def test_train_mode_terminal_condition(self):
"""Loss of life is not considered terminal condition in `test` mode"""
ale = ALE(
rom='breakout',
mode='train',
width=None,
height=None,
grayscale=True,
repeat_action=1,
random_start=None,
)
outcome = ale.reset()
lives_before = outcome.info['lives']
while True:
outcome = ale.step(1)
if ale._ale.game_over():
break
lives_after = outcome.info['lives']
if lives_before == lives_after:
continue
lives_before = lives_after
self.assertTrue(
outcome.terminal,
'A life loss must be considered as terminal in `train` mode')
def test_buffer_frame_grayscale(self):
"""_frame_buffer contains the last raw frames given as buffer_frames"""
buffer_frames = 4
ale = ALE(
rom='breakout',
mode='train',
repeat_action=1,
buffer_frames=buffer_frames,
grayscale=True,
)
ale.reset()
frame = ale._get_raw_screen()
for i in range(1, buffer_frames):
ale.step(1)
frame = ale._get_raw_screen()
frame = frame[:, :, 0]
self.assertTrue((frame == ale._frame_buffer[i]).all())
for _ in range(10):
for i in range(buffer_frames):
ale.step(1)
frame = ale._get_raw_screen()
frame = frame[:, :, 0]
self.assertTrue((frame == ale._frame_buffer[i]).all())
def test_train_reset(self):
"""reset does not reset game in train mode when game is not yet over"""
ale = ALE(
'breakout',
mode='train',
random_start=0,
)
ale.reset()
while True:
outcome = ale.step(action=1)
if not outcome.terminal:
continue
if ale._ale.game_over():
break
fr0 = outcome.info['episode_frame_number']
outcome = ale.reset()
fr1 = outcome.info['episode_frame_number']
self.assertEqual(
fr1, fr0, 'New episode should not start at reset '
'when mode==train and not game_over')
outcome = ale.reset()
fr = outcome.info['episode_frame_number']
self.assertEqual(
fr, 1,
'New episode should be started when `reset` is called on game_over'
)
def test_grayscale_observation_color_channel_without_resize(self):
"""Observation has color channel when grayscale=True"""
ale = ALE(
rom='breakout',
grayscale=True,
)
ale.reset()
outcome = ale.step(1)
observation = outcome.state
self.assertTrue(len(observation.shape) == 2)
def test_rgb_observation_color_channel_without_resize(self):
"""Observation has color channel when grayscale=False"""
ale = ALE(
rom='breakout',
grayscale=False,
)
ale.reset()
outcome = ale.step(1)
observation = outcome.state
self.assertTrue(
len(observation.shape) == 3, 'Color channel is missing')
self.assertTrue(observation.shape[2] == 3,
'Incorrect number of color channels')
def test_no_resize_rgb(self):
"""Observation width and height equal to the original screen size"""
ale = ALE(
rom='breakout',
width=None,
height=None,
grayscale=False,
)
ale.reset()
outcome = ale.step(1)
expected = (210, 160)
found = outcome.state.shape[:2]
self.assertEqual(
expected, found,
'Observation shape must equal to the original screen size, '
'when `width` and `height` are `None`.')
def test_resize_height_rgb(self):
"""Observation size equals to the given size"""
height = 84
ale = ALE(
rom='breakout',
width=None,
height=height,
grayscale=False,
)
ale.reset()
outcome = ale.step(1)
expected = (height, 160)
found = outcome.state.shape[:2]
self.assertEqual(expected, found,
'Observation must be resized when height is given.')
def test_resize_width_grayscale(self):
"""Observation size equals to the given size"""
width = 84
ale = ALE(
rom='breakout',
width=width,
height=None,
grayscale=True,
)
ale.reset()
outcome = ale.step(1)
expected = (210, width)
found = outcome.state.shape
self.assertEqual(expected, found,
'Observation must be resized when width is given.')
def test_resize_width_and_height_grayscale(self):
"""Observation size equals to the given size"""
width, height = 84, 84
ale = ALE(
rom='breakout',
width=width,
height=height,
grayscale=True,
)
ale.reset()
outcome = ale.step(1)
expected = (height, width)
found = outcome.state.shape
self.assertEqual(
expected, found,
'Observation must be resized when both width and height are given.'
)
def test_random_start(self):
"""Episode starts from frame number in range of [1, `random_start`]"""
random_start = 30
ale = ALE(
rom='breakout',
random_start=random_start,
)
n_try = 100000
checked = [False] * random_start
for _ in range(n_try):
outcome = ale.reset()
frame = outcome.info['episode_frame_number']
checked[frame - 1] = True
if sum(checked) == random_start:
break
else:
self.fail(
'Not all starting frame (1 - {}) was observed after {} reset.'.
format(random_start, n_try))
def test_repeat_action(self):
"""`step` advances the number of frames given as `repeat_action`"""
for repeat_action in [1, 4]:
repeat_action = 1
ale = ALE(
rom='breakout',
mode='test',
repeat_action=repeat_action,
)
ale.reset()
last_frame = 1
while True:
outcome = ale.step(1)
if outcome.terminal:
break
frame = outcome.info['episode_frame_number']
self.assertEqual(frame - last_frame, repeat_action)
last_frame = frame
