File size: 17,511 Bytes
079c32c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 |
from collections import namedtuple
from typing import Optional, Tuple
import torch
import torch.nn as nn
from torch.distributions import Independent, Normal
from ding.hpc_rl import hpc_wrapper
ppo_data = namedtuple(
'ppo_data', ['logit_new', 'logit_old', 'action', 'value_new', 'value_old', 'adv', 'return_', 'weight']
)
ppo_data_continuous = namedtuple(
'ppo_data_continuous',
['mu_sigma_new', 'mu_sigma_old', 'action', 'value_new', 'value_old', 'adv', 'return_', 'weight']
)
ppo_policy_data = namedtuple('ppo_policy_data', ['logit_new', 'logit_old', 'action', 'adv', 'weight'])
ppo_policy_data_continuous = namedtuple(
'ppo_policy_data_continuous', ['mu_sigma_new', 'mu_sigma_old', 'action', 'adv', 'weight']
)
ppo_value_data = namedtuple('ppo_value_data', ['value_new', 'value_old', 'return_', 'weight'])
ppo_loss = namedtuple('ppo_loss', ['policy_loss', 'value_loss', 'entropy_loss'])
ppo_policy_loss = namedtuple('ppo_policy_loss', ['policy_loss', 'entropy_loss'])
ppo_info = namedtuple('ppo_info', ['approx_kl', 'clipfrac'])
def shape_fn_ppo(args, kwargs):
r"""
Overview:
Return shape of ppo for hpc
Returns:
shape: [B, N]
"""
if len(args) <= 0:
tmp = kwargs['data'].logit_new.shape
else:
tmp = args[0].logit_new.shape
return tmp
@hpc_wrapper(
shape_fn=shape_fn_ppo,
namedtuple_data=True,
include_args=[0, 1, 2, 3],
include_kwargs=['data', 'clip_ratio', 'use_value_clip', 'dual_clip']
)
def ppo_error(
data: namedtuple,
clip_ratio: float = 0.2,
use_value_clip: bool = True,
dual_clip: Optional[float] = None
) -> Tuple[namedtuple, namedtuple]:
"""
Overview:
Implementation of Proximal Policy Optimization (arXiv:1707.06347) with value_clip and dual_clip
Arguments:
- data (:obj:`namedtuple`): the ppo input data with fieids shown in ``ppo_data``
- clip_ratio (:obj:`float`): the ppo clip ratio for the constraint of policy update, defaults to 0.2
- use_value_clip (:obj:`bool`): whether to use clip in value loss with the same ratio as policy
- dual_clip (:obj:`float`): a parameter c mentioned in arXiv:1912.09729 Equ. 5, shoule be in [1, inf),\
defaults to 5.0, if you don't want to use it, set this parameter to None
Returns:
- ppo_loss (:obj:`namedtuple`): the ppo loss item, all of them are the differentiable 0-dim tensor
- ppo_info (:obj:`namedtuple`): the ppo optim information for monitoring, all of them are Python scalar
Shapes:
- logit_new (:obj:`torch.FloatTensor`): :math:`(B, N)`, where B is batch size and N is action dim
- logit_old (:obj:`torch.FloatTensor`): :math:`(B, N)`
- action (:obj:`torch.LongTensor`): :math:`(B, )`
- value_new (:obj:`torch.FloatTensor`): :math:`(B, )`
- value_old (:obj:`torch.FloatTensor`): :math:`(B, )`
- adv (:obj:`torch.FloatTensor`): :math:`(B, )`
- return (:obj:`torch.FloatTensor`): :math:`(B, )`
- weight (:obj:`torch.FloatTensor` or :obj:`None`): :math:`(B, )`
- policy_loss (:obj:`torch.FloatTensor`): :math:`()`, 0-dim tensor
- value_loss (:obj:`torch.FloatTensor`): :math:`()`
- entropy_loss (:obj:`torch.FloatTensor`): :math:`()`
Examples:
>>> action_dim = 4
>>> data = ppo_data(
>>> logit_new=torch.randn(3, action_dim),
>>> logit_old=torch.randn(3, action_dim),
>>> action=torch.randint(0, action_dim, (3,)),
>>> value_new=torch.randn(3),
>>> value_old=torch.randn(3),
>>> adv=torch.randn(3),
>>> return_=torch.randn(3),
>>> weight=torch.ones(3),
>>> )
>>> loss, info = ppo_error(data)
.. note::
adv is already normalized value (adv - adv.mean()) / (adv.std() + 1e-8), and there are many
ways to calculate this mean and std, like among data buffer or train batch, so we don't couple
this part into ppo_error, you can refer to our examples for different ways.
"""
assert dual_clip is None or dual_clip > 1.0, "dual_clip value must be greater than 1.0, but get value: {}".format(
dual_clip
)
logit_new, logit_old, action, value_new, value_old, adv, return_, weight = data
policy_data = ppo_policy_data(logit_new, logit_old, action, adv, weight)
policy_output, policy_info = ppo_policy_error(policy_data, clip_ratio, dual_clip)
value_data = ppo_value_data(value_new, value_old, return_, weight)
value_loss = ppo_value_error(value_data, clip_ratio, use_value_clip)
return ppo_loss(policy_output.policy_loss, value_loss, policy_output.entropy_loss), policy_info
def ppo_policy_error(data: namedtuple,
clip_ratio: float = 0.2,
dual_clip: Optional[float] = None) -> Tuple[namedtuple, namedtuple]:
'''
Overview:
Get PPO policy loss
Arguments:
- data (:obj:`namedtuple`): ppo input data with fieids shown in ``ppo_policy_data``
- clip_ratio (:obj:`float`): clip value for ratio
- dual_clip (:obj:`float`): a parameter c mentioned in arXiv:1912.09729 Equ. 5, shoule be in [1, inf),\
defaults to 5.0, if you don't want to use it, set this parameter to None
Returns:
- ppo_policy_loss (:obj:`namedtuple`): the ppo policy loss item, all of them are the differentiable 0-dim tensor
- ppo_info (:obj:`namedtuple`): the ppo optim information for monitoring, all of them are Python scalar
Shapes:
- logit_new (:obj:`torch.FloatTensor`): :math:`(B, N)`, where B is batch size and N is action dim
- logit_old (:obj:`torch.FloatTensor`): :math:`(B, N)`
- action (:obj:`torch.LongTensor`): :math:`(B, )`
- adv (:obj:`torch.FloatTensor`): :math:`(B, )`
- weight (:obj:`torch.FloatTensor` or :obj:`None`): :math:`(B, )`
- policy_loss (:obj:`torch.FloatTensor`): :math:`()`, 0-dim tensor
- entropy_loss (:obj:`torch.FloatTensor`): :math:`()`
Examples:
>>> action_dim = 4
>>> data = ppo_policy_data(
>>> logit_new=torch.randn(3, action_dim),
>>> logit_old=torch.randn(3, action_dim),
>>> action=torch.randint(0, action_dim, (3,)),
>>> adv=torch.randn(3),
>>> weight=torch.ones(3),
>>> )
>>> loss, info = ppo_policy_error(data)
'''
logit_new, logit_old, action, adv, weight = data
if weight is None:
weight = torch.ones_like(adv)
dist_new = torch.distributions.categorical.Categorical(logits=logit_new)
dist_old = torch.distributions.categorical.Categorical(logits=logit_old)
logp_new = dist_new.log_prob(action)
logp_old = dist_old.log_prob(action)
dist_new_entropy = dist_new.entropy()
if dist_new_entropy.shape != weight.shape:
dist_new_entropy = dist_new.entropy().mean(dim=1)
entropy_loss = (dist_new_entropy * weight).mean()
# policy_loss
ratio = torch.exp(logp_new - logp_old)
if ratio.shape != adv.shape:
ratio = ratio.mean(dim=1)
surr1 = ratio * adv
surr2 = ratio.clamp(1 - clip_ratio, 1 + clip_ratio) * adv
if dual_clip is not None:
clip1 = torch.min(surr1, surr2)
clip2 = torch.max(clip1, dual_clip * adv)
# only use dual_clip when adv < 0
policy_loss = -(torch.where(adv < 0, clip2, clip1) * weight).mean()
