zjowowen's picture
init space
079c32c
raw
history blame
69.7 kB
from typing import List, Dict, Any, Tuple, Union
from collections import namedtuple
import copy
import torch
from torch.utils.data import Dataset, DataLoader
from ding.utils import POLICY_REGISTRY, split_data_generator, RunningMeanStd
from ding.utils.data import default_collate, default_decollate
from ding.torch_utils import Adam, to_device
from ding.rl_utils import get_gae_with_default_last_value, get_train_sample, gae, gae_data, get_gae, \
ppo_policy_data, ppo_policy_error, ppo_value_data, ppo_value_error, ppg_data, ppg_joint_error
from ding.model import model_wrap
from .base_policy import Policy
class ExperienceDataset(Dataset):
"""
Overview:
A dataset class for storing and accessing experience data.
Interface:
``__init__``, ``__len__``, ``__getitem__``.
"""
def __init__(self, data):
"""
Arguments:
- data (:obj:`dict`): A dictionary containing the experience data, where the keys represent the data types \
and the values are the corresponding data arrays.
"""
super().__init__()
self.data = data
def __len__(self):
return list(self.data.values())[0].shape[0]
def __getitem__(self, ind):
data = {}
for key in self.data.keys():
data[key] = self.data[key][ind]
return data
def create_shuffled_dataloader(data, batch_size):
ds = ExperienceDataset(data)
return DataLoader(ds, batch_size=batch_size, shuffle=True)
@POLICY_REGISTRY.register('ppg')
class PPGPolicy(Policy):
"""
Overview:
Policy class of PPG algorithm. PPG is a policy gradient algorithm with auxiliary phase training. \
The auxiliary phase training is proposed to distill the value into the policy network, \
while making sure the policy network does not change the action predictions (kl div loss). \
Paper link: https://arxiv.org/abs/2009.04416.
Interface:
``_init_learn``, ``_data_preprocess_learn``, ``_forward_learn``, ``_state_dict_learn``, \
``_load_state_dict_learn``, ``_init_collect``, ``_forward_collect``, ``_process_transition``, \
``_get_train_sample``, ``_get_batch_size``, ``_init_eval``, ``_forward_eval``, ``default_model``, \
``_monitor_vars_learn``, ``learn_aux``.
Config:
== ==================== ======== ============== ======================================== =======================
ID Symbol Type Default Value Description Other(Shape)
== ==================== ======== ============== ======================================== =======================
1 ``type`` str ppg | RL policy register name, refer to | this arg is optional,
| registry ``POLICY_REGISTRY`` | a placeholder
2 ``cuda`` bool False | Whether to use cuda for network | this arg can be diff-
| erent from modes
3 ``on_policy`` bool True | Whether the RL algorithm is on-policy
| or off-policy
4. ``priority`` bool False | Whether use priority(PER) | priority sample,
| update priority
5 | ``priority_`` bool False | Whether use Importance Sampling | IS weight
| ``IS_weight`` | Weight to correct biased update.
6 | ``learn.update`` int 5 | How many updates(iterations) to train | this args can be vary
| ``_per_collect`` | after collector's one collection. Only | from envs. Bigger val
| valid in serial training | means more off-policy
7 | ``learn.value_`` float 1.0 | The loss weight of value network | policy network weight
| ``weight`` | is set to 1
8 | ``learn.entropy_`` float 0.01 | The loss weight of entropy | policy network weight
| ``weight`` | regularization | is set to 1
9 | ``learn.clip_`` float 0.2 | PPO clip ratio
| ``ratio``
10 | ``learn.adv_`` bool False | Whether to use advantage norm in
| ``norm`` | a whole training batch
11 | ``learn.aux_`` int 5 | The frequency(normal update times)
| ``freq`` | of auxiliary phase training
12 | ``learn.aux_`` int 6 | The training epochs of auxiliary
| ``train_epoch`` | phase
13 | ``learn.aux_`` int 1 | The loss weight of behavioral_cloning
| ``bc_weight`` | in auxiliary phase
14 | ``collect.dis`` float 0.99 | Reward's future discount factor, aka. | may be 1 when sparse
| ``count_factor`` | gamma | reward env
15 | ``collect.gae_`` float 0.95 | GAE lambda factor for the balance
| ``lambda`` | of bias and variance(1-step td and mc)
== ==================== ======== ============== ======================================== =======================
"""
config = dict(
# (str) RL policy register name (refer to function "POLICY_REGISTRY").
type='ppg',
# (bool) Whether to use cuda for network.
cuda=False,
# (bool) Whether the RL algorithm is on-policy or off-policy. (Note: in practice PPO can be off-policy used)
on_policy=True,
priority=False,
# (bool) Whether use Importance Sampling Weight to correct biased update. If True, priority must be True.
priority_IS_weight=False,
learn=dict(
actor_epoch_per_collect=1,
critic_epoch_per_collect=1,
batch_size=64,
learning_rate=0.001,
# ==============================================================
# The following configs is algorithm-specific
# ==============================================================
# (float) The loss weight of value network, policy network weight is set to 1
value_weight=0.5,
# (float) The loss weight of entropy regularization, policy network weight is set to 1
entropy_weight=0.01,
# (float) PPO clip ratio, defaults to 0.2
clip_ratio=0.2,
value_norm=False,
# (bool) Whether to use advantage norm in a whole training batch
adv_norm=False,
# (int) The frequency(normal update times) of auxiliary phase training
aux_freq=8,
# (int) The training epochs of auxiliary phase
aux_train_epoch=6,
# (int) The loss weight of behavioral_cloning in auxiliary phase
aux_bc_weight=1,
grad_clip_type='clip_norm',
grad_clip_value=10,
ignore_done=False,
),
collect=dict(
# n_sample=64,
unroll_len=1,
# ==============================================================
# The following configs is algorithm-specific
# ==============================================================
# (float) Reward's future discount factor, aka. gamma.
discount_factor=0.99,
# (float) GAE lambda factor for the balance of bias and variance(1-step td and mc)
gae_lambda=0.95,
),
eval=dict(),
)
def default_model(self) -> Tuple[str, List[str]]:
"""
Overview:
Return this algorithm default neural network model setting for demonstration. ``__init__`` method will \
automatically call this method to get the default model setting and create model.
Returns:
- model_info (:obj:`Tuple[str, List[str]]`): The registered model name and model's import_names.
"""
return 'ppg', ['ding.model.template.ppg']
def _init_learn(self) -> None:
"""
Overview:
Initialize the learn mode of policy, including related attributes and modules. For PPG, it mainly \
contains optimizer, algorithm-specific arguments such as aux_bc_weight and aux_train_epoch. This method \
also executes some special network initializations and prepares running mean/std monitor for value. \
This method will be called in ``__init__`` method if ``learn`` field is in ``enable_field``.
