-
Notifications
You must be signed in to change notification settings - Fork 41
/
eupg.py
338 lines (288 loc) · 11.9 KB
/
eupg.py
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
"""EUPG is an ESR algorithm based on Policy Gradient (REINFORCE like)."""
import time
from copy import deepcopy
from typing import Callable, List, Optional, Union
from typing_extensions import override
import gymnasium as gym
import numpy as np
import torch as th
import torch.nn as nn
import torch.optim as optim
import wandb
from torch.distributions import Categorical
from morl_baselines.common.accrued_reward_buffer import AccruedRewardReplayBuffer
from morl_baselines.common.evaluation import log_episode_info
from morl_baselines.common.morl_algorithm import MOAgent, MOPolicy
from morl_baselines.common.networks import layer_init, mlp
class PolicyNet(nn.Module):
"""Policy network."""
def __init__(self, obs_shape, action_dim, rew_dim, net_arch):
"""Initialize the policy network.
Args:
obs_shape: Observation shape
action_dim: Action dimension
rew_dim: Reward dimension
net_arch: Number of units per layer
"""
super().__init__()
self.obs_shape = obs_shape
self.action_dim = action_dim
self.rew_dim = rew_dim
# Conditioned on accrued reward, so input takes reward
input_dim = obs_shape[0] + rew_dim
# |S|+|R| -> ... -> |A|
self.net = mlp(input_dim, action_dim, net_arch, activation_fn=nn.Tanh)
self.apply(layer_init)
def forward(self, obs: th.Tensor, acc_reward: th.Tensor):
"""Forward pass.
Args:
obs: Observation
acc_reward: accrued reward
Returns: Probability of each action
"""
input = th.cat((obs, acc_reward), dim=acc_reward.dim() - 1)
pi = self.net(input)
# Normalized sigmoid
x_exp = th.sigmoid(pi)
probas = x_exp / th.sum(x_exp)
return probas.view(-1, self.action_dim) # Batch Size x |Actions|
def distribution(self, obs: th.Tensor, acc_reward: th.Tensor):
"""Categorical distribution based on the action probabilities.
Args:
obs: observation
acc_reward: accrued reward
Returns: action distribution.
"""
probas = self.forward(obs, acc_reward)
distribution = Categorical(probas)
return distribution
class EUPG(MOPolicy, MOAgent):
"""Expected Utility Policy Gradient Algorithm.
The idea is to condition the network on the accrued reward and to scalarize the rewards based on the episodic return (accrued + future rewards)
Paper: D. Roijers, D. Steckelmacher, and A. Nowe, Multi-objective Reinforcement Learning for the Expected Utility of the Return. 2018.
"""
def __init__(
self,
env: gym.Env,
scalarization: Callable[[np.ndarray, np.ndarray], float],
weights: np.ndarray = np.ones(2),
id: Optional[int] = None,
buffer_size: int = int(1e5),
net_arch: List = [50],
gamma: float = 0.99,
learning_rate: float = 1e-3,
project_name: str = "MORL-Baselines",
experiment_name: str = "EUPG",
wandb_entity: Optional[str] = None,
log: bool = True,
log_every: int = 1000,
device: Union[th.device, str] = "auto",
seed: Optional[int] = None,
parent_rng: Optional[np.random.Generator] = None,
):
"""Initialize the EUPG algorithm.
Args:
env: Environment
scalarization: Scalarization function to use (can be non-linear)
weights: Weights to use for the scalarization function
id: Id of the agent (for logging)
buffer_size: Size of the replay buffer
net_arch: Number of units per layer
gamma: Discount factor
learning_rate: Learning rate (alpha)
project_name: Name of the project (for logging)
experiment_name: Name of the experiment (for logging)
wandb_entity: Entity to use for wandb
log: Whether to log or not
log_every: Log every n episodes
device: Device to use for NN. Can be "cpu", "cuda" or "auto".
seed: Seed for the random number generator
parent_rng: Parent random number generator (for reproducibility)
"""
MOAgent.__init__(self, env, device, seed=seed)
MOPolicy.__init__(self, None, device)
# Seeding
self.seed = seed
self.parent_rng = parent_rng
if parent_rng is not None:
self.np_random = parent_rng
else:
self.np_random = np.random.default_rng(self.seed)
self.env = env
self.id = id
# RL
self.scalarization = scalarization
self.weights = weights
self.gamma = gamma
# Learning
self.buffer_size = buffer_size
self.net_arch = net_arch
self.learning_rate = learning_rate
self.buffer = AccruedRewardReplayBuffer(
obs_shape=self.observation_shape,
action_shape=self.action_shape,
rew_dim=self.reward_dim,
max_size=self.buffer_size,
obs_dtype=np.int32,
action_dtype=np.int32,
)
self.net = PolicyNet(
obs_shape=self.observation_shape,
rew_dim=self.reward_dim,
action_dim=self.action_dim,
net_arch=self.net_arch,
).to(self.device)
self.optimizer = optim.Adam(self.net.parameters(), lr=self.learning_rate)
# Logging
self.project_name = project_name
self.experiment_name = experiment_name
self.log = log
self.log_every = log_every
if log and parent_rng is None:
self.setup_wandb(self.project_name, self.experiment_name, wandb_entity)
def __deepcopy__(self, memo):
"""Deep copy the policy."""
copied_net = deepcopy(self.net)
copied = type(self)(
self.env,
self.scalarization,
self.weights,
self.id,
self.buffer_size,
self.net_arch,
self.gamma,
self.learning_rate,
self.project_name,
self.experiment_name,
log=self.log,
device=self.device,
parent_rng=self.parent_rng,
)
copied.global_step = self.global_step
copied.optimizer = optim.Adam(copied_net.parameters(), lr=self.learning_rate)
copied.buffer = deepcopy(self.buffer)
return copied
@override
def get_policy_net(self) -> nn.Module:
return self.net
@override
def get_buffer(self):
return self.buffer
@override
def set_buffer(self, buffer):
raise Exception("On-policy algorithms should not share buffer.")
@override
def set_weights(self, weights: np.ndarray):
self.weights = weights
@th.no_grad()
@override
def eval(self, obs: np.ndarray, accrued_reward: Optional[np.ndarray]) -> Union[int, np.ndarray]:
if type(obs) is int:
obs = th.as_tensor([obs]).to(self.device)
else:
obs = th.as_tensor(obs).to(self.device)
accrued_reward = th.as_tensor(accrued_reward).float().to(self.device)
return self.__choose_action(obs, accrued_reward)
@th.no_grad()
def __choose_action(self, obs: th.Tensor, accrued_reward: th.Tensor) -> int:
action = self.net.distribution(obs, accrued_reward)
action = action.sample().detach().item()
return action
@override
def update(self):
(
obs,
accrued_rewards,
actions,
rewards,
next_obs,
terminateds,
) = self.buffer.get_all_data(to_tensor=True, device=self.device)
episodic_return = th.sum(rewards, dim=0)
scalarized_return = self.scalarization(episodic_return.cpu().numpy(), self.weights)
scalarized_return = th.scalar_tensor(scalarized_return).to(self.device)
discounted_forward_rewards = self._forward_cumulative_rewards(rewards)
scalarized_values = self.scalarization(discounted_forward_rewards)
# For each sample in the batch, get the distribution over actions
current_distribution = self.net.distribution(obs, accrued_rewards)
# Policy gradient
log_probs = current_distribution.log_prob(actions.squeeze())
loss = -th.mean(log_probs * scalarized_values)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if self.log:
log_str = f"_{self.id}" if self.id is not None else ""
wandb.log(
{
f"losses{log_str}/loss": loss,
f"metrics{log_str}/scalarized_episodic_return": scalarized_return,
"global_step": self.global_step,
},
)
def _forward_cumulative_rewards(self, rewards):
flip_rewards = rewards.flip(dims=[0])
cumulative_rewards = th.zeros(self.reward_dim).to(self.device)
for i in range(len(rewards)):
cumulative_rewards = self.gamma * cumulative_rewards + flip_rewards[i]
flip_rewards[i] = cumulative_rewards
forward_rewards = flip_rewards.flip(dims=[0])
return forward_rewards
def train(self, total_timesteps: int, eval_env: Optional[gym.Env] = None, eval_freq: int = 1000, start_time=None):
"""Train the agent.
Args:
total_timesteps: Number of timesteps to train for
eval_env: Environment to run policy evaluation on
eval_freq: Frequency of policy evaluation
start_time: Start time of the training (for SPS)
"""
if start_time is None:
start_time = time.time()
# Init
(
obs,
_,
) = self.env.reset()
accrued_reward_tensor = th.zeros(self.reward_dim, dtype=th.float32).float().to(self.device)
# Training loop
for _ in range(1, total_timesteps + 1):
self.global_step += 1
with th.no_grad():
# For training, takes action according to the policy
action = self.__choose_action(th.Tensor([obs]).to(self.device), accrued_reward_tensor)
next_obs, vec_reward, terminated, truncated, info = self.env.step(action)
# Memory update
self.buffer.add(obs, accrued_reward_tensor.cpu().numpy(), action, vec_reward, next_obs, terminated)
accrued_reward_tensor += th.from_numpy(vec_reward).to(self.device)
if eval_env is not None and self.log and self.global_step % eval_freq == 0:
self.policy_eval_esr(eval_env, scalarization=self.scalarization, weights=self.weights, log=self.log)
if terminated or truncated:
# NN is updated at the end of each episode
self.update()
self.buffer.cleanup()
obs, _ = self.env.reset()
self.num_episodes += 1
accrued_reward_tensor = th.zeros(self.reward_dim).float().to(self.device)
if self.log and self.num_episodes % self.log_every == 0 and "episode" in info.keys():
log_episode_info(
info=info["episode"],
scalarization=self.scalarization,
weights=self.weights,
id=self.id,
global_timestep=self.global_step,
)
else:
obs = next_obs
if self.log and self.global_step % 1000 == 0:
print("SPS:", int(self.global_step / (time.time() - start_time)))
wandb.log({"charts/SPS": int(self.global_step / (time.time() - start_time)), "global_step": self.global_step})
@override
def get_config(self) -> dict:
return {
"env_id": self.env.unwrapped.spec.id,
"learning_rate": self.learning_rate,
"buffer_size": self.buffer_size,
"gamma": self.gamma,
"net_arch": self.net_arch,
"seed": self.seed,
}