else:
policy_loss = (-torch.min(surr1, surr2) * weight).mean()
with torch.no_grad():
approx_kl = (logp_old - logp_new).mean().item()
clipped = ratio.gt(1 + clip_ratio) | ratio.lt(1 - clip_ratio)
clipfrac = torch.as_tensor(clipped).float().mean().item()
return ppo_policy_loss(policy_loss, entropy_loss), ppo_info(approx_kl, clipfrac)
def ppo_value_error(
data: namedtuple,
clip_ratio: float = 0.2,
use_value_clip: bool = True,
) -> torch.Tensor:
'''
Overview:
Get PPO value loss
Arguments:
- data (:obj:`namedtuple`): ppo input data with fieids shown in ``ppo_value_data``
- clip_ratio (:obj:`float`): clip value for ratio
- use_value_clip (:obj:`bool`): whether use value clip
Returns:
- value_loss (:obj:`torch.FloatTensor`): the ppo value loss item, \
all of them are the differentiable 0-dim tensor
Shapes:
- value_new (:obj:`torch.FloatTensor`): :math:`(B, )`, where B is batch size
- value_old (:obj:`torch.FloatTensor`): :math:`(B, )`
- return (:obj:`torch.FloatTensor`): :math:`(B, )`
- weight (:obj:`torch.FloatTensor` or :obj:`None`): :math:`(B, )`
- value_loss (:obj:`torch.FloatTensor`): :math:`()`, 0-dim tensor
Examples:
>>> action_dim = 4
>>> data = ppo_value_data(
>>> value_new=torch.randn(3),
>>> value_old=torch.randn(3),
>>> return_=torch.randn(3),
>>> weight=torch.ones(3),
>>> )
>>> loss, info = ppo_value_error(data)
'''
value_new, value_old, return_, weight = data
if weight is None:
weight = torch.ones_like(value_old)
# value_loss
if use_value_clip:
value_clip = value_old + (value_new - value_old).clamp(-clip_ratio, clip_ratio)
v1 = (return_ - value_new).pow(2)
v2 = (return_ - value_clip).pow(2)
value_loss = 0.5 * (torch.max(v1, v2) * weight).mean()
else:
value_loss = 0.5 * ((return_ - value_new).pow(2) * weight).mean()
return value_loss
def ppo_error_continuous(
data: namedtuple,
clip_ratio: float = 0.2,
use_value_clip: bool = True,
dual_clip: Optional[float] = None
) -> Tuple[namedtuple, namedtuple]:
"""
Overview:
Implementation of Proximal Policy Optimization (arXiv:1707.06347) with value_clip and dual_clip
Arguments:
- data (:obj:`namedtuple`): the ppo input data with fieids shown in ``ppo_data``
- clip_ratio (:obj:`float`): the ppo clip ratio for the constraint of policy update, defaults to 0.2
- use_value_clip (:obj:`bool`): whether to use clip in value loss with the same ratio as policy
- dual_clip (:obj:`float`): a parameter c mentioned in arXiv:1912.09729 Equ. 5, shoule be in [1, inf),\
defaults to 5.0, if you don't want to use it, set this parameter to None
Returns:
- ppo_loss (:obj:`namedtuple`): the ppo loss item, all of them are the differentiable 0-dim tensor
- ppo_info (:obj:`namedtuple`): the ppo optim information for monitoring, all of them are Python scalar
Shapes:
- mu_sigma_new (:obj:`tuple`): :math:`((B, N), (B, N))`, where B is batch size and N is action dim
