args.num_traj_buffers,
args.num_steps,
)
next_obs = share_memory(np.zeros(dimensions[:-1]+env.observation_space.shape, dtype=np.float32))
next_done = share_memory(np.zeros(dimensions[:-1], dtype=np.float32))
obs = share_memory(np.zeros(dimensions+env.observation_space.shape, dtype=np.float32))
actions = share_memory(np.zeros(dimensions+env.action_space.shape, dtype=env.action_space.dtype))
logprobs = share_memory(np.zeros(dimensions, dtype=np.float32))
rewards = share_memory(np.zeros(dimensions, dtype=np.float32))
dones = share_memory(np.zeros(dimensions, dtype=np.float32))
values = share_memory(np.zeros(dimensions, dtype=np.float32))
traj_availables = share_memory(np.ones(dimensions, dtype=np.float32))
actor_processes = []
policy_workers = []
stats_queue = MpQueue()
rollout_task_queues = [MpQueue() for i in range(args.num_rollout_workers)]
policy_request_queues = [MpQueue() for i in range(args.num_policy_workers)]
learner_request_queue = MpQueue()
data_process_back_queues = []
for i in range(args.num_rollout_workers):
actor = mp.Process(
target=act,
args=[[args, experiment_name, i, lock, stats_queue, 0,
next_obs, next_done, obs, actions, logprobs, rewards, dones, values, traj_availables,
rollout_task_queues[i], policy_request_queues, learner_request_queue]],
)
actor.start()
num_steps = 32
lock = mp.Lock()
dimensions = (
mp.cpu_count(),
num_envs,
1,
1,
num_steps,
)
# obs = share_memory_numpy(np.zeros(dimensions + (84,84,3)))
obs = share_memory_torch_numpy_mixed(np.zeros(dimensions + (84, 84, 3)), 5)
# raise
actor_processes = []
ctx = mp.get_context("forkserver")
stats_queue = MpQueue()
for i in range(num_cpus):
actor = mp.Process(
target=act,
args=[obs, num_envs, num_steps],
)
actor.start()
actor_processes.append(actor)
import timeit
timer = timeit.default_timer
existing_video_files = []
global_step = 0
global_step_increment = 0
start_time = time.time()
update_step = 0
try: