def kl_and_loss_stats(policy: Policy,
train_batch: SampleBatch) -> Dict[str, TensorType]:
"""Stats function for PPO. Returns a dict with important KL and 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.
"""
return {
"cur_kl_coeff":
policy.kl_coeff,
"cur_lr":
policy.cur_lr,
"total_loss":
policy._total_loss,
"policy_loss":
policy._mean_policy_loss,
"vf_loss":
policy._mean_vf_loss,
"vf_explained_var":
explained_variance(train_batch[Postprocessing.VALUE_TARGETS],
policy.model.value_function()),
"kl":
policy._mean_kl,
"entropy":
policy._mean_entropy,
"entropy_coeff":
policy.entropy_coeff,
}
def marwil_loss(policy, model, dist_class, train_batch):
model_out, _ = model.from_batch(train_batch)
action_dist = dist_class(model_out, model)
state_values = model.value_function()
advantages = train_batch[Postprocessing.ADVANTAGES]
actions = train_batch[SampleBatch.ACTIONS]
# Advantage estimation.
adv = advantages - state_values
adv_squared = torch.mean(torch.pow(adv, 2.0))
# Value loss.
policy.v_loss = 0.5 * adv_squared
# Policy loss.
# Update averaged advantage norm.
policy.ma_adv_norm.add_(1e-6 * (adv_squared - policy.ma_adv_norm))
# Exponentially weighted advantages.
exp_advs = torch.exp(policy.config["beta"] *
(adv / (1e-8 + torch.pow(policy.ma_adv_norm, 0.5))))
# log\pi_\theta(a|s)
logprobs = action_dist.logp(actions)
policy.p_loss = -1.0 * torch.mean(exp_advs.detach() * logprobs)
# Combine both losses.
policy.total_loss = policy.p_loss + policy.config["vf_coeff"] * \
policy.v_loss
explained_var = explained_variance(advantages, state_values)
policy.explained_variance = torch.mean(explained_var)
return policy.total_loss
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])),
}
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 kl_and_loss_stats(policy, train_batch):
return {
"cur_kl_coeff": policy.kl_coeff,
"cur_lr": policy.cur_lr,
"total_loss": policy.loss_obj.loss,
"policy_loss": policy.loss_obj.mean_policy_loss,
"vf_loss": policy.loss_obj.mean_vf_loss,
"vf_explained_var": explained_variance(
train_batch[Postprocessing.VALUE_TARGETS],
policy.model.value_function()),
"kl": policy.loss_obj.mean_kl,
"entropy": policy.loss_obj.mean_entropy,
"entropy_coeff": policy.entropy_coeff,
}
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"])
return {
"cur_lr": policy.cur_lr,
"policy_loss": policy.loss.pi_loss,
"entropy": policy.loss.mean_entropy,
"entropy_coeff": policy.entropy_coeff,
"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])),
}
def marwil_loss(policy, model, dist_class, train_batch):
model_out, _ = model.from_batch(train_batch)
print("model_out.shape", model_out.shape)
state_values = model.value_function()
advantages = train_batch[Postprocessing.ADVANTAGES]
actions = train_batch[SampleBatch.ACTIONS]
print("actions", actions[0:2])
for i in range(0, model_out.shape[0]):
if actions[i][0] == 0:
model_out[i][-6:] = 0
elif actions[i][0] == 1:
model_out[i][-4:] = 0
model_out[i][3:7] = 0
elif actions[i][0] == 2:
model_out[i][3:-4] = 0
else:
pass
print("model_out", model_out[0:2])
action_dist = dist_class(model_out, model)
print("action_dist", action_dist.input_lens)
# Value loss.
policy.v_loss = 0.5 * torch.mean(torch.pow(state_values - advantages, 2.0))