.. note::
For the member variables that need to be saved and loaded, please refer to the ``_state_dict_learn`` \
and ``_load_state_dict_learn`` methods.
.. note::
For the member variables that need to be monitored, please refer to the ``_monitor_vars_learn`` method.
.. note::
If you want to set some spacial member variables in ``_init_learn`` method, you'd better name them \
with prefix ``_learn_`` to avoid conflict with other modes, such as ``self._learn_attr1``.
"""
# Optimizer
self._optimizer_ac = Adam(self._model.actor_critic.parameters(), lr=self._cfg.learn.learning_rate)
self._optimizer_aux_critic = Adam(self._model.aux_critic.parameters(), lr=self._cfg.learn.learning_rate)
self._learn_model = model_wrap(self._model, wrapper_name='base')
# Algorithm config
self._priority = self._cfg.priority
self._priority_IS_weight = self._cfg.priority_IS_weight
assert not self._priority and not self._priority_IS_weight, "Priority is not implemented in PPG"
self._value_weight = self._cfg.learn.value_weight
self._entropy_weight = self._cfg.learn.entropy_weight
self._value_norm = self._cfg.learn.value_norm
if self._value_norm:
self._running_mean_std = RunningMeanStd(epsilon=1e-4, device=self._device)
self._clip_ratio = self._cfg.learn.clip_ratio
self._adv_norm = self._cfg.learn.adv_norm
# Main model
self._learn_model.reset()
# Auxiliary memories
self._aux_train_epoch = self._cfg.learn.aux_train_epoch
self._train_iteration = 0
self._aux_memories = []
self._aux_bc_weight = self._cfg.learn.aux_bc_weight
def _data_preprocess_learn(self, data: List[Any]) -> dict:
"""
Overview:
Preprocess the data to fit the required data format for learning, including \
collate(stack data into batch), ignore done(in some fake terminate env),\
prepare loss weight per training sample, and cpu tensor to cuda.
Arguments:
- data (:obj:`List[Dict[str, Any]]`): The data collected from collect function.
Returns:
- data (:obj:`Dict[str, Any]`): The processed data, including at least ['done', 'weight'].
"""
# data preprocess
data = default_collate(data)
ignore_done = self._cfg.learn.ignore_done
if ignore_done:
data['done'] = None
else:
data['done'] = data['done'].float()
data['weight'] = None
if self._cuda:
data = to_device(data, self._device)
return data
def _forward_learn(self, data: dict) -> Dict[str, Any]:
"""
Overview:
Forward and backward function of learn mode.
Arguments:
- data (:obj:`Dict[str, Any]`): Input data used for policy forward, including the \
collected training samples from replay buffer. For each element in dict, the key of the \
dict is the name of data items and the value is the corresponding data. Usually, the value is \
torch.Tensor or np.ndarray or there dict/list combinations. In the ``_forward_learn`` method, data \
often need to first be stacked in the batch dimension by some utility functions such as \
``default_preprocess_learn``. \
For PPG, each element in list is a dict containing at least the following keys: ``obs``, ``action``, \
``reward``, ``logit``, ``value``, ``done``. Sometimes, it also contains other keys such as ``weight``.
Returns:
- info_dict (:obj:`Dict[str, Any]`): Dict type data, a info dict indicated training result, which will be \
recorded in text log and tensorboard, values are python scalar or a list of scalars. \
For the detailed definition of the dict, refer to the code of ``_monitor_vars_learn`` method.
.. note::
The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \
For the data type that not supported, the main reason is that the corresponding model does not support it. \
You can implement you own model rather than use the default model. For more information, please raise an \
issue in GitHub repo and we will continue to follow up.
.. note::
For more detailed examples, please refer to our unittest for PPGPolicy: ``ding.policy.tests.test_ppgs``.
"""
data = self._data_preprocess_learn(data)
# ====================
# PPG forward
# ====================
self._learn_model.train()
return_infos = []
if self._value_norm:
unnormalized_return = data['adv'] + data['value'] * self._running_mean_std.std
data['return'] = unnormalized_return / self._running_mean_std.std
self._running_mean_std.update(unnormalized_return.cpu().numpy())
else:
data['return'] = data['adv'] + data['value']
for epoch in range(self._cfg.learn.actor_epoch_per_collect):
for policy_data in split_data_generator(data, self._cfg.learn.batch_size, shuffle=True):
policy_adv = policy_data['adv']
if self._adv_norm:
# Normalize advantage in a total train_batch
policy_adv = (policy_adv - policy_adv.mean()) / (policy_adv.std() + 1e-8)
# Policy Phase(Policy)
policy_output = self._learn_model.forward(policy_data['obs'], mode='compute_actor')
policy_error_data = ppo_policy_data(
policy_output['logit'], policy_data['logit'], policy_data['action'], policy_adv,
policy_data['weight']
)
ppo_policy_loss, ppo_info = ppo_policy_error(policy_error_data, self._clip_ratio)
policy_loss = ppo_policy_loss.policy_loss - self._entropy_weight * ppo_policy_loss.entropy_loss
self._optimizer_ac.zero_grad()
policy_loss.backward()
self._optimizer_ac.step()
for epoch in range(self._cfg.learn.critic_epoch_per_collect):
for value_data in split_data_generator(data, self._cfg.learn.batch_size, shuffle=True):
value_adv = value_data['adv']
return_ = value_data['return']
if self._adv_norm:
# Normalize advantage in a total train_batch
value_adv = (value_adv - value_adv.mean()) / (value_adv.std() + 1e-8)
# Policy Phase(Value)
value_output = self._learn_model.forward(value_data['obs'], mode='compute_critic')
value_error_data = ppo_value_data(
value_output['value'], value_data['value'], return_, value_data['weight']
)
value_loss = self._value_weight * ppo_value_error(value_error_data, self._clip_ratio)
self._optimizer_aux_critic.zero_grad()
value_loss.backward()
self._optimizer_aux_critic.step()
data['return_'] = data['return']
self._aux_memories.append(copy.deepcopy(data))
self._train_iteration += 1
# ====================
# PPG update
# use aux loss after iterations and reset aux_memories
# ====================
# Auxiliary Phase
# record data for auxiliary head
if self._train_iteration % self._cfg.learn.aux_freq == 0:
aux_loss, bc_loss, aux_value_loss = self.learn_aux()
return {
'policy_cur_lr': self._optimizer_ac.defaults['lr'],
'value_cur_lr': self._optimizer_aux_critic.defaults['lr'],
'policy_loss': ppo_policy_loss.policy_loss.item(),
'value_loss': value_loss.item(),
'entropy_loss': ppo_policy_loss.entropy_loss.item(),
'policy_adv_abs_max': policy_adv.abs().max().item(),
'approx_kl': ppo_info.approx_kl,
'clipfrac': ppo_info.clipfrac,
'aux_value_loss': aux_value_loss,
'auxiliary_loss': aux_loss,
'behavioral_cloning_loss': bc_loss,
}
else:
return {
'policy_cur_lr': self._optimizer_ac.defaults['lr'],
'value_cur_lr': self._optimizer_aux_critic.defaults['lr'],
'policy_loss': ppo_policy_loss.policy_loss.item(),
'value_loss': value_loss.item(),
'entropy_loss': ppo_policy_loss.entropy_loss.item(),
'policy_adv_abs_max': policy_adv.abs().max().item(),
'approx_kl': ppo_info.approx_kl,
'clipfrac': ppo_info.clipfrac,
}
def _state_dict_learn(self) -> Dict[str, Any]:
"""
Overview:
Return the state_dict of learn mode, usually including model and optimizer.
