def test_video_callable_records_videos():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
env = Monitor(env, temp)
env.reset()
env.close()
results = monitoring.load_results(temp)
assert len(results['videos']) == 1, "Videos: {}".format(results['videos'])
def test_semisuper_succeeds():
"""Regression test. Ensure that this can write"""
with helpers.tempdir() as temp:
env = gym.make('SemisuperPendulumDecay-v0')
env = Monitor(env, temp)
env.reset()
env.step(env.action_space.sample())
env.close()
def test_monitor_filename():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
env = Monitor(env, directory=temp)
env.close()
manifests = glob.glob(os.path.join(temp, '*.manifest.*'))
assert len(manifests) == 1
def test_video_callable_false_does_not_record():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
env = Monitor(env, temp, video_callable=False)
env.reset()
env.close()
results = monitoring.load_results(temp)
assert len(results['videos']) == 0
def test_video_callable_true_not_allowed():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
try:
env = Monitor(env, temp, video_callable=True)
except error.Error:
pass
else:
assert False
def test_text_envs():
env = gym.make('FrozenLake-v0')
video = VideoRecorder(env)
try:
env.reset()
video.capture_frame()
video.close()
finally:
os.remove(video.path)
def test_semisuper_true_rewards():
env = gym.make('SemisuperPendulumNoise-v0')
env.reset()
observation, perceived_reward, done, info = env.step(
env.action_space.sample())
true_reward = info['true_reward']
# The noise in the reward should ensure these are different. If we get spurious errors, we can remove this check
assert perceived_reward != true_reward
def test_record_simple():
env = gym.make("CartPole-v1")
rec = VideoRecorder(env)
env.reset()
rec.capture_frame()
rec.close()
assert not rec.empty
assert not rec.broken
assert os.path.exists(rec.path)
f = open(rec.path)
assert os.fstat(f.fileno()).st_size > 100
def test_write_upon_reset_false():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
env = Monitor(env, directory=temp, video_callable=False, write_upon_reset=False)
env.reset()
files = glob.glob(os.path.join(temp, '*'))
assert not files, "Files: {}".format(files)
env.close()
files = glob.glob(os.path.join(temp, '*'))
assert len(files) > 0
def test_write_upon_reset_true():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
# TODO: Fix Cartpole to not configure itself automatically
# assert not env._configured
env = Monitor(env, directory=temp, video_callable=False, write_upon_reset=True)
env.configure()
env.reset()
files = glob.glob(os.path.join(temp, '*'))
assert len(files) > 0, "Files: {}".format(files)
env.close()
files = glob.glob(os.path.join(temp, '*'))
assert len(files) > 0
def test_steps_limit_restart():
with helpers.tempdir() as temp:
env = gym.make('test.StepsLimitCartpole-v0')
env = Monitor(env, temp, video_callable=False)
env.reset()
# Episode has started
_, _, done, info = env.step(env.action_space.sample())
assert done == False
# Limit reached, now we get a done signal and the env resets itself
_, _, done, info = env.step(env.action_space.sample())
assert done == True
assert env.episode_id == 1
env.close()
def test():
benchmark = registration.Benchmark(id='MyBenchmark-v0',
scorer=scoring.ClipTo01ThenAverage(),
tasks=[{
'env_id': 'CartPole-v0',
'trials': 1,
'max_timesteps': 5
}, {
'env_id': 'CartPole-v0',
'trials': 1,
'max_timesteps': 100,
}])
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
env = wrappers.Monitor(env, directory=temp, video_callable=False)
env.seed(0)
env.set_monitor_mode('evaluation')
rollout(env)
env.set_monitor_mode('training')
for i in range(2):
rollout(env)
env.set_monitor_mode('evaluation')
rollout(env, good=True)
env.close()
results = monitoring.load_results(temp)
evaluation_score = benchmark.score_evaluation(
'CartPole-v0', results['data_sources'],
results['initial_reset_timestamps'], results['episode_lengths'],
results['episode_rewards'], results['episode_types'],
results['timestamps'])
benchmark_score = benchmark.score_benchmark({
'CartPole-v0':
evaluation_score['scores'],
})
assert np.all(
np.isclose(evaluation_score['scores'],
[0.00089999999999999998, 0.0054000000000000003
])), "evaluation_score={}".format(evaluation_score)
assert np.isclose(
benchmark_score,
0.00315), "benchmark_score={}".format(benchmark_score)
def test_only_complete_episodes_written():
with helpers.tempdir() as temp:
env = gym.make('CartPole-v0')
env = Monitor(env, temp, video_callable=False)
env.reset()
d = False
while not d:
_, _, d, _ = env.step(env.action_space.sample())
env.reset()
env.step(env.action_space.sample())
env.close()
# Only 1 episode should be written
results = monitoring.load_results(temp)
assert len(results['episode_lengths']) == 1, "Found {} episodes written; expecting 1".format(len(results['episode_lengths']))
def test_env_reuse():
with helpers.tempdir() as temp:
env = gym.make('Autoreset-v0')
env = Monitor(env, temp)
env.reset()
_, _, done, _ = env.step(None)
assert not done
_, _, done, _ = env.step(None)
assert done
_, _, done, _ = env.step(None)
assert not done
_, _, done, _ = env.step(None)
assert done
env.close()
def test_no_monitor_reset_unless_done():
def assert_reset_raises(env):
errored = False
try:
env.reset()
except error.Error:
errored = True
assert errored, "Env allowed a reset when it shouldn't have"
with helpers.tempdir() as temp:
# Make sure we can reset as we please without monitor
env = gym.make('CartPole-v0')
env.reset()
env.step(env.action_space.sample())
env.step(env.action_space.sample())
env.reset()
# can reset once as soon as we start
env = Monitor(env, temp, video_callable=False)
env.reset()
# can reset multiple times in a row
env.reset()
env.reset()
env.step(env.action_space.sample())
env.step(env.action_space.sample())
assert_reset_raises(env)
# should allow resets after the episode is done
d = False
while not d:
_, _, d, _ = env.step(env.action_space.sample())
env.reset()
env.reset()
env.step(env.action_space.sample())
assert_reset_raises(env)
env.close()
def callback(self, obs_t, obs_tp1, action, rew, done, info):
points = self.data_callback(obs_t, obs_tp1, action, rew, done, info)
for point, data_series in zip(points, self.data):
data_series.append(point)
self.t += 1
xmin, xmax = max(0, self.t - self.horizon_timesteps), self.t
for i, plot in enumerate(self.cur_plot):
if plot is not None:
plot.remove()
self.cur_plot[i] = self.ax[i].scatter(range(xmin, xmax),
list(self.data[i]))
self.ax[i].set_xlim(xmin, xmax)
plt.pause(0.000001)
if __name__ == '__main__':
from rl_algs.common.atari_wrappers import wrap_deepmind
def callback(obs_t, obs_tp1, action, rew, done, info):
return [rew, obs_t.mean()]
env_plotter = EnvPlotter(callback, 30 * 5, ["reward", "mean intensity"])
env = gym.make("MontezumaRevengeNoFrameskip-v3")
env = wrap_deepmind(env)
play_env(env, zoom=4, callback=env_plotter.callback, fps=30)