# Policy loss.
# Advantage estimation.
adv = advantages - state_values
# Update averaged advantage norm.
policy.ma_adv_norm.add_(
1e-6 * (torch.mean(torch.pow(adv, 2.0)) - policy.ma_adv_norm))
# #xponentially weighted advantages.
exp_advs = torch.exp(policy.config["beta"] *
(adv / (1e-8 + torch.pow(policy.ma_adv_norm, 0.5))))
# log\pi_\theta(a|s)
logprobs = action_dist.logp(actions)
policy.p_loss = -1.0 * torch.mean(exp_advs.detach() * logprobs)
# Combine both losses.
policy.total_loss = policy.p_loss + policy.config["vf_coeff"] * \
policy.v_loss
explained_var = explained_variance(advantages, state_values)
policy.explained_variance = torch.mean(explained_var)
return policy.total_loss
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 marwil_loss(policy: Policy, model: ModelV2, dist_class: ActionDistribution,
train_batch: SampleBatch) -> TensorType:
model_out, _ = model.from_batch(train_batch)
action_dist = dist_class(model_out, model)
actions = train_batch[SampleBatch.ACTIONS]
# log\pi_\theta(a|s)
logprobs = action_dist.logp(actions)
# Advantage estimation.
if policy.config["beta"] != 0.0:
cumulative_rewards = train_batch[Postprocessing.ADVANTAGES]
state_values = model.value_function()
adv = cumulative_rewards - state_values
adv_squared_mean = torch.mean(torch.pow(adv, 2.0))
explained_var = explained_variance(cumulative_rewards, state_values)
policy.explained_variance = torch.mean(explained_var)
# Policy loss.
# Update averaged advantage norm.
rate = policy.config["moving_average_sqd_adv_norm_update_rate"]
policy._moving_average_sqd_adv_norm.add_(
rate * (adv_squared_mean - policy._moving_average_sqd_adv_norm))
# Exponentially weighted advantages.
exp_advs = torch.exp(
policy.config["beta"] *
(adv /
(1e-8 + torch.pow(policy._moving_average_sqd_adv_norm, 0.5))))
policy.p_loss = -torch.mean(exp_advs.detach() * logprobs)
# Value loss.
policy.v_loss = 0.5 * adv_squared_mean
else:
# Policy loss (simple BC loss term).
policy.p_loss = -1.0 * torch.mean(logprobs)
# Value loss.
policy.v_loss = 0.0
# Combine both losses.
policy.total_loss = policy.p_loss + policy.config["vf_coeff"] * \
policy.v_loss
return policy.total_loss
def ppo_surrogate_loss(
policy: Policy, model: ModelV2,
dist_class: Type[TorchDistributionWrapper],
train_batch: SampleBatch) -> Union[TensorType, List[TensorType]]:
"""Constructs the loss for Proximal Policy Objective.
Args:
policy (Policy): The Policy to calculate the loss for.
model (ModelV2): The Model to calculate the loss for.
dist_class (Type[ActionDistribution]: The action distr. class.
train_batch (SampleBatch): The training data.
Returns:
Union[TensorType, List[TensorType]]: A single loss tensor or a list
of loss tensors.
"""
logits, state = model.from_batch(train_batch, is_training=True)
curr_action_dist = dist_class(logits, model)
# RNN case: Mask away 0-padded chunks at end of time axis.
if state:
B = len(train_batch["seq_lens"])
max_seq_len = logits.shape[0] // B
mask = sequence_mask(train_batch["seq_lens"],
max_seq_len,
time_major=model.is_time_major())
mask = torch.reshape(mask, [-1])
num_valid = torch.sum(mask)
def reduce_mean_valid(t):
return torch.sum(t[mask]) / num_valid
# non-RNN case: No masking.
else:
mask = None
reduce_mean_valid = torch.mean
prev_action_dist = dist_class(train_batch[SampleBatch.ACTION_DIST_INPUTS],
model)
logp_ratio = torch.exp(
curr_action_dist.logp(train_batch[SampleBatch.ACTIONS]) -
train_batch[SampleBatch.ACTION_LOGP])
action_kl = prev_action_dist.kl(curr_action_dist)
mean_kl = reduce_mean_valid(action_kl)
curr_entropy = curr_action_dist.entropy()
mean_entropy = reduce_mean_valid(curr_entropy)
surrogate_loss = torch.min(
train_batch[Postprocessing.ADVANTAGES] * logp_ratio,
train_batch[Postprocessing.ADVANTAGES] *
torch.clamp(logp_ratio, 1 - policy.config["clip_param"],
1 + policy.config["clip_param"]))
mean_policy_loss = reduce_mean_valid(-surrogate_loss)
if policy.config["use_gae"]:
prev_value_fn_out = train_batch[SampleBatch.VF_PREDS]
value_fn_out = model.value_function()
vf_loss1 = torch.pow(
value_fn_out - train_batch[Postprocessing.VALUE_TARGETS], 2.0)
vf_clipped = prev_value_fn_out + torch.clamp(
value_fn_out - prev_value_fn_out, -policy.config["vf_clip_param"],
policy.config["vf_clip_param"])
vf_loss2 = torch.pow(
vf_clipped - train_batch[Postprocessing.VALUE_TARGETS], 2.0)
vf_loss = torch.max(vf_loss1, vf_loss2)
mean_vf_loss = reduce_mean_valid(vf_loss)
total_loss = reduce_mean_valid(-surrogate_loss +
policy.kl_coeff * action_kl +
policy.config["vf_loss_coeff"] *
vf_loss -
policy.entropy_coeff * curr_entropy)
else:
mean_vf_loss = 0.0
total_loss = reduce_mean_valid(-surrogate_loss +
policy.kl_coeff * action_kl -
policy.entropy_coeff * curr_entropy)