- mu_sigma_old (:obj:`tuple`): :math:`((B, N), (B, N))`, where B is batch size and N is action dim
- action (:obj:`torch.LongTensor`): :math:`(B, )`
- value_new (:obj:`torch.FloatTensor`): :math:`(B, )`
- value_old (:obj:`torch.FloatTensor`): :math:`(B, )`
- adv (:obj:`torch.FloatTensor`): :math:`(B, )`
- return (:obj:`torch.FloatTensor`): :math:`(B, )`
- weight (:obj:`torch.FloatTensor` or :obj:`None`): :math:`(B, )`
- policy_loss (:obj:`torch.FloatTensor`): :math:`()`, 0-dim tensor
- value_loss (:obj:`torch.FloatTensor`): :math:`()`
- entropy_loss (:obj:`torch.FloatTensor`): :math:`()`
Examples:
>>> action_dim = 4
>>> data = ppo_data_continuous(
>>> mu_sigma_new= dict(mu=torch.randn(3, action_dim), sigma=torch.randn(3, action_dim)**2),
>>> mu_sigma_old= dict(mu=torch.randn(3, action_dim), sigma=torch.randn(3, action_dim)**2),
>>> action=torch.randn(3, action_dim),
>>> value_new=torch.randn(3),
>>> value_old=torch.randn(3),
>>> adv=torch.randn(3),
>>> return_=torch.randn(3),
>>> weight=torch.ones(3),
>>> )
>>> loss, info = ppo_error(data)
.. note::
adv is already normalized value (adv - adv.mean()) / (adv.std() + 1e-8), and there are many
ways to calculate this mean and std, like among data buffer or train batch, so we don't couple
this part into ppo_error, you can refer to our examples for different ways.
"""
assert dual_clip is None or dual_clip > 1.0, "dual_clip value must be greater than 1.0, but get value: {}".format(
dual_clip
)
mu_sigma_new, mu_sigma_old, action, value_new, value_old, adv, return_, weight = data
if weight is None:
weight = torch.ones_like(adv)
dist_new = Independent(Normal(mu_sigma_new['mu'], mu_sigma_new['sigma']), 1)
if len(mu_sigma_old['mu'].shape) == 1:
dist_old = Independent(Normal(mu_sigma_old['mu'].unsqueeze(-1), mu_sigma_old['sigma'].unsqueeze(-1)), 1)
else:
dist_old = Independent(Normal(mu_sigma_old['mu'], mu_sigma_old['sigma']), 1)
logp_new = dist_new.log_prob(action)
logp_old = dist_old.log_prob(action)
entropy_loss = (dist_new.entropy() * weight).mean()
# policy_loss
ratio = torch.exp(logp_new - logp_old)
surr1 = ratio * adv
surr2 = ratio.clamp(1 - clip_ratio, 1 + clip_ratio) * adv
if dual_clip is not None:
policy_loss = (-torch.max(torch.min(surr1, surr2), dual_clip * adv) * weight).mean()
else:
policy_loss = (-torch.min(surr1, surr2) * weight).mean()
with torch.no_grad():
approx_kl = (logp_old - logp_new).mean().item()
clipped = ratio.gt(1 + clip_ratio) | ratio.lt(1 - clip_ratio)
clipfrac = torch.as_tensor(clipped).float().mean().item()
# value_loss
if use_value_clip:
value_clip = value_old + (value_new - value_old).clamp(-clip_ratio, clip_ratio)
v1 = (return_ - value_new).pow(2)
v2 = (return_ - value_clip).pow(2)
value_loss = 0.5 * (torch.max(v1, v2) * weight).mean()
else:
value_loss = 0.5 * ((return_ - value_new).pow(2) * weight).mean()
return ppo_loss(policy_loss, value_loss, entropy_loss), ppo_info(approx_kl, clipfrac)
def ppo_policy_error_continuous(data: namedtuple,
