processes = []
train_agent_method = None
total_args = args
train_agent_method = train_rewards
for rank in range(0, args.num_processes):
if rank == 0:
p = mp.Process(target=train_agent_method,
args=(rank, total_args, shared_model, enc,
optimizer, tb_log_dir, d_module))
else:
def main(method):
params = {
'obs_size': (160, 100), # screen size of cv2 window
'dt': 0.025, # time interval between two frames
'ego_vehicle_filter':
'vehicle.lincoln*', # filter for defining ego vehicle
'port': 2000, # connection port
'task_mode':
'Straight', # mode of the task, [random, roundabout (only for Town03)]
'code_mode': 'train',
'max_time_episode': 100, # maximum timesteps per episode
'desired_speed': 15, # desired speed (m/s)
'max_ego_spawn_times': 100, # maximum times to spawn ego vehicle
}
args = built_parser(method=method)
env = gym.make(args.env_name, params=params)
state_dim = env.state_space.shape
action_dim = env.action_space.shape[0]
args.state_dim = state_dim
args.action_dim = action_dim
action_high = env.action_space.high
action_low = env.action_space.low
args.action_high = action_high.tolist()
args.action_low = action_low.tolist()
args.seed = np.random.randint(0, 30)
args.init_time = time.time()
num_cpu = mp.cpu_count()
print(state_dim, action_dim, action_high, num_cpu)
if args.alpha == 'auto' and args.target_entropy == 'auto':
delta_a = np.array(args.action_high, dtype=np.float32) - np.array(
args.action_low, dtype=np.float32)
args.target_entropy = -1 * args.action_dim # + sum(np.log(delta_a/2))
Q_net1 = QNet(args)
Q_net1.train()
Q_net1.share_memory()
Q_net1_target = QNet(args)
Q_net1_target.train()
Q_net1_target.share_memory()
Q_net2 = QNet(args)
Q_net2.train()
Q_net2.share_memory()
Q_net2_target = QNet(args)
Q_net2_target.train()
Q_net2_target.share_memory()
actor1 = PolicyNet(args)
print("Network inited")
if args.code_model == "eval":
actor1.load_state_dict(
torch.load('./' + args.env_name + '/method_' + str(args.method) +
'/model/policy_' + str(args.max_train) + '.pkl'))
actor1.train()
actor1.share_memory()
actor1_target = PolicyNet(args)
actor1_target.train()
actor1_target.share_memory()
actor2 = PolicyNet(args)
actor2.train()
actor2.share_memory()
actor2_target = PolicyNet(args)
actor2_target.train()
actor2_target.share_memory()
print("Network set")
Q_net1_target.load_state_dict(Q_net1.state_dict())
Q_net2_target.load_state_dict(Q_net2.state_dict())
actor1_target.load_state_dict(actor1.state_dict())
actor2_target.load_state_dict(actor2.state_dict())
print("Network loaded!")