# Store stats in policy for stats_fn.
policy._total_loss = total_loss
policy._mean_policy_loss = mean_policy_loss
policy._mean_vf_loss = mean_vf_loss
policy._vf_explained_var = explained_variance(
train_batch[Postprocessing.VALUE_TARGETS],
policy.model.value_function())
policy._mean_entropy = mean_entropy
policy._mean_kl = mean_kl
return total_loss
def appo_surrogate_loss(policy: Policy, model: ModelV2,
dist_class: Type[TorchDistributionWrapper],
train_batch: SampleBatch) -> TensorType:
"""Constructs the loss for APPO.
With IS modifications and V-trace for Advantage Estimation.
Args:
policy (Policy): The Policy to calculate the loss for.
model (ModelV2): The Model to calculate the loss for.
dist_class (Type[ActionDistribution]): The action distr. class.
train_batch (SampleBatch): The training data.
Returns:
Union[TensorType, List[TensorType]]: A single loss tensor or a list
of loss tensors.
"""
target_model = policy.target_models[model]
model_out, _ = model.from_batch(train_batch)
action_dist = dist_class(model_out, model)
if isinstance(policy.action_space, gym.spaces.Discrete):
is_multidiscrete = False
output_hidden_shape = [policy.action_space.n]
elif isinstance(policy.action_space,
gym.spaces.multi_discrete.MultiDiscrete):
is_multidiscrete = True
output_hidden_shape = policy.action_space.nvec.astype(np.int32)
else:
is_multidiscrete = False
output_hidden_shape = 1
def _make_time_major(*args, **kw):
return make_time_major(policy, train_batch.get("seq_lens"), *args,
**kw)
actions = train_batch[SampleBatch.ACTIONS]
dones = train_batch[SampleBatch.DONES]
rewards = train_batch[SampleBatch.REWARDS]
behaviour_logits = train_batch[SampleBatch.ACTION_DIST_INPUTS]
target_model_out, _ = target_model.from_batch(train_batch)
prev_action_dist = dist_class(behaviour_logits, model)
values = model.value_function()
values_time_major = _make_time_major(values)
if policy.is_recurrent():
max_seq_len = torch.max(train_batch["seq_lens"])
mask = sequence_mask(train_batch["seq_lens"], max_seq_len)
mask = torch.reshape(mask, [-1])
mask = _make_time_major(mask, drop_last=policy.config["vtrace"])
num_valid = torch.sum(mask)
def reduce_mean_valid(t):
return torch.sum(t[mask]) / num_valid
else:
reduce_mean_valid = torch.mean
if policy.config["vtrace"]:
logger.debug("Using V-Trace surrogate loss (vtrace=True)")
old_policy_behaviour_logits = target_model_out.detach()
old_policy_action_dist = dist_class(old_policy_behaviour_logits, model)
if isinstance(output_hidden_shape, (list, tuple, np.ndarray)):
unpacked_behaviour_logits = torch.split(behaviour_logits,
list(output_hidden_shape),
dim=1)
unpacked_old_policy_behaviour_logits = torch.split(
old_policy_behaviour_logits, list(output_hidden_shape), dim=1)
else:
unpacked_behaviour_logits = torch.chunk(behaviour_logits,
output_hidden_shape,
dim=1)
unpacked_old_policy_behaviour_logits = torch.chunk(
old_policy_behaviour_logits, output_hidden_shape, dim=1)
