-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathppo_update.py
182 lines (168 loc) · 8.67 KB
/
ppo_update.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
# Created by Andrew Silva, andrew.silva@gatech.edu
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.distributions import Categorical
import numpy as np
class PPO:
def __init__(self, actor_critic_arr, two_nets=True):
lr = 1e-3
eps = 1e-5
self.clip_param = 0.2
self.ppo_epoch = 8
self.num_mini_batch = 4
self.value_loss_coef = 0.05
self.entropy_coef = 0.01
self.max_grad_norm = 0.5
if two_nets:
self.actor = actor_critic_arr[0]
self.critic = actor_critic_arr[1]
if self.actor.input_dim > 100:
self.actor_opt = optim.RMSprop(self.actor.parameters(), lr=1e-4)
self.critic_opt = optim.RMSprop(self.critic.parameters(), lr=1e-4)
elif self.actor.input_dim >= 8:
self.actor_opt = optim.RMSprop(self.actor.parameters(), lr=1e-3)
self.critic_opt = optim.RMSprop(self.critic.parameters(), lr=1e-3)
else:
self.actor_opt = optim.RMSprop(self.actor.parameters(), lr=1e-2)
self.critic_opt = optim.RMSprop(self.critic.parameters(), lr=1e-2)
else:
self.actor = actor_critic_arr
self.actor_opt = optim.Adam(self.actor.parameters(), lr=lr, eps=eps)
self.two_nets = two_nets
self.epoch_counter = 0
def cartpole_update(self, rollouts, agent_in, go_deeper=False):
if self.actor.input_dim < 10:
batch_size = max(rollouts.step // 32, 1)
num_iters = rollouts.step // batch_size
else:
num_iters = 8
batch_size = 4
total_action_loss = torch.Tensor([0])
total_value_loss = torch.Tensor([0])
for iteration in range(num_iters):
total_action_loss = torch.Tensor([0])
total_value_loss = torch.Tensor([0])
if go_deeper:
deep_total_action_loss = torch.Tensor([0])
deep_total_value_loss = torch.Tensor([0])
for b in range(batch_size):
sample = rollouts.sample()
if not sample:
break
state = sample['state']
action_probs = sample['action_prob']
adv_targ = torch.Tensor([sample['advantage']])
reward = sample['reward']
old_action_probs = sample['full_prob_vector']
if np.isnan(adv_targ) or np.isnan(reward) or True in np.isnan(old_action_probs):
continue
action_taken = sample['action_taken']
hidden_state = sample['hidden_state']
if hidden_state is not None:
new_action_probs, _ = self.actor(*state, hidden_state[0])
new_value, _ = self.critic(*state, hidden_state[1])
else:
new_action_probs = self.actor(*state)
new_value = self.critic(*state)
if go_deeper:
deep_action_probs = sample['deeper_action_prob']
deep_adv = torch.Tensor([sample['deeper_advantage']])
deeper_old_probs = sample['deeper_full_prob_vector']
new_deep_probs = agent_in.deeper_action_network(*state)
new_deep_vals = agent_in.deeper_value_network(*state)
deep_dist = Categorical(new_deep_probs)
deeper_probs = deep_dist.log_prob(action_taken)
deeper_val = new_deep_vals[action_taken.item()]
deeper_entropy = deep_dist.entropy().mean() * self.entropy_coef
# deep_ratio = torch.div(deeper_probs, deep_action_probs)
deep_ratio = torch.nn.functional.kl_div(new_deep_probs, deeper_old_probs, reduction='batchmean')
deep_surr1 = deep_ratio.mul(deep_adv).mul(deeper_probs)
deep_surr2 = torch.clamp(deep_ratio, 1.0 - self.clip_param,
1.0 + self.clip_param).pow(-1).mul(deep_adv).mul(deeper_probs)
deep_action_loss = -torch.min(deep_surr1, deep_surr2).mean()
