def _save(self, agent_context: core.AgentContext):
"""Saves the current policy in directory."""
assert agent_context
assert agent_context.train, "TrainContext not set."
tc = agent_context.train
min_rewards, avg_reward, max_rewards = tc.eval_rewards[
tc.episodes_done_in_training]
current_dir = f'episode_{tc.episodes_done_in_training}-avg_reward_{avg_reward}'
current_dir = os.path.join(self.directory, current_dir)
agent_context._agent_saver(directory=current_dir)
self.saved_agents.append(
(tc.episodes_done_in_training, avg_reward, current_dir))
def on_play_end(self, agent_context: core.AgentContext):
if agent_context._is_plot_ready(core.PlotType.TRAIN_EVAL):
self._display_plots(agent_context)
if agent_context.is_play:
self._display_plots(agent_context)
if on_play_end_clear_jupyter_display:
self._clear_jupyter_plots(agent_context, wait=False)
def _refresh_subplot(self, agent_context: core.AgentContext,
plot_type: core.PlotType):
"""Sets this axes active and calls plot if this plot callback is registered on at least 1 plot
out of plot_type."""
assert self.axes is not None
plot_type = plot_type & self._plot_type
if agent_context._is_plot_ready(plot_type):
pyc = agent_context.pyplot
if not pyc.is_jupyter_active:
plt.figure(pyc.figure.number)
if plt.gcf() is pyc.figure:
plt.sca(self.axes)
self.plot(agent_context, plot_type)
def on_play_end(self, agent_context: core.AgentContext):
if agent_context._is_plot_ready(core.PlotType.PLAY_EPISODE):
self._close(agent_context)
def on_train_iteration_end(self, agent_context: core.AgentContext):
if agent_context._is_plot_ready(core.PlotType.TRAIN_ITERATION):
self._write_figure_to_video(agent_context)
def on_play_step_end(self, agent_context: core.AgentContext, action,
step_result: Tuple):
if agent_context._is_plot_ready(core.PlotType.PLAY_STEP):
self._write_figure_to_video(agent_context)
def on_play_episode_end(self, agent_context: core.AgentContext):
if agent_context._is_plot_ready(core.PlotType.PLAY_EPISODE
| core.PlotType.TRAIN_EVAL):
self._write_figure_to_video(agent_context)
def on_train_iteration_end(self, agent_context: core.AgentContext):
if agent_context._is_plot_ready(core.PlotType.TRAIN_ITERATION):
self._display_plots(agent_context)
def on_train_iteration_begin(self, agent_context: core.AgentContext):
# display initial evaluation before training starts.
if agent_context.train.iterations_done_in_training == 0 and \
agent_context._is_plot_ready(core.PlotType.TRAIN_EVAL):
self._display_plots(agent_context)
def on_play_step_end(self, agent_context: core.AgentContext, action,
step_result: Tuple):
if agent_context._is_plot_ready(core.PlotType.PLAY_STEP):
self._display_plots(agent_context)
def on_play_episode_end(self, agent_context: core.AgentContext):
if agent_context._is_plot_ready(core.PlotType.PLAY_EPISODE):
self._display_plots(agent_context)
def on_play_begin(self, agent_context: core.AgentContext):
agent_context.play.max_steps_per_episode = self._max_steps_per_episode
if isinstance(agent_context, core.EpisodesTrainContext):
agent_context.num_episodes_per_iteration = self._num_episodes_per_iteration