# Prepare actions for loss.
loss_actions = actions if is_multidiscrete else torch.unsqueeze(
actions, dim=1)
# Prepare KL for loss.
action_kl = _make_time_major(old_policy_action_dist.kl(action_dist),
drop_last=True)
# Compute vtrace on the CPU for better perf.
vtrace_returns = vtrace.multi_from_logits(
behaviour_policy_logits=_make_time_major(unpacked_behaviour_logits,
drop_last=True),
target_policy_logits=_make_time_major(
unpacked_old_policy_behaviour_logits, drop_last=True),
actions=torch.unbind(_make_time_major(loss_actions,
drop_last=True),
dim=2),
discounts=(1.0 - _make_time_major(dones, drop_last=True).float()) *
policy.config["gamma"],
rewards=_make_time_major(rewards, drop_last=True),
values=values_time_major[:-1], # drop-last=True
bootstrap_value=values_time_major[-1],
dist_class=TorchCategorical if is_multidiscrete else dist_class,
model=model,
clip_rho_threshold=policy.config["vtrace_clip_rho_threshold"],
clip_pg_rho_threshold=policy.config["vtrace_clip_pg_rho_threshold"]
)
actions_logp = _make_time_major(action_dist.logp(actions),
drop_last=True)
prev_actions_logp = _make_time_major(prev_action_dist.logp(actions),
drop_last=True)
old_policy_actions_logp = _make_time_major(
old_policy_action_dist.logp(actions), drop_last=True)
is_ratio = torch.clamp(
torch.exp(prev_actions_logp - old_policy_actions_logp), 0.0, 2.0)
logp_ratio = is_ratio * torch.exp(actions_logp - prev_actions_logp)
policy._is_ratio = is_ratio
advantages = vtrace_returns.pg_advantages.to(logp_ratio.device)
surrogate_loss = torch.min(
advantages * logp_ratio,
advantages *
torch.clamp(logp_ratio, 1 - policy.config["clip_param"],
1 + policy.config["clip_param"]))
mean_kl = reduce_mean_valid(action_kl)
mean_policy_loss = -reduce_mean_valid(surrogate_loss)
# The value function loss.
value_targets = vtrace_returns.vs.to(values_time_major.device)
delta = values_time_major[:-1] - value_targets
mean_vf_loss = 0.5 * reduce_mean_valid(torch.pow(delta, 2.0))
# The entropy loss.
mean_entropy = reduce_mean_valid(
_make_time_major(action_dist.entropy(), drop_last=True))
else:
logger.debug("Using PPO surrogate loss (vtrace=False)")
# Prepare KL for Loss
action_kl = _make_time_major(prev_action_dist.kl(action_dist))
actions_logp = _make_time_major(action_dist.logp(actions))
prev_actions_logp = _make_time_major(prev_action_dist.logp(actions))
logp_ratio = torch.exp(actions_logp - prev_actions_logp)
advantages = _make_time_major(train_batch[Postprocessing.ADVANTAGES])
surrogate_loss = torch.min(
advantages * logp_ratio,
advantages *
torch.clamp(logp_ratio, 1 - policy.config["clip_param"],
1 + policy.config["clip_param"]))
mean_kl = reduce_mean_valid(action_kl)
mean_policy_loss = -reduce_mean_valid(surrogate_loss)
# The value function loss.
value_targets = _make_time_major(
train_batch[Postprocessing.VALUE_TARGETS])
delta = values_time_major - value_targets
mean_vf_loss = 0.5 * reduce_mean_valid(torch.pow(delta, 2.0))