Q_net1_optimizer = my_optim.SharedAdam(Q_net1.parameters(),
lr=args.critic_lr)
Q_net1_optimizer.share_memory()
Q_net2_optimizer = my_optim.SharedAdam(Q_net2.parameters(),
lr=args.critic_lr)
Q_net2_optimizer.share_memory()
actor1_optimizer = my_optim.SharedAdam(actor1.parameters(),
lr=args.actor_lr)
actor1_optimizer.share_memory()
actor2_optimizer = my_optim.SharedAdam(actor2.parameters(),
lr=args.actor_lr)
actor2_optimizer.share_memory()
log_alpha = torch.zeros(1, dtype=torch.float32, requires_grad=True)
log_alpha.share_memory_()
alpha_optimizer = my_optim.SharedAdam([log_alpha], lr=args.alpha_lr)
alpha_optimizer.share_memory()
print("Optimizer done")
share_net = [
Q_net1, Q_net1_target, Q_net2, Q_net2_target, actor1, actor1_target,
actor2, actor2_target, log_alpha
]
share_optimizer = [
Q_net1_optimizer, Q_net2_optimizer, actor1_optimizer, actor2_optimizer,
alpha_optimizer
]
experience_in_queue = []
experience_out_queue = []
for i in range(args.num_buffers):
experience_in_queue.append(Queue(maxsize=10))
experience_out_queue.append(Queue(maxsize=10))
shared_queue = [experience_in_queue, experience_out_queue]
step_counter = mp.Value('i', 0)
stop_sign = mp.Value('i', 0)
iteration_counter = mp.Value('i', 0)
shared_value = [step_counter, stop_sign, iteration_counter]
lock = mp.Lock()
procs = []
if args.code_model == "train":
for i in range(args.num_learners):
if i % 2 == 0:
device = torch.device("cuda:1")
else:
device = torch.device("cuda:0")
# device = torch.device("cpu")
procs.append(
Process(target=leaner_agent,
args=(args, shared_queue, shared_value, share_net,
share_optimizer, device, lock, i)))
for i in range(args.num_actors):
procs.append(
Process(target=actor_agent,
args=(args, shared_queue, shared_value,
[actor1, Q_net1], lock, i)))
for i in range(args.num_buffers):
procs.append(
Process(target=buffer,
args=(args, shared_queue, shared_value, i)))
procs.append(
Process(target=evaluate_agent,
args=(args, shared_value, share_net)))
elif args.code_model == "simu":
procs.append(Process(target=simu_agent, args=(args, shared_value)))
for p in procs:
p.start()
for p in procs:
p.join()
self.seed = 1
self.num_processes = 16
self.num_steps = 20
self.max_episode_length = 10000
self.env_name = 'Breakout-v0'
# Main run
os.environ['OMP_NUM_THREADS'] = '1' # 1 thread per core
params = Params() # creating the params object from the Params class, that sets all the model parameters
torch.manual_seed(params.seed) # setting the seed (not essential)
env = create_atari_env(params.env_name) # we create an optimized environment thanks to universe
shared_model = ActorCritic(env.observation_space.shape[0],
env.action_space) # shared_model is the model shared by the different agents (different threads in different cores)
shared_model.share_memory() # storing the model in the shared memory of the computer, which allows the threads to have access to this shared memory even if they are in different cores
optimizer = my_optim.SharedAdam(shared_model.parameters(),
lr=params.lr) # the optimizer is also shared because it acts on the shared model
optimizer.share_memory() # same, we store the optimizer in the shared memory so that all the agents can have access to this shared memory to optimize the model
processes = [] # initializing the processes with an empty list
p = mp.Process(target=test, args=(params.num_processes, params,
shared_model)) # allowing to create the 'test' process with some arguments 'args' passed to the 'test' target function - the 'test' process doesn't update the shared model but uses it on a part of it - torch.multiprocessing.Process runs a function in an independent thread
p.start() # starting the created process p
processes.append(p) # adding the created process p to the list of processes
for rank in range(0,
params.num_processes): # making a loop to run all the other processes that will be trained by updating the shared model
p = mp.Process(target=train, args=(rank, params, shared_model, optimizer))
p.start()
processes.append(p)
