def stats(policy, train_batch):
values_batched = make_time_major(policy,
train_batch.get("seq_lens"),
policy.model.value_function(),
drop_last=policy.config["vtrace"])
stats_dict = {
"cur_lr":
policy.cur_lr,
"policy_loss":
policy.loss.pi_loss,
"entropy":
policy.loss.entropy,
"var_gnorm":
global_norm(policy.model.trainable_variables()),
"vf_loss":
policy.loss.vf_loss,
"vf_explained_var":
explained_variance(torch.reshape(policy.loss.value_targets, [-1]),
torch.reshape(values_batched, [-1]),
framework="torch"),
}
if policy.config["vtrace"]:
is_stat_mean = torch.mean(policy.loss.is_ratio, [0, 1])
is_stat_var = torch.var(policy.loss.is_ratio, [0, 1])
stats_dict.update({"mean_IS": is_stat_mean})
stats_dict.update({"var_IS": is_stat_var})
if policy.config["use_kl_loss"]:
stats_dict.update({"kl": policy.loss.mean_kl})
stats_dict.update({"KL_Coeff": policy.kl_coeff})
return stats_dict
def stats(policy: Policy, train_batch: SampleBatch):
"""Stats function for APPO. Returns a dict with important loss stats.
Args:
policy (Policy): The Policy to generate stats for.
train_batch (SampleBatch): The SampleBatch (already) used for training.
Returns:
Dict[str, TensorType]: The stats dict.
"""
values_batched = make_time_major(policy,
train_batch.get("seq_lens"),
policy.model.value_function(),
drop_last=policy.config["vtrace"])
stats_dict = {
"cur_lr":
policy.cur_lr,
"policy_loss":
policy._mean_policy_loss,
"entropy":
policy._mean_entropy,
"var_gnorm":
global_norm(policy.model.trainable_variables()),
"vf_loss":
policy._mean_vf_loss,
"vf_explained_var":
explained_variance(torch.reshape(policy._value_targets, [-1]),
torch.reshape(values_batched, [-1])),
}
if policy.config["vtrace"]:
is_stat_mean = torch.mean(policy._is_ratio, [0, 1])
is_stat_var = torch.var(policy._is_ratio, [0, 1])
stats_dict.update({"mean_IS": is_stat_mean})
stats_dict.update({"var_IS": is_stat_var})
if policy.config["use_kl_loss"]:
stats_dict.update({"kl": policy._mean_kl})
stats_dict.update({"KL_Coeff": policy.kl_coeff})
return stats_dict
def _make_time_major(*args, **kw):
return make_time_major(policy, train_batch.get("seq_lens"), *args,
**kw)
def _make_time_major(*args, **kwargs):
return make_time_major(policy, train_batch.get(SampleBatch.SEQ_LENS),
*args, **kwargs)