Returns:
- state_dict (:obj:`Dict[str, Any]`): the dict of current policy learn state, for saving and restoring.
"""
return {
'model': self._learn_model.state_dict(),
'optimizer_ac': self._optimizer_ac.state_dict(),
'optimizer_aux_critic': self._optimizer_aux_critic.state_dict(),
}
def _load_state_dict_learn(self, state_dict: Dict[str, Any]) -> None:
"""
Overview:
Load the state_dict variable into policy learn mode.
Arguments:
- state_dict (:obj:`Dict[str, Any]`): the dict of policy learn state saved before.\
When the value is distilled into the policy network, we need to make sure the policy \
network does not change the action predictions, we need two optimizers, \
_optimizer_ac is used in policy net, and _optimizer_aux_critic is used in value net.
.. tip::
If you want to only load some parts of model, you can simply set the ``strict`` argument in \
load_state_dict to ``False``, or refer to ``ding.torch_utils.checkpoint_helper`` for more \
complicated operation.
"""
self._learn_model.load_state_dict(state_dict['model'])
self._optimizer_ac.load_state_dict(state_dict['optimizer_ac'])
self._optimizer_aux_critic.load_state_dict(state_dict['optimizer_aux_critic'])
def _init_collect(self) -> None:
"""
Overview:
Initialize the collect mode of policy, including related attributes and modules. For PPG, it contains the \
collect_model to balance the exploration and exploitation (e.g. the multinomial sample mechanism in \
discrete action space), and other algorithm-specific arguments such as unroll_len and gae_lambda.
This method will be called in ``__init__`` method if ``collect`` field is in ``enable_field``.
.. note::
If you want to set some spacial member variables in ``_init_collect`` method, you'd better name them \
with prefix ``_collect_`` to avoid conflict with other modes, such as ``self._collect_attr1``.
"""
self._unroll_len = self._cfg.collect.unroll_len
self._collect_model = model_wrap(self._model, wrapper_name='multinomial_sample')
# TODO continuous action space exploration
self._collect_model.reset()
self._gamma = self._cfg.collect.discount_factor
self._gae_lambda = self._cfg.collect.gae_lambda
def _forward_collect(self, data: dict) -> dict:
"""
Overview:
Policy forward function of collect mode (collecting training data by interacting with envs). Forward means \
that the policy gets some necessary data (mainly observation) from the envs and then returns the output \
data, such as the action to interact with the envs.
Arguments:
- data (:obj:`Dict[str, Any]`): Dict type data, stacked env data for predicting policy_output(action), \
values are torch.Tensor or np.ndarray or dict/list combinations, keys are env_id indicated by integer.
Returns:
- output (:obj:`Dict[int, Any]`): The output data of policy forward, including at least the action and \
other necessary data (action logit and value) for learn mode defined in \
``self._process_transition`` method. The key of the dict is the same as the input data, \
i.e. environment id.
.. tip::
If you want to add more tricks on this policy, like temperature factor in multinomial sample, you can pass \
related data as extra keyword arguments of this method.
.. note::
The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \
For the data type that not supported, the main reason is that the corresponding model does not support it. \
You can implement you own model rather than use the default model. For more information, please raise an \
issue in GitHub repo and we will continue to follow up.
.. note::
For more detailed examples, please refer to our unittest for PPGPolicy: ``ding.policy.tests.test_ppg``.
"""
data_id = list(data.keys())
data = default_collate(list(data.values()))
if self._cuda:
data = to_device(data, self._device)
self._collect_model.eval()
with torch.no_grad():
output = self._collect_model.forward(data, mode='compute_actor_critic')
if self._cuda:
output = to_device(output, 'cpu')
output = default_decollate(output)
return {i: d for i, d in zip(data_id, output)}
def _process_transition(self, obs: Any, model_output: dict, timestep: namedtuple) -> dict:
"""
Overview:
Process and pack one timestep transition data into a dict, which can be directly used for training and \
saved in replay buffer. For PPG, it contains obs, next_obs, action, reward, done, logit, value.
Arguments:
- obs (:obj:`Any`): Env observation
- model_output (:obj:`dict`): The output of the policy network with the observation \
as input. For PPG, it contains the state value, action and the logit of the action.
- timestep (:obj:`namedtuple`): The execution result namedtuple returned by the environment step \
method, except all the elements have been transformed into tensor data. Usually, it contains the next \
obs, reward, done, info, etc.
Returns:
- transition (:obj:`dict`): The processed transition data of the current timestep.
.. note::
``next_obs`` is used to calculate nstep return when necessary, so we place in into transition by default. \
You can delete this field to save memory occupancy if you do not need nstep return.
"""
transition = {
'obs': obs,
'next_obs': timestep.obs,
'logit': model_output['logit'],
'action': model_output['action'],
'value': model_output['value'],
'reward': timestep.reward,
'done': timestep.done,
}
return transition
def _get_train_sample(self, data: List[Dict[str, Any]]) -> Union[None, List[Any]]:
"""
Overview:
For a given trajectory (transitions, a list of transition) data, process it into a list of sample that \
can be used for training directly. In PPG, a train sample is a processed transition with new computed \
``adv`` field. This method is usually used in collectors to execute necessary. \
RL data preprocessing before training, which can help learner amortize revelant time consumption. \
In addition, you can also implement this method as an identity function and do the data processing \
in ``self._forward_learn`` method.
Arguments:
- data (:obj:`List[Dict[str, Any]]`): The trajectory data (a list of transition), each element is \
the same format as the return value of ``self._process_transition`` method.
Returns:
- samples (:obj:`dict`): The processed train samples, each element is the similar format \
as input transitions, but may contain more data for training, such as GAE advantage.