for p in processes: # creating a pointer that will allow to kill all the threads when at least one of the threads, or main.py will be killed, allowing to stop the program safely
print('working')
p.join()
parser.add_argument('--test', action='store_true',
help='test ')
parser.add_argument('--feature', type=int, default=96,
help='features num')
if __name__ == '__main__':
args = parser.parse_args()
os.environ['OMP_NUM_THREADS'] = '1'
torch.manual_seed(args.seed)
num_inputs = args.feature
num_actions = 9
ac_net = ActorCritic(num_inputs, num_actions)
opt_ac = my_optim.SharedAdam(ac_net.parameters(), lr=args.lr)
if args.resume:
print("=> loading checkpoint ")
checkpoint = torch.load('../models/kankan/best.t7')
#args.start_epoch = checkpoint['epoch']
#best_prec1 = checkpoint['best_prec1']
ac_net.load_state_dict(checkpoint['state_dict'])
#opt_ac.load_state_dict(checkpoint['optimizer'])
print(ac_net)
print("=> loaded checkpoint (epoch {})"
.format(checkpoint['epoch']))
ac_net.share_memory()
#opt_ac = my_optim.SharedAdam(ac_net.parameters(), lr=args.lr)
opt_ac.share_memory()
operative_temp = [all_parameter[8]] + all_parameter[10:21]
cost_flex = all_parameter[2:8] + [all_parameter[9]]
state_num = all_parameter[10:109] + all_parameter[0:6] + [
all_parameter[7]
] + predictionFlat(params.file_path_prediction, (time_step_update) % 8760)
state = np.array(state_normalization(params.file_path_norm, state_num))
state = torch.from_numpy(state).float()
cx = torch.zeros(1, params.hidden_layer
) # the cell states of the LSTM are reinitialized to zero
hx = torch.zeros(1, params.hidden_layer
) # the hidden states of the LSTM are reinitialized to zero
model = ActorCritic(178, params.output_space)
optimizer = my_optim.SharedAdam(model.parameters(), lr=params.lr)
value, action_values, (hx, cx) = model(
(state.unsqueeze(0), (hx, cx))
) # getting from the model the output V(S) of the critic, the output Q(S,A) of the actor, and the new hidden & cell states
prob = F.softmax(
action_values, dim=1
) # generating a distribution of probabilities of the Q-values according to the softmax: prob(a) = exp(prob(a))/sum_b(exp(prob(b)))
log_prob = F.log_softmax(
action_values, dim=1
) # generating a distribution of log probabilities of the Q-values according to the log softmax: log_prob(a) = log(prob(a))
entropy = -(log_prob * prob).sum(1) # H(p) = - sum_x p(x).log(p(x))
action = prob.multinomial(
1
).data # selecting an action by taking a random draw from the prob distribution
log_prob = log_prob.gather(
self.env_name = 'Pendulum-v0'
if __name__ == '__main__':
os.environ['OMP_NUM_THREADS'] = '1'
params = Params()
torch.manual_seed(params.seed)
env = gym.make(params.env_name)
num_inputs = env.observation_space.shape[0]
num_outputs = env.action_space.shape[0]
shared_p = Policy(num_inputs, num_outputs)
shared_v = Value(num_inputs)
shared_p.share_memory()
shared_v.share_memory()
optimizer_p = my_optim.SharedAdam(shared_p.parameters(), lr=params.lr)
optimizer_v = my_optim.SharedAdam(shared_v.parameters(), lr=params.lr)
processes = []
p = mp.Process(target=test, args=(params.num_processes, params, shared_p))
p.start()
processes.append(p)
for rank in range(0, params.num_processes):
p = mp.Process(target=train,
args=(rank, params, shared_p, shared_v, optimizer_p,
optimizer_v))
p.start()
processes.append(p)
for p in processes:
p.join()
cx = torch.zeros(1, 256)
hx = torch.zeros(1, 256)
state = env.reset()
state = torch.from_numpy(state)
# <---ICM---
shared_curiosity = IntrinsicCuriosityModule2(args.num_stack,
env.action_space,
args.epsilon)
shared_curiosity.share_memory()
# ---ICM--->
if args.no_shared:
optimizer = None
else:
optimizer = my_optim.SharedAdam(shared_curiosity.parameters(),
lr=args.lr)
optimizer.share_memory()
if args.curiosity_file is not None:
logging.info("Load curiosity")
shared_curiosity.load_state_dict(torch.load(args.curiosity_file),
strict=False)
if args.optimizer_file is not None:
logging.info("Load optimizer")
optimizer.load_state_dict(torch.load(args.optimizer_file))
if args.new_curiosity:
logging.info("Bayesian curiosity")
processes = []