def __init__(
self,
actions: List[str],
report_interval: int = 100,
target_action_distribution: Optional[List[float]] = None,
recent_window_size: int = 100,
):
self.value_list_observers = {
"cpe_results": ValueListObserver("cpe_details")
}
self.aggregating_observers = OrderedDict(
(name, IntervalAggregatingObserver(report_interval, aggregator))
for name, aggregator in [
("td_loss", agg.MeanAggregator("td_loss")),
("reward_loss", agg.MeanAggregator("reward_loss")),
(
"model_values",
agg.FunctionsByActionAggregator("model_values", actions, {
"mean": torch.mean,
"std": torch.std
}),
),
("logged_action",
agg.ActionCountAggregator("logged_actions", actions)),
(
"model_action",
agg.ActionCountAggregator("model_action_idxs", actions),
),
("recent_rewards",
agg.RecentValuesAggregator("logged_rewards")),
] + [(f"{key}_tb",
agg.TensorBoardActionCountAggregator(key, title, actions))
for key, title in [
("logged_actions", "logged"),
("model_action_idxs", "model"),
]] +
[(f"{key}_tb",
agg.TensorBoardHistogramAndMeanAggregator(key, log_key))
for key, log_key in [
("td_loss", "td_loss"),
("reward_loss", "reward_loss"),
("logged_propensities", "propensities/logged"),
("logged_rewards", "reward/logged"),
]] + [(
f"{key}_tb",
agg.TensorBoardActionHistogramAndMeanAggregator(
key, category, title, actions),
) for key, category, title in [
("model_propensities", "propensities", "model"),
("model_rewards", "reward", "model"),
("model_values", "value", "model"),
]])
self.last_epoch_end_num_batches = 0
self.num_data_points_per_epoch = None
epoch_end_observer = EpochEndObserver(self._epoch_end_callback)
super().__init__(
list(self.value_list_observers.values()) +
list(self.aggregating_observers.values()) + [epoch_end_observer])
self.target_action_distribution = target_action_distribution
self.recent_window_size = recent_window_size
def aggregating_observers(self):
return {
name: IntervalAggregatingObserver(self.report_interval, aggregator)
for name, aggregator in itertools.chain(
[
("mse_loss_per_batch", agg.MeanAggregator("mse_loss")),
(
"step_entropy_loss_per_batch",
agg.MeanAggregator("step_entropy_loss"),
),
(
"q_values_per_batch",
agg.FunctionsByActionAggregator(
"q_values", self.action_names,
{"mean": torch.mean}),
),
("eval_mse_loss_per_batch",
agg.MeanAggregator("eval_mse_loss")),
(
"eval_step_entropy_loss_per_batch",
agg.MeanAggregator("eval_step_entropy_loss"),
),
(
"eval_q_values_per_batch",
agg.FunctionsByActionAggregator(
"eval_q_values", self.action_names,
{"mean": torch.mean}),
),
(
"eval_action_distribution_per_batch",
agg.FunctionsByActionAggregator(
"eval_action_distribution",
self.action_names,
{"mean": torch.mean},
),
),
],
[(
f"{key}_tb",
agg.TensorBoardHistogramAndMeanAggregator(key, log_key),
) for key, log_key in [
("mse_loss", "mse_loss"),
("step_entropy_loss", "step_entropy_loss"),
("eval_mse_loss", "eval_mse_loss"),
("eval_step_entropy_loss", "eval_step_entropy_loss"),
]],
[(
f"{key}_tb",
agg.TensorBoardActionHistogramAndMeanAggregator(
key, category, title, self.action_names),
) for key, category, title in [
("q_values", "q_values", "training"),
("eval_q_values", "q_values", "eval"),
("eval_action_distribution", "action_distribution",
"eval"),
]],
)
}
def __init__(
self,
actions: List[str],
report_interval: int = 100,
target_action_distribution: Optional[List[float]] = None,
recent_window_size: int = 100,
):
self.value_list_observers = {}
self.aggregating_observers = {
**{
"cpe_results":
IntervalAggregatingObserver(1, agg.ListAggregator("cpe_details")),
},
**{
name: IntervalAggregatingObserver(report_interval, aggregator)
for name, aggregator in itertools.chain(
[
("td_loss", agg.MeanAggregator("td_loss")),
("reward_loss", agg.MeanAggregator("reward_loss")),
(
"model_values",
agg.FunctionsByActionAggregator(
"model_values",
actions,
{
"mean": torch.mean,
"std": torch.std
},
),
),
(
"logged_action",
agg.ActionCountAggregator("logged_actions", actions),
),
(
"model_action",
agg.ActionCountAggregator("model_action_idxs", actions),
),
(
"recent_rewards",
agg.RecentValuesAggregator("logged_rewards"),
),
],
[(
f"{key}_tb",
agg.TensorBoardActionCountAggregator(
key, title, actions),
) for key, title in [
("logged_actions", "logged"),
("model_action_idxs", "model"),
]],
[(
f"{key}_tb",
agg.TensorBoardHistogramAndMeanAggregator(
key, log_key),
) for key, log_key in [
("td_loss", "td_loss"),
("reward_loss", "reward_loss"),
("logged_propensities", "propensities/logged"),
("logged_rewards", "reward/logged"),
]],
[(
f"{key}_tb",
agg.TensorBoardActionHistogramAndMeanAggregator(
key, category, title, actions),
) for key, category, title in [
("model_propensities", "propensities", "model"),
("model_rewards", "reward", "model"),
("model_values", "value", "model"),
]],
)
},
}
super().__init__(self.value_list_observers, self.aggregating_observers)
self.target_action_distribution = target_action_distribution
self.recent_window_size = recent_window_size