"""
data = to_device(data, self._device)
if self._cfg.learn.ignore_done:
data[-1]['done'] = False
if data[-1]['done']:
last_value = torch.zeros_like(data[-1]['value'])
else:
with torch.no_grad():
last_value = self._collect_model.forward(
data[-1]['next_obs'].unsqueeze(0), mode='compute_actor_critic'
)['value']
if self._value_norm:
last_value *= self._running_mean_std.std
for i in range(len(data)):
data[i]['value'] *= self._running_mean_std.std
data = get_gae(
data,
to_device(last_value, self._device),
gamma=self._gamma,
gae_lambda=self._gae_lambda,
cuda=False,
)
if self._value_norm:
for i in range(len(data)):
data[i]['value'] /= self._running_mean_std.std
return get_train_sample(data, self._unroll_len)
def _get_batch_size(self) -> Dict[str, int]:
"""
Overview:
Get learn batch size. In the PPG algorithm, different networks require different data.\
We need to get data['policy'] and data['value'] to train policy net and value net,\
this function is used to get the batch size of data['policy'] and data['value'].
Returns:
- output (:obj:`dict[str, int]`): Dict type data, including str type batch size and int type batch size.
"""
bs = self._cfg.learn.batch_size
return {'policy': bs, 'value': bs}
def _init_eval(self) -> None:
"""
Overview:
Initialize the eval mode of policy, including related attributes and modules. For PPG, it contains the \
eval model to select optimial action (e.g. greedily select action with argmax mechanism in discrete \
action). This method will be called in ``__init__`` method if ``eval`` field is in ``enable_field``.
.. note::
If you want to set some spacial member variables in ``_init_eval`` method, you'd better name them \
with prefix ``_eval_`` to avoid conflict with other modes, such as ``self._eval_attr1``.
"""
self._eval_model = model_wrap(self._model, wrapper_name='argmax_sample')
self._eval_model.reset()
def _forward_eval(self, data: dict) -> dict:
"""
Overview:
Policy forward function of eval mode (evaluation policy performance by interacting with envs). Forward \
means that the policy gets some necessary data (mainly observation) from the envs and then returns the \
action to interact with the envs. ``_forward_eval`` in PPG often uses deterministic sample method to get \
actions while ``_forward_collect`` usually uses stochastic sample method for balance exploration and \
exploitation.
Arguments:
- data (:obj:`Dict[str, Any]`): The input data used for policy forward, including at least the obs. The \
key of the dict is environment id and the value is the corresponding data of the env.
Returns:
- output (:obj:`Dict[int, Any]`): The output data of policy forward, including at least the action. The \
key of the dict is the same as the input data, i.e. environment id.
.. note::
The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \
For the data type that not supported, the main reason is that the corresponding model does not support it. \
You can implement you own model rather than use the default model. For more information, please raise an \
issue in GitHub repo and we will continue to follow up.
.. note::
For more detailed examples, please refer to our unittest for PPGPolicy: ``ding.policy.tests.test_ppg``.
"""
data_id = list(data.keys())
data = default_collate(list(data.values()))
if self._cuda:
data = to_device(data, self._device)
self._eval_model.eval()
with torch.no_grad():
output = self._eval_model.forward(data, mode='compute_actor')
if self._cuda:
output = to_device(output, 'cpu')
output = default_decollate(output)
return {i: d for i, d in zip(data_id, output)}
def _monitor_vars_learn(self) -> List[str]:
"""
Overview:
Return the necessary keys for logging the return dict of ``self._forward_learn``. The logger module, such \
as text logger, tensorboard logger, will use these keys to save the corresponding data.
Returns:
- vars (:obj:`List[str]`): The list of the necessary keys to be logged.
"""
return [
'policy_cur_lr',
'value_cur_lr',
'policy_loss',
'value_loss',
'entropy_loss',
'policy_adv_abs_max',
'approx_kl',
'clipfrac',
'aux_value_loss',
'auxiliary_loss',
'behavioral_cloning_loss',
]
def learn_aux(self) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Overview:
The auxiliary phase training, where the value is distilled into the policy network. In PPG algorithm, \
we use the value function loss as the auxiliary objective, thereby sharing features between the policy \
and value function while minimizing distortions to the policy. We also use behavioral cloning loss to \
optimize the auxiliary objective while otherwise preserving the original policy.
Returns:
- aux_loss (:obj:`Tuple[torch.Tensor, torch.Tensor, torch.Tensor]`): Including average auxiliary loss\
average behavioral cloning loss, and average auxiliary value loss.
"""
aux_memories = self._aux_memories
# gather states and target values into one tensor
data = {}
states = []
actions = []
return_ = []
old_values = []
weights = []
for memory in aux_memories:
# for memory in memories:
states.append(memory['obs'])
actions.append(memory['action'])
return_.append(memory['return_'])
old_values.append(memory['value'])
if memory['weight'] is None:
weight = torch.ones_like(memory['action'])
else:
weight = torch.tensor(memory['weight'])
weights.append(weight)
data['obs'] = torch.cat(states)
data['action'] = torch.cat(actions)
data['return_'] = torch.cat(return_)
data['value'] = torch.cat(old_values)
data['weight'] = torch.cat(weights).float()
# compute current policy logit_old
with torch.no_grad():
data['logit_old'] = self._model.forward(data['obs'], mode='compute_actor')['logit']
# prepared dataloader for auxiliary phase training
dl = create_shuffled_dataloader(data, self._cfg.learn.batch_size)
# the proposed auxiliary phase training
# where the value is distilled into the policy network,
# while making sure the policy network does not change the action predictions (kl div loss)
i = 0
auxiliary_loss_ = 0
behavioral_cloning_loss_ = 0
value_loss_ = 0
for epoch in range(self._aux_train_epoch):
for data in dl:
policy_output = self._model.forward(data['obs'], mode='compute_actor_critic')
# Calculate ppg error 'logit_new', 'logit_old', 'action', 'value_new', 'value_old', 'return_', 'weight'
data_ppg = ppg_data(
policy_output['logit'], data['logit_old'], data['action'], policy_output['value'], data['value'],
data['return_'], data['weight']
)
ppg_joint_loss = ppg_joint_error(data_ppg, self._clip_ratio)
wb = self._aux_bc_weight
total_loss = ppg_joint_loss.auxiliary_loss + wb * ppg_joint_loss.behavioral_cloning_loss
# # policy network loss copmoses of both the kl div loss as well as the auxiliary loss
# aux_loss = clipped_value_loss(policy_values, rewards, old_values, self.value_clip)
# loss_kl = F.kl_div(action_logprobs, old_action_probs, reduction='batchmean')
# policy_loss = aux_loss + loss_kl
self._optimizer_ac.zero_grad()
total_loss.backward()
self._optimizer_ac.step()
# paper says it is important to train the value network extra during the auxiliary phase
# Calculate ppg error 'value_new', 'value_old', 'return_', 'weight'
values = self._model.forward(data['obs'], mode='compute_critic')['value']
data_aux = ppo_value_data(values, data['value'], data['return_'], data['weight'])
value_loss = ppo_value_error(data_aux, self._clip_ratio)
self._optimizer_aux_critic.zero_grad()
value_loss.backward()
self._optimizer_aux_critic.step()
auxiliary_loss_ += ppg_joint_loss.auxiliary_loss.item()
behavioral_cloning_loss_ += ppg_joint_loss.behavioral_cloning_loss.item()
value_loss_ += value_loss.item()
i += 1
self._aux_memories = []
return auxiliary_loss_ / i, behavioral_cloning_loss_ / i, value_loss_ / i
@POLICY_REGISTRY.register('ppg_offpolicy')
class PPGOffPolicy(Policy):
"""
Overview:
Policy class of PPG algorithm with off-policy training mode. Off-policy PPG contains two different data \
max_use buffers. The policy buffer offers data for policy phase , while the value buffer provides auxiliary \
phase's data. The whole training procedure is similar to off-policy PPO but execute additional auxiliary \
phase with a fixed frequency.