deep_total_action_loss = deep_total_action_loss + deep_action_loss - deeper_entropy
deeper_val = deeper_val.view(-1, 1)
copy_reward = torch.FloatTensor([reward]).view(-1, 1)
deeper_value_loss = F.mse_loss(copy_reward, deeper_val)
deep_total_value_loss = deep_total_value_loss + deeper_value_loss
update_m = Categorical(new_action_probs)
update_log_probs = update_m.log_prob(action_taken)
new_value = new_value[action_taken.item()]
entropy = update_m.entropy().mean() * self.entropy_coef
# ratio = torch.div(update_log_probs, action_probs)
ratio = torch.nn.functional.kl_div(new_action_probs, old_action_probs, reduction='batchmean')
clipped = torch.clamp(ratio, 1.0 - self.clip_param,
1.0 + self.clip_param).mul(adv_targ).mul(update_log_probs)
ratio = ratio.mul(adv_targ).mul(update_log_probs)
action_loss = -torch.min(ratio, clipped).mean()
new_value = new_value.view(-1, 1)
reward = torch.FloatTensor([reward]).view(-1, 1)
value_loss = F.mse_loss(reward, new_value)
total_value_loss = total_value_loss + value_loss
total_action_loss = total_action_loss + action_loss - entropy
if total_value_loss != 0:
nn.utils.clip_grad_norm_(self.critic.parameters(), self.max_grad_norm)
self.critic_opt.zero_grad()
total_value_loss.backward()
self.critic_opt.step()
if total_action_loss != 0:
nn.utils.clip_grad_norm_(self.actor.parameters(), self.max_grad_norm)
self.actor_opt.zero_grad()
total_action_loss.backward()
self.actor_opt.step()
if go_deeper:
if deep_total_value_loss != 0:
nn.utils.clip_grad_norm_(agent_in.deeper_value_network.parameters(), self.max_grad_norm)
agent_in.deeper_value_opt.zero_grad()
deep_total_value_loss.backward()
agent_in.deeper_value_opt.step()
if deep_total_action_loss != 0:
nn.utils.clip_grad_norm_(agent_in.deeper_action_network.parameters(), self.max_grad_norm)
agent_in.deeper_actor_opt.zero_grad()
deep_total_action_loss.backward()
agent_in.deeper_actor_opt.step()
agent_in.deepen_networks()
agent_in.reset()
return total_action_loss.item(), total_value_loss.item()
def sl_updates(self, rollouts, agent_in, heuristic_teacher):
if self.actor.input_dim < 10:
batch_size = max(rollouts.step // 32, 1)
num_iters = rollouts.step // batch_size
else:
num_iters = 8
batch_size = 4
aggregate_actor_loss = 0
for iteration in range(num_iters):
total_action_loss = torch.Tensor([0])
total_value_loss = torch.Tensor([0])
for b in range(batch_size):
sample = rollouts.sample()
if not sample:
break
state = sample['state']
reward = sample['reward']
if np.isnan(reward):
continue
new_action_probs = self.actor(*state).view(1, -1)
new_value = self.critic(*state)
label = torch.LongTensor([heuristic_teacher.get_action(state[0].detach().clone().data.cpu().numpy()[0])])
action_loss = torch.nn.functional.cross_entropy(new_action_probs, label)
new_value = new_value.view(-1, 1)
reward = torch.Tensor([reward]).view(-1, 1)
value_loss = F.mse_loss(reward, new_value)
total_value_loss = total_value_loss + value_loss
total_action_loss = total_action_loss + action_loss
if total_value_loss != 0:
self.critic_opt.zero_grad()
total_value_loss.backward()
self.critic_opt.step()
if total_action_loss != 0:
self.actor_opt.zero_grad()
total_action_loss.backward()
self.actor_opt.step()
aggregate_actor_loss += total_action_loss.item()
aggregate_actor_loss /= float(num_iters*batch_size)
agent_in.reset()
return aggregate_actor_loss