clip_ratio: float = 0.2,
dual_clip: Optional[float] = None) -> Tuple[namedtuple, namedtuple]:
"""
Overview:
Implementation of Proximal Policy Optimization (arXiv:1707.06347) with dual_clip
Arguments:
- data (:obj:`namedtuple`): the ppo input data with fieids shown in ``ppo_data``
- clip_ratio (:obj:`float`): the ppo clip ratio for the constraint of policy update, defaults to 0.2
- dual_clip (:obj:`float`): a parameter c mentioned in arXiv:1912.09729 Equ. 5, shoule be in [1, inf),\
defaults to 5.0, if you don't want to use it, set this parameter to None
Returns:
- ppo_loss (:obj:`namedtuple`): the ppo loss item, all of them are the differentiable 0-dim tensor
- ppo_info (:obj:`namedtuple`): the ppo optim information for monitoring, all of them are Python scalar
Shapes:
- mu_sigma_new (:obj:`tuple`): :math:`((B, N), (B, N))`, where B is batch size and N is action dim
- mu_sigma_old (:obj:`tuple`): :math:`((B, N), (B, N))`, where B is batch size and N is action dim
- action (:obj:`torch.LongTensor`): :math:`(B, )`
- adv (:obj:`torch.FloatTensor`): :math:`(B, )`
- weight (:obj:`torch.FloatTensor` or :obj:`None`): :math:`(B, )`
- policy_loss (:obj:`torch.FloatTensor`): :math:`()`, 0-dim tensor
- entropy_loss (:obj:`torch.FloatTensor`): :math:`()`
Examples:
>>> action_dim = 4
>>> data = ppo_policy_data_continuous(
>>> mu_sigma_new=dict(mu=torch.randn(3, action_dim), sigma=torch.randn(3, action_dim)**2),
>>> mu_sigma_old=dict(mu=torch.randn(3, action_dim), sigma=torch.randn(3, action_dim)**2),
>>> action=torch.randn(3, action_dim),
>>> adv=torch.randn(3),
>>> weight=torch.ones(3),
>>> )
>>> loss, info = ppo_policy_error_continuous(data)
"""
assert dual_clip is None or dual_clip > 1.0, "dual_clip value must be greater than 1.0, but get value: {}".format(
dual_clip
)
mu_sigma_new, mu_sigma_old, action, adv, weight = data
if weight is None:
weight = torch.ones_like(adv)
dist_new = Independent(Normal(mu_sigma_new['mu'], mu_sigma_new['sigma']), 1)
if len(mu_sigma_old['mu'].shape) == 1:
dist_old = Independent(Normal(mu_sigma_old['mu'].unsqueeze(-1), mu_sigma_old['sigma'].unsqueeze(-1)), 1)
else:
dist_old = Independent(Normal(mu_sigma_old['mu'], mu_sigma_old['sigma']), 1)
logp_new = dist_new.log_prob(action)
logp_old = dist_old.log_prob(action)
entropy_loss = (dist_new.entropy() * weight).mean()
# policy_loss
ratio = torch.exp(logp_new - logp_old)
surr1 = ratio * adv
surr2 = ratio.clamp(1 - clip_ratio, 1 + clip_ratio) * adv
if dual_clip is not None:
policy_loss = (-torch.max(torch.min(surr1, surr2), dual_clip * adv) * weight).mean()
else:
policy_loss = (-torch.min(surr1, surr2) * weight).mean()
with torch.no_grad():
approx_kl = (logp_old - logp_new).mean().item()
clipped = ratio.gt(1 + clip_ratio) | ratio.lt(1 - clip_ratio)
clipfrac = torch.as_tensor(clipped).float().mean().item()
return ppo_policy_loss(policy_loss, entropy_loss), ppo_info(approx_kl, clipfrac)
|