Interface:
``_init_learn``, ``_data_preprocess_learn``, ``_forward_learn``, ``_state_dict_learn``, \
``_load_state_dict_learn``, ``_init_collect``, ``_forward_collect``, ``_process_transition``, \
``_get_train_sample``, ``_get_batch_size``, ``_init_eval``, ``_forward_eval``, ``default_model``, \
``_monitor_vars_learn``, ``learn_aux``.
Config:
== ==================== ======== ============== ======================================== =======================
ID Symbol Type Default Value Description Other(Shape)
== ==================== ======== ============== ======================================== =======================
1 ``type`` str ppg | RL policy register name, refer to | this arg is optional,
| registry ``POLICY_REGISTRY`` | a placeholder
2 ``cuda`` bool False | Whether to use cuda for network | this arg can be diff-
| erent from modes
3 ``on_policy`` bool True | Whether the RL algorithm is on-policy
| or off-policy
4. ``priority`` bool False | Whether use priority(PER) | priority sample,
| update priority
5 | ``priority_`` bool False | Whether use Importance Sampling | IS weight
| ``IS_weight`` | Weight to correct biased update.
6 | ``learn.update`` int 5 | How many updates(iterations) to train | this args can be vary
| ``_per_collect`` | after collector's one collection. Only | from envs. Bigger val
| valid in serial training | means more off-policy
7 | ``learn.value_`` float 1.0 | The loss weight of value network | policy network weight
| ``weight`` | is set to 1
8 | ``learn.entropy_`` float 0.01 | The loss weight of entropy | policy network weight
| ``weight`` | regularization | is set to 1
9 | ``learn.clip_`` float 0.2 | PPO clip ratio
| ``ratio``
10 | ``learn.adv_`` bool False | Whether to use advantage norm in
| ``norm`` | a whole training batch
11 | ``learn.aux_`` int 5 | The frequency(normal update times)
| ``freq`` | of auxiliary phase training
12 | ``learn.aux_`` int 6 | The training epochs of auxiliary
| ``train_epoch`` | phase
13 | ``learn.aux_`` int 1 | The loss weight of behavioral_cloning
| ``bc_weight`` | in auxiliary phase
14 | ``collect.dis`` float 0.99 | Reward's future discount factor, aka. | may be 1 when sparse
| ``count_factor`` | gamma | reward env
15 | ``collect.gae_`` float 0.95 | GAE lambda factor for the balance
| ``lambda`` | of bias and variance(1-step td and mc)
== ==================== ======== ============== ======================================== =======================
"""
config = dict(
# (str) RL policy register name (refer to function "POLICY_REGISTRY").
type='ppg_offpolicy',
# (bool) Whether to use cuda for network.
cuda=False,
# (bool) Whether the RL algorithm is on-policy or off-policy. (Note: in practice PPO can be off-policy used)
on_policy=False,
priority=False,
# (bool) Whether use Importance Sampling Weight to correct biased update. If True, priority must be True.
priority_IS_weight=False,
# (bool) Whether to need policy data in process transition
transition_with_policy_data=True,
learn=dict(
update_per_collect=5,
batch_size=64,
learning_rate=0.001,
# ==============================================================
# The following configs is algorithm-specific
# ==============================================================
# (float) The loss weight of value network, policy network weight is set to 1
value_weight=0.5,
# (float) The loss weight of entropy regularization, policy network weight is set to 1
entropy_weight=0.01,
# (float) PPO clip ratio, defaults to 0.2
clip_ratio=0.2,
# (bool) Whether to use advantage norm in a whole training batch
adv_norm=False,
# (int) The frequency(normal update times) of auxiliary phase training
aux_freq=5,
# (int) The training epochs of auxiliary phase
aux_train_epoch=6,
# (int) The loss weight of behavioral_cloning in auxiliary phase
aux_bc_weight=1,
ignore_done=False,
),
collect=dict(
# n_sample=64,
unroll_len=1,
# ==============================================================
# The following configs is algorithm-specific
# ==============================================================
# (float) Reward's future discount factor, aka. gamma.
discount_factor=0.99,
# (float) GAE lambda factor for the balance of bias and variance(1-step td and mc)
gae_lambda=0.95,
),
eval=dict(),
other=dict(
replay_buffer=dict(
# PPG use two separate buffer for different reuse
multi_buffer=True,
policy=dict(replay_buffer_size=1000, ),
value=dict(replay_buffer_size=1000, ),
),
),
)
def default_model(self) -> Tuple[str, List[str]]:
"""
Overview:
Return this algorithm default neural network model setting for demonstration. ``__init__`` method will \
automatically call this method to get the default model setting and create model.
Returns:
- model_info (:obj:`Tuple[str, List[str]]`): The registered model name and model's import_names.
.. note::
The user can define and use customized network model but must obey the same inferface definition indicated \
by import_names path.
"""
return 'ppg', ['ding.model.template.ppg']
def _init_learn(self) -> None:
"""
Overview:
Initialize the learn mode of policy, including related attributes and modules. For PPG, it mainly \
contains optimizer, algorithm-specific arguments such as aux_bc_weight and aux_train_epoch. This method \
also executes some special network initializations and prepares running mean/std monitor for value. \
This method will be called in ``__init__`` method if ``learn`` field is in ``enable_field``.
.. note::
For the member variables that need to be saved and loaded, please refer to the ``_state_dict_learn`` \
and ``_load_state_dict_learn`` methods.
.. note::
For the member variables that need to be monitored, please refer to the ``_monitor_vars_learn`` method.
.. note::
If you want to set some spacial member variables in ``_init_learn`` method, you'd better name them \
with prefix ``_learn_`` to avoid conflict with other modes, such as ``self._learn_attr1``.
