def process__buffer(q_aggr, qs_dist, args, **_kwargs):
max_memo = args.max_memo
env_name = args.env_name
max_step = args.max_step
batch_size = args.batch_size
repeat_times = 2
# reward_scale = args.reward_scale
# gamma = args.gamma
'''init'''
env = gym.make(env_name)
state_dim, action_dim, max_action, target_reward, is_discrete = get_env_info(env, is_print=False)
buffer = BufferArray(max_memo, state_dim, action_dim) # experiment replay buffer
workers_num = len(qs_dist)
'''loop'''
is_training = True
while is_training:
for i in range(workers_num):
memo_array, is_solved = q_aggr.get()
buffer.extend_memo(memo_array)
if is_solved:
is_training = False
buffer.init_before_sample()
for i in range(max_step * repeat_times):
# batch_arrays = buffer.random_sample(batch_size, device=None) # faster but worse
for q_dist in qs_dist:
batch_arrays = buffer.random_sample(batch_size, device=None) # slower but better
q_dist.put(batch_arrays)
print('|| Exit: process__buffer')
def process__buffer(q_aggr, qs_dist, args, **_kwargs):
max_memo = args.max_memo
env_name = args.env_name
max_step = args.max_step
batch_size = args.batch_size
repeat_times = 2
reward_scale = args.reward_scale
gamma = args.gamma
'''init'''
env = gym.make(env_name)
state_dim, action_dim, max_action, target_reward = get_env_info(
env, be_quiet=False)
buffer = BufferArray(max_memo, state_dim,
action_dim) # experiment replay buffer
workers_num = len(qs_dist)
'''loop'''
with torch.no_grad(): # update replay buffer
# rewards, steps = agent.update_buffer(
# env, buffer, max_step, max_action, reward_scale, gamma)
rewards, steps = initial_exploration(env, buffer, max_step, max_action,
reward_scale, gamma, action_dim)
while True:
for _ in range(workers_num):
memo_array = q_aggr.get()
buffer.extend_memo(memo_array)
buffer.init_before_sample()
for _ in range(max_step * repeat_times):
for q_dist in qs_dist:
batch_arrays = buffer.random_sample(batch_size, device=None)
q_dist.put(batch_arrays)