def evaluate_experiment(env_id, experiment_name, num_envs=96):
# fixed seeds
random.seed(0)
np.random.seed(0)
tf.random.set_random_seed(0)
params = AgentTMAX.Params(experiment_name)
params = params.load()
params.seed = 0
# for faster evaluation
params.num_envs = num_envs
params.num_workers = 32 if num_envs >= 32 else num_envs
def make_env_func():
e = create_env(env_id, skip_frames=True)
e.seed(0)
return e
agent = AgentTMAX(make_env_func, params)
agent.initialize()
rate, speed = 0, -1
multi_env = None
try:
multi_env = MultiEnv(
params.num_envs,
params.num_workers,
make_env_func=make_env_func,
stats_episodes=params.stats_episodes,
)
success, avg_speed = evaluate_locomotion_agent(agent, multi_env)
log.info('Finished evaluating experiment %s', experiment_name)
rate = np.mean(success)
speed = -1
avg_speed = [s for s in avg_speed if s > 0]
if len(avg_speed) > 0:
speed = np.mean(avg_speed)
log.info('Success rate %.1f%%, avg. speed %.2f edges/frame', rate * 100, speed)
except (Exception, KeyboardInterrupt, SystemExit):
log.exception('Interrupt...')
finally:
log.info('Closing env...')
if multi_env is not None:
multi_env.close()
agent.finalize()
return rate, speed
def test_dist_training(self):
t = Timing()
def make_env():
return make_doom_env(doom_env_by_name(TEST_ENV_NAME))
params = AgentTMAX.Params('__test_dist_train__')
params.distance_target_buffer_size = 1000
with t.timeit('generate_data'):
# first: generate fake random data
buffer = Buffer()
obs1 = np.full([84, 84, 3], 0, dtype=np.uint8)
obs1[:, :, 1] = 255
obs2 = np.full([84, 84, 3], 0, dtype=np.uint8)
obs2[:, :, 2] = 255
data_size = params.distance_target_buffer_size
for i in range(data_size):
same = i % 2 == 0
if same:
if random.random() < 0.5:
obs_first = obs_second = obs1
else:
obs_first = obs_second = obs2
else:
obs_first, obs_second = obs1, obs2
if random.random() < 0.5:
obs_second, obs_first = obs_first, obs_second
buffer.add(obs_first=obs_first,
obs_second=obs_second,
labels=0 if same else 1)
with t.timeit('init'):
agent = AgentTMAX(make_env, params)
agent.initialize()
params.distance_train_epochs = 1
params.distance_batch_size = 256
agent.distance.train(buffer, 1, agent)
with t.timeit('train'):
params.distance_train_epochs = 2
params.distance_batch_size = 64
agent.distance.train(buffer, 1, agent, t)
agent.finalize()
log.info('Timing: %s', t)
shutil.rmtree(params.experiment_dir())