"""
# Optimizer
self._optimizer_ac = Adam(self._model.actor_critic.parameters(), lr=self._cfg.learn.learning_rate)
self._optimizer_aux_critic = Adam(self._model.aux_critic.parameters(), lr=self._cfg.learn.learning_rate)
self._learn_model = model_wrap(self._model, wrapper_name='base')
# Algorithm config
self._priority = self._cfg.priority
self._priority_IS_weight = self._cfg.priority_IS_weight
assert not self._priority and not self._priority_IS_weight, "Priority is not implemented in PPG"
self._value_weight = self._cfg.learn.value_weight
self._entropy_weight = self._cfg.learn.entropy_weight
self._clip_ratio = self._cfg.learn.clip_ratio
self._adv_norm = self._cfg.learn.adv_norm
# Main model
self._learn_model.reset()
# Auxiliary memories
self._aux_train_epoch = self._cfg.learn.aux_train_epoch
self._train_iteration = 0
self._aux_memories = []
self._aux_bc_weight = self._cfg.learn.aux_bc_weight
def _data_preprocess_learn(self, data: List[Any]) -> dict:
"""
Overview:
Preprocess the data to fit the required data format for learning, including \
collate(stack data into batch), ignore done(in some fake terminate env),\
prepare loss weight per training sample, and cpu tensor to cuda.
Arguments:
- data (:obj:`List[Dict[str, Any]]`): The data collected from collect function.
Returns:
- data (:obj:`Dict[str, Any]`): The processed data, including at least ['done', 'weight'].
"""
# data preprocess
for k, data_item in data.items():
data_item = default_collate(data_item)
ignore_done = self._cfg.learn.ignore_done
if ignore_done:
data_item['done'] = None
else:
data_item['done'] = data_item['done'].float()
data_item['weight'] = None
data[k] = data_item
if self._cuda:
data = to_device(data, self._device)
return data
def _forward_learn(self, data: dict) -> Dict[str, Any]:
"""
Overview:
Forward and backward function of learn mode.
Arguments:
- data (:obj:`Dict[str, Any]`): Input data used for policy forward, including the \
collected training samples from replay buffer. For each element in dict, the key of the \
dict is the name of data items and the value is the corresponding data. Usually, \
the class type of value is either torch.Tensor or np.ndarray, or a dict/list containing \
either torch.Tensor or np.ndarray items In the ``_forward_learn`` method, data \
often need to first be stacked in the batch dimension by some utility functions such as \
``default_preprocess_learn``. \
For PPGOff, each element in list is a dict containing at least the following keys: ``obs``, \
``action``, ``reward``, ``logit``, ``value``, ``done``. Sometimes, it also contains other keys \
such as ``weight``.
Returns:
- info_dict (:obj:`Dict[str, Any]`): Dict type data, a info dict indicated training result, which will be \
recorded in text log and tensorboard, values are python scalar or a list of scalars. \
For the detailed definition of the dict, refer to the code of ``_monitor_vars_learn`` method.
ReturnsKeys:
- necessary: "current lr", "total_loss", "policy_loss", "value_loss", "entropy_loss", \
"adv_abs_max", "approx_kl", "clipfrac", \
"aux_value_loss", "auxiliary_loss", "behavioral_cloning_loss".
- current_lr (:obj:`float`): Current learning rate.
- total_loss (:obj:`float`): The calculated loss.
- policy_loss (:obj:`float`): The policy(actor) loss of ppg.
- value_loss (:obj:`float`): The value(critic) loss of ppg.
- entropy_loss (:obj:`float`): The entropy loss.
- auxiliary_loss (:obj:`float`): The auxiliary loss, we use the value function loss \
as the auxiliary objective, thereby sharing features between the policy and value function\
while minimizing distortions to the policy.
- aux_value_loss (:obj:`float`): The auxiliary value loss, we need to train the value network extra \
during the auxiliary phase, it's the value loss we train the value network during auxiliary phase.
- behavioral_cloning_loss (:obj:`float`): The behavioral cloning loss, used to optimize the auxiliary\
objective while otherwise preserving the original policy.
"""
data = self._data_preprocess_learn(data)
# ====================
# PPG forward
# ====================
self._learn_model.train()
policy_data, value_data = data['policy'], data['value']
policy_adv, value_adv = policy_data['adv'], value_data['adv']
return_ = value_data['value'] + value_adv
if self._adv_norm:
# Normalize advantage in a total train_batch
policy_adv = (policy_adv - policy_adv.mean()) / (policy_adv.std() + 1e-8)
value_adv = (value_adv - value_adv.mean()) / (value_adv.std() + 1e-8)
# Policy Phase(Policy)
policy_output = self._learn_model.forward(policy_data['obs'], mode='compute_actor')
policy_error_data = ppo_policy_data(
policy_output['logit'], policy_data['logit'], policy_data['action'], policy_adv, policy_data['weight']
)
ppo_policy_loss, ppo_info = ppo_policy_error(policy_error_data, self._clip_ratio)
policy_loss = ppo_policy_loss.policy_loss - self._entropy_weight * ppo_policy_loss.entropy_loss
self._optimizer_ac.zero_grad()
policy_loss.backward()
self._optimizer_ac.step()
# Policy Phase(Value)
value_output = self._learn_model.forward(value_data['obs'], mode='compute_critic')
value_error_data = ppo_value_data(value_output['value'], value_data['value'], return_, value_data['weight'])
value_loss = self._value_weight * ppo_value_error(value_error_data, self._clip_ratio)
self._optimizer_aux_critic.zero_grad()
value_loss.backward()
self._optimizer_aux_critic.step()
# ====================
# PPG update
# use aux loss after iterations and reset aux_memories
# ====================
# Auxiliary Phase
# record data for auxiliary head
data = data['value']
data['return_'] = return_.data
self._aux_memories.append(copy.deepcopy(data))
self._train_iteration += 1
total_loss = policy_loss + value_loss
if self._train_iteration % self._cfg.learn.aux_freq == 0:
aux_loss, bc_loss, aux_value_loss = self.learn_aux()
total_loss += aux_loss + bc_loss + aux_value_loss
return {
'policy_cur_lr': self._optimizer_ac.defaults['lr'],
'value_cur_lr': self._optimizer_aux_critic.defaults['lr'],
'policy_loss': ppo_policy_loss.policy_loss.item(),
'value_loss': value_loss.item(),
'entropy_loss': ppo_policy_loss.entropy_loss.item(),
'policy_adv_abs_max': policy_adv.abs().max().item(),
'approx_kl': ppo_info.approx_kl,
'clipfrac': ppo_info.clipfrac,
'aux_value_loss': aux_value_loss,
'auxiliary_loss': aux_loss,
'behavioral_cloning_loss': bc_loss,
'total_loss': total_loss.item(),
}
else:
return {
'policy_cur_lr': self._optimizer_ac.defaults['lr'],
'value_cur_lr': self._optimizer_aux_critic.defaults['lr'],
'policy_loss': ppo_policy_loss.policy_loss.item(),
'value_loss': value_loss.item(),
'entropy_loss': ppo_policy_loss.entropy_loss.item(),
'policy_adv_abs_max': policy_adv.abs().max().item(),
'approx_kl': ppo_info.approx_kl,
'clipfrac': ppo_info.clipfrac,
'total_loss': total_loss.item(),
}
def _state_dict_learn(self) -> Dict[str, Any]:
"""
Overview:
Return the state_dict of learn mode, usually including model and optimizer.
