def test_single_impala_agent_functionality(self):
"""
Creates a single IMPALAAgent and runs it for a few steps in a DeepMindLab Env to test
all steps of the actor and learning process.
"""
try:
from rlgraph.environments.deepmind_lab import DeepmindLabEnv
except ImportError:
print("Deepmind Lab not installed: Will skip this test.")
return
agent_config = config_from_path(
"configs/impala_agent_for_deepmind_lab_env.json")
env_spec = dict(level_id="lt_hallway_slope",
observations=["RGB_INTERLEAVED", "INSTR"],
frameskip=4)
dummy_env = DeepmindLabEnv.from_spec(env_spec)
agent = SingleIMPALAAgent.from_spec(
default_dict(dict(type="single-impala-agent"), agent_config),
architecture="large",
environment_spec=default_dict(dict(type="deepmind-lab"), env_spec),
state_space=dummy_env.state_space,
action_space=dummy_env.action_space,
# TODO: automate this (by lookup from NN).
internal_states_space=IMPALAAgent.default_internal_states_space,
# Summarize time-steps to have an overview of the env-stepping speed.
summary_spec=dict(summary_regexp="time-step",
directory="/home/rlgraph/"),
dynamic_batching=False,
num_workers=4)
# Count items in the queue.
print("Items in queue: {}".format(
agent.call_api_method("get_queue_size")))
updates = 5
update_times = list()
print("Updating from queue ...")
for _ in range(updates):
start_time = time.monotonic()
agent.update()
update_times.append(time.monotonic() - start_time)
print(
"Updates per second (including waiting for enqueued items): {}/s".
format(updates / np.sum(update_times)))
time.sleep(5)
agent.terminate()
def test_isolated_impala_actor_agent_functionality(self):
"""
Creates a non-distributed IMPALAAgent (actor) and runs it for a few steps in a DeepMindLab Env to test
all steps of the learning process.
"""
try:
from rlgraph.environments.deepmind_lab import DeepmindLabEnv
except ImportError:
print("Deepmind Lab not installed: Will skip this test.")
return
agent_config = config_from_path(
"configs/impala_agent_for_deepmind_lab_env.json")
env_spec = dict(level_id="seekavoid_arena_01",
observations=["RGB_INTERLEAVED", "INSTR"],
frameskip=4)
dummy_env = DeepmindLabEnv.from_spec(env_spec)
agent = IMPALAAgent.from_spec(
agent_config,
type="actor",
architecture="large",
environment_spec=default_dict(dict(type="deepmind-lab"), env_spec),
state_space=dummy_env.state_space,
action_space=dummy_env.action_space,
# TODO: automate this (by lookup from NN).
internal_states_space=IMPALAAgent.default_internal_states_space,
execution_spec=dict(
#mode="distributed",
#distributed_spec=dict(job="actor", task_index=0, cluster_spec=self.cluster_spec_single_actor),
disable_monitoring=True),
# Need large queue to be able to fill it up (don't have a learner).
fifo_queue_spec=dict(capacity=10000))
# Start Specifiable Server with Env manually (monitoring is disabled).
agent.environment_stepper.environment_server.start_server()
time_start = time.perf_counter()
steps = 5
for _ in range(steps):
agent.call_api_method("perform_n_steps_and_insert_into_fifo")
time_total = time.perf_counter() - time_start
print(
"Done running {}x{} steps in Deepmind Lab env using IMPALA network in {}sec ({} actions/sec)."
.format(steps, agent.worker_sample_size, time_total,
agent.worker_sample_size * steps / time_total))
agent.environment_stepper.environment_server.stop_server()
agent.terminate()