# The entropy loss.
mean_entropy = reduce_mean_valid(
_make_time_major(action_dist.entropy()))
# The summed weighted loss
total_loss = mean_policy_loss + \
mean_vf_loss * policy.config["vf_loss_coeff"] - \
mean_entropy * policy.config["entropy_coeff"]
# Optional additional KL Loss
if policy.config["use_kl_loss"]:
total_loss += policy.kl_coeff * mean_kl
policy._total_loss = total_loss
policy._mean_policy_loss = mean_policy_loss
policy._mean_kl = mean_kl
policy._mean_vf_loss = mean_vf_loss
policy._mean_entropy = mean_entropy
policy._value_targets = value_targets
policy._vf_explained_var = explained_variance(
torch.reshape(value_targets, [-1]),
torch.reshape(
values_time_major[:-1]
if policy.config["vtrace"] else values_time_major, [-1]),
)
return total_loss
def build_vtrace_loss(policy, model, dist_class, train_batch):
model_out, _ = model.from_batch(train_batch)
action_dist = dist_class(model_out, model)
if isinstance(policy.action_space, gym.spaces.Discrete):
is_multidiscrete = False
output_hidden_shape = [policy.action_space.n]
elif isinstance(policy.action_space, gym.spaces.MultiDiscrete):
is_multidiscrete = True
output_hidden_shape = policy.action_space.nvec.astype(np.int32)
else:
is_multidiscrete = False
output_hidden_shape = 1
def _make_time_major(*args, **kw):
return make_time_major(policy, train_batch.get(SampleBatch.SEQ_LENS),
*args, **kw)
actions = train_batch[SampleBatch.ACTIONS]
dones = train_batch[SampleBatch.DONES]
rewards = train_batch[SampleBatch.REWARDS]
behaviour_action_logp = train_batch[SampleBatch.ACTION_LOGP]
behaviour_logits = train_batch[SampleBatch.ACTION_DIST_INPUTS]
if isinstance(output_hidden_shape, (list, tuple, np.ndarray)):
unpacked_behaviour_logits = torch.split(behaviour_logits,
list(output_hidden_shape),
dim=1)
unpacked_outputs = torch.split(model_out,
list(output_hidden_shape),
dim=1)
else:
unpacked_behaviour_logits = torch.chunk(behaviour_logits,
output_hidden_shape,
dim=1)
unpacked_outputs = torch.chunk(model_out, output_hidden_shape, dim=1)
values = model.value_function()
if policy.is_recurrent():
max_seq_len = torch.max(train_batch[SampleBatch.SEQ_LENS])
mask_orig = sequence_mask(train_batch[SampleBatch.SEQ_LENS],
max_seq_len)
mask = torch.reshape(mask_orig, [-1])
else:
mask = torch.ones_like(rewards)
# Prepare actions for loss.
loss_actions = actions if is_multidiscrete else torch.unsqueeze(actions,
dim=1)
# Inputs are reshaped from [B * T] => [T - 1, B] for V-trace calc.
loss = VTraceLoss(
actions=_make_time_major(loss_actions, drop_last=True),
actions_logp=_make_time_major(action_dist.logp(actions),
drop_last=True),
actions_entropy=_make_time_major(action_dist.entropy(),
drop_last=True),
dones=_make_time_major(dones, drop_last=True),
behaviour_action_logp=_make_time_major(behaviour_action_logp,
drop_last=True),
behaviour_logits=_make_time_major(unpacked_behaviour_logits,
drop_last=True),
target_logits=_make_time_major(unpacked_outputs, drop_last=True),
discount=policy.config["gamma"],
rewards=_make_time_major(rewards, drop_last=True),
values=_make_time_major(values, drop_last=True),
bootstrap_value=_make_time_major(values)[-1],
dist_class=TorchCategorical if is_multidiscrete else dist_class,
model=model,
valid_mask=_make_time_major(mask, drop_last=True),
config=policy.config,
vf_loss_coeff=policy.config["vf_loss_coeff"],
entropy_coeff=policy.entropy_coeff,
clip_rho_threshold=policy.config["vtrace_clip_rho_threshold"],
clip_pg_rho_threshold=policy.config["vtrace_clip_pg_rho_threshold"])
# Store values for stats function in model (tower), such that for
# multi-GPU, we do not override them during the parallel loss phase.
model.tower_stats["pi_loss"] = loss.pi_loss
model.tower_stats["vf_loss"] = loss.vf_loss
model.tower_stats["entropy"] = loss.entropy
model.tower_stats["mean_entropy"] = loss.mean_entropy
model.tower_stats["total_loss"] = loss.total_loss
values_batched = make_time_major(policy,
train_batch.get(SampleBatch.SEQ_LENS),
values,
drop_last=policy.config["vtrace"])
model.tower_stats["vf_explained_var"] = explained_variance(
torch.reshape(loss.value_targets, [-1]),
torch.reshape(values_batched, [-1]))
return loss.total_loss