Returns:
- state_dict (:obj:`Dict[str, Any]`): the dict of current policy learn state, for saving and restoring.
"""
return {
'model': self._learn_model.state_dict(),
'optimizer_ac': self._optimizer_ac.state_dict(),
'optimizer_aux_critic': self._optimizer_aux_critic.state_dict(),
}
def _load_state_dict_learn(self, state_dict: Dict[str, Any]) -> None:
"""
Overview:
Load the state_dict variable into policy learn mode.
Arguments:
- state_dict (:obj:`Dict[str, Any]`): the dict of policy learn state saved before.\
When the value is distilled into the policy network, we need to make sure the policy \
network does not change the action predictions, we need two optimizers, \
_optimizer_ac is used in policy net, and _optimizer_aux_critic is used in value net.
.. tip::
If you want to only load some parts of model, you can simply set the ``strict`` argument in \
load_state_dict to ``False``, or refer to ``ding.torch_utils.checkpoint_helper`` for more \
complicated operation.
"""
self._learn_model.load_state_dict(state_dict['model'])
self._optimizer_ac.load_state_dict(state_dict['optimizer_ac'])
self._optimizer_aux_critic.load_state_dict(state_dict['optimizer_aux_critic'])
def _init_collect(self) -> None:
"""
Overview:
Initialize the collect mode of policy, including related attributes and modules. For PPO, it contains the \
collect_model to balance the exploration and exploitation (e.g. the multinomial sample mechanism in \
discrete action space), and other algorithm-specific arguments such as unroll_len and gae_lambda.
This method will be called in ``__init__`` method if ``collect`` field is in ``enable_field``.
.. note::
If you want to set some spacial member variables in ``_init_collect`` method, you'd better name them \
with prefix ``_collect_`` to avoid conflict with other modes, such as ``self._collect_attr1``.
"""
self._unroll_len = self._cfg.collect.unroll_len
self._collect_model = model_wrap(self._model, wrapper_name='multinomial_sample')
# TODO continuous action space exploration
self._collect_model.reset()
self._gamma = self._cfg.collect.discount_factor
self._gae_lambda = self._cfg.collect.gae_lambda
def _forward_collect(self, data: dict) -> dict:
"""
Overview:
Policy forward function of collect mode (collecting training data by interacting with envs). Forward means \
that the policy gets some necessary data (mainly observation) from the envs and then returns the output \
data, such as the action to interact with the envs.
Arguments:
- data (:obj:`Dict[str, Any]`): Dict type data, stacked env data for predicting policy_output(action), \
values are torch.Tensor or np.ndarray or dict/list combinations, keys are env_id indicated by integer.
Returns:
- output (:obj:`Dict[int, Any]`): The output data of policy forward, including at least the action and \
other necessary data (action logit and value) for learn mode defined in \
``self._process_transition`` method. The key of the dict is the same as the input data, \
i.e. environment id.
.. tip::
If you want to add more tricks on this policy, like temperature factor in multinomial sample, you can pass \
related data as extra keyword arguments of this method.
.. note::
The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \
For the data type that not supported, the main reason is that the corresponding model does not support it. \
You can implement you own model rather than use the default model. For more information, please raise an \
issue in GitHub repo and we will continue to follow up.
.. note::
For more detailed examples, please refer to our unittest for PPGOffPolicy: ``ding.policy.tests.test_ppg``.
"""
data_id = list(data.keys())
data = default_collate(list(data.values()))
if self._cuda:
data = to_device(data, self._device)
self._collect_model.eval()
with torch.no_grad():
output = self._collect_model.forward(data, mode='compute_actor_critic')
if self._cuda:
output = to_device(output, 'cpu')
output = default_decollate(output)
return {i: d for i, d in zip(data_id, output)}
def _process_transition(self, obs: Any, model_output: dict, timestep: namedtuple) -> dict:
"""
Overview:
Process and pack one timestep transition data into a dict, which can be directly used for training and \
saved in replay buffer. For PPG, it contains obs, next_obs, action, reward, done, logit, value.
Arguments:
- obs (:obj:`Any`): Env observation
- model_output (:obj:`dict`): The output of the policy network with the observation \
as input. For PPG, it contains the state value, action and the logit of the action.
- timestep (:obj:`namedtuple`): The execution result namedtuple returned by the environment step \
method, except all the elements have been transformed into tensor data. Usually, it contains the next \
obs, reward, done, info, etc.
Returns:
- transition (:obj:`dict`): The processed transition data of the current timestep.
.. note::
``next_obs`` is used to calculate nstep return when necessary, so we place in into transition by default. \
You can delete this field to save memory occupancy if you do not need nstep return.
"""
transition = {
'obs': obs,
'next_obs': timestep.obs,
'logit': model_output['logit'],
'action': model_output['action'],
'value': model_output['value'],
'reward': timestep.reward,
'done': timestep.done,
}
return transition
def _get_train_sample(self, data: list) -> Union[None, List[Any]]:
"""
Overview:
For a given trajectory (transitions, a list of transition) data, process it into a list of sample that \
can be used for training directly. In PPG, a train sample is a processed transition with new computed \
``adv`` field. This method is usually used in collectors to execute necessary. \
RL data preprocessing before training, which can help learner amortize revelant time consumption. \
In addition, you can also implement this method as an identity function and do the data processing \
in ``self._forward_learn`` method.
Arguments:
- data (:obj:`list`): The trajectory data (a list of transition), each element is \
the same format as the return value of ``self._process_transition`` method.
Returns:
- samples (:obj:`dict`): The processed train samples, each element is the similar format \
as input transitions, but may contain more data for training, such as GAE advantage.
"""
data = get_gae_with_default_last_value(
data,
data[-1]['done'],
gamma=self._gamma,
gae_lambda=self._gae_lambda,
cuda=False,
)
data = get_train_sample(data, self._unroll_len)
for d in data:
d['buffer_name'] = ["policy", "value"]
return data
def _get_batch_size(self) -> Dict[str, int]:
"""
Overview:
Get learn batch size. In the PPG algorithm, different networks require different data.\
We need to get data['policy'] and data['value'] to train policy net and value net,\
this function is used to get the batch size of data['policy'] and data['value'].
