def _evaluation_paths(self, policy, evaluation_env):
if self._eval_n_episodes < 1: return ()
# TODO(hartikainen): I don't like this way of handling evaluation mode
# for the policies. We should instead have two separete policies for
# training and evaluation.
with policy.evaluation_mode():
paths = rollouts(self._eval_n_episodes,
evaluation_env,
policy,
self.sampler._max_path_length,
render_kwargs=self._eval_render_kwargs)
should_save_video = (
self._video_save_frequency > 0
and (self._epoch == 0 or
(self._epoch + 1) % self._video_save_frequency == 0))
if should_save_video:
fps = 1 // getattr(self._training_environment, 'dt', 1 / 30)
for i, path in enumerate(paths):
video_frames = path.pop('images')
video_file_name = f'evaluation_path_{self._epoch}_{i}.mp4'
video_file_path = os.path.join(os.getcwd(), 'videos',
video_file_name)
save_video(video_frames, video_file_path, fps=fps)
return paths
def _evaluation_paths(self, policy, evaluation_env):
if self._eval_n_episodes < 1: return ()
with policy.set_deterministic(self._eval_deterministic):
paths = rollouts(self._eval_n_episodes,
evaluation_env,
policy,
self.sampler._max_path_length,
render_kwargs=self._eval_render_kwargs)
should_save_video = (
self._video_save_frequency > 0
and (self._epoch == 0 or
(self._epoch + 1) % self._video_save_frequency == 0))
if should_save_video:
fps = 1 // getattr(self._training_environment, 'dt', 1 / 30)
for i, path in enumerate(paths):
video_frames = path.pop('images')
video_file_name = f'evaluation_path_{self._epoch}_{i}.mp4'
video_file_path = os.path.join(os.getcwd(), 'videos',
video_file_name)
save_video(video_frames, video_file_path, fps=fps)
return paths
def simulate_policy(checkpoint_path,
num_rollouts,
max_path_length,
render_kwargs,
video_save_path=None,
evaluation_environment_params=None):
checkpoint_path = os.path.abspath(checkpoint_path.rstrip('/'))
variant, progress, metadata = load_variant_progress_metadata(
checkpoint_path)
environment = load_environment(variant)
policy = load_policy(checkpoint_path, variant, environment)
render_kwargs = {**DEFAULT_RENDER_KWARGS, **render_kwargs}
paths = rollouts(num_rollouts,
environment,
policy,
path_length=max_path_length,
render_kwargs=render_kwargs)
if video_save_path and render_kwargs.get('mode') == 'rgb_array':
fps = 1 // getattr(environment, 'dt', 1 / 30)
for i, path in enumerate(paths):
video_save_dir = os.path.expanduser('/tmp/simulate_policy/')
video_save_path = os.path.join(video_save_dir, f'episode_{i}.mp4')
save_video(path['images'], video_save_path, fps=fps)
return paths