Returns:
- output (:obj:`dict[str, int]`): Dict type data, including str type batch size and int type batch size.
"""
bs = self._cfg.learn.batch_size
return {'policy': bs, 'value': bs}
def _init_eval(self) -> None:
"""
Overview:
Initialize the eval mode of policy, including related attributes and modules. For PPG, it contains the \
eval model to select optimial action (e.g. greedily select action with argmax mechanism in discrete \
action). This method will be called in ``__init__`` method if ``eval`` field is in ``enable_field``.
.. note::
If you want to set some spacial member variables in ``_init_eval`` method, you'd better name them \
with prefix ``_eval_`` to avoid conflict with other modes, such as ``self._eval_attr1``.
"""
self._eval_model = model_wrap(self._model, wrapper_name='argmax_sample')
self._eval_model.reset()
def _forward_eval(self, data: dict) -> dict:
r"""
Overview:
Policy forward function of eval mode (evaluation policy performance by interacting with envs). Forward \
means that the policy gets some necessary data (mainly observation) from the envs and then returns the \
action to interact with the envs. ``_forward_eval`` in PPG often uses deterministic sample method to get \
actions while ``_forward_collect`` usually uses stochastic sample method for balance exploration and \
exploitation.
Arguments:
- data (:obj:`Dict[str, Any]`): The input data used for policy forward, including at least the obs. The \
key of the dict is environment id and the value is the corresponding data of the env.
Returns:
- output (:obj:`Dict[int, Any]`): The output data of policy forward, including at least the action. The \
key of the dict is the same as the input data, i.e. environment id.
.. note::
The input value can be torch.Tensor or dict/list combinations and current policy supports all of them. \
For the data type that not supported, the main reason is that the corresponding model does not support it. \
You can implement you own model rather than use the default model. For more information, please raise an \
issue in GitHub repo and we will continue to follow up.
.. note::
For more detailed examples, please refer to our unittest for PPGOffPolicy: ``ding.policy.tests.test_ppg``.
"""
data_id = list(data.keys())
data = default_collate(list(data.values()))
if self._cuda:
data = to_device(data, self._device)
self._eval_model.eval()
with torch.no_grad():
output = self._eval_model.forward(data, mode='compute_actor')
if self._cuda:
output = to_device(output, 'cpu')
output = default_decollate(output)
return {i: d for i, d in zip(data_id, output)}
def _monitor_vars_learn(self) -> List[str]:
"""
Overview:
Return the necessary keys for logging the return dict of ``self._forward_learn``. The logger module, such \
as text logger, tensorboard logger, will use these keys to save the corresponding data.
Returns:
- vars (:obj:`List[str]`): The list of the necessary keys to be logged.
"""
return [
'policy_cur_lr',
'value_cur_lr',
'policy_loss',
'value_loss',
'entropy_loss',
'policy_adv_abs_max',
'approx_kl',
'clipfrac',
'aux_value_loss',
'auxiliary_loss',
'behavioral_cloning_loss',
]
def learn_aux(self) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Overview:
The auxiliary phase training, where the value is distilled into the policy network. In PPG algorithm, \
we use the value function loss as the auxiliary objective, thereby sharing features between the policy \
and value function while minimizing distortions to the policy. We also use behavioral cloning loss to \
optimize the auxiliary objective while otherwise preserving the original policy.
Returns:
- aux_loss (:obj:`Tuple[torch.Tensor, torch.Tensor, torch.Tensor]`): Including average auxiliary loss\
average behavioral cloning loss, and average auxiliary value loss.
"""
aux_memories = self._aux_memories
# gather states and target values into one tensor
data = {}
states = []
actions = []
return_ = []
old_values = []
weights = []
for memory in aux_memories:
# for memory in memories:
states.append(memory['obs'])
actions.append(memory['action'])
return_.append(memory['return_'])
old_values.append(memory['value'])
if memory['weight'] is None:
weight = torch.ones_like(memory['action'])
else:
weight = torch.tensor(memory['weight'])
weights.append(weight)
data['obs'] = torch.cat(states)
data['action'] = torch.cat(actions)
data['return_'] = torch.cat(return_)
data['value'] = torch.cat(old_values)
data['weight'] = torch.cat(weights)
# compute current policy logit_old
with torch.no_grad():
data['logit_old'] = self._model.forward(data['obs'], mode='compute_actor')['logit']
# prepared dataloader for auxiliary phase training
dl = create_shuffled_dataloader(data, self._cfg.learn.batch_size)
# the proposed auxiliary phase training
# where the value is distilled into the policy network,
# while making sure the policy network does not change the action predictions (kl div loss)
i = 0
auxiliary_loss_ = 0
behavioral_cloning_loss_ = 0
value_loss_ = 0
for epoch in range(self._aux_train_epoch):
for data in dl:
policy_output = self._model.forward(data['obs'], mode='compute_actor_critic')
# Calculate ppg error 'logit_new', 'logit_old', 'action', 'value_new', 'value_old', 'return_', 'weight'
data_ppg = ppg_data(
policy_output['logit'], data['logit_old'], data['action'], policy_output['value'], data['value'],
data['return_'], data['weight']
)
ppg_joint_loss = ppg_joint_error(data_ppg, self._clip_ratio)
wb = self._aux_bc_weight
total_loss = ppg_joint_loss.auxiliary_loss + wb * ppg_joint_loss.behavioral_cloning_loss
# # policy network loss copmoses of both the kl div loss as well as the auxiliary loss
# aux_loss = clipped_value_loss(policy_values, rewards, old_values, self.value_clip)
# loss_kl = F.kl_div(action_logprobs, old_action_probs, reduction='batchmean')
# policy_loss = aux_loss + loss_kl
self._optimizer_ac.zero_grad()
total_loss.backward()
self._optimizer_ac.step()
# paper says it is important to train the value network extra during the auxiliary phase
# Calculate ppg error 'value_new', 'value_old', 'return_', 'weight'
values = self._model.forward(data['obs'], mode='compute_critic')['value']
data_aux = ppo_value_data(values, data['value'], data['return_'], data['weight'])
value_loss = ppo_value_error(data_aux, self._clip_ratio)
self._optimizer_aux_critic.zero_grad()
value_loss.backward()
self._optimizer_aux_critic.step()
auxiliary_loss_ += ppg_joint_loss.auxiliary_loss.item()
behavioral_cloning_loss_ += ppg_joint_loss.behavioral_cloning_loss.item()
value_loss_ += value_loss.item()
i += 1
self._aux_memories = []
return auxiliary_loss_ / i, behavioral_cloning_loss_ / i, value_loss_ / i