-
Notifications
You must be signed in to change notification settings - Fork 47
/
train_tools.py
164 lines (140 loc) · 8.69 KB
/
train_tools.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
import os
import torch.optim as optim
import numpy as np
import tools
import time
from collections import deque
from storage import PCTRolloutStorage
from kfac import KFACOptimizer
import random
import torch
np.set_printoptions(threshold=np.inf)
class train_tools(object):
def __init__(self, writer, timeStr, PCT_policy, args):
self.writer = writer
self.timeStr = timeStr
self.step_counter = 0
self.PCT_policy = PCT_policy
self.use_acktr = args.use_acktr
seed = args.seed
if self.use_acktr:
self.policy_optim = KFACOptimizer(self.PCT_policy) # For ACKTR method. （https://proceedings.neurips.cc/paper/2017/hash/361440528766bbaaaa1901845cf4152b-Abstract.html）
else:
self.policy_optim = optim.Adam(self.PCT_policy.parameters(), lr=args.learning_rate) # For naive A2C method.
if seed is not None:
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
def train_n_steps(self, envs, args, device):
model_save_path = os.path.join(args.model_save_path, self.timeStr)
sub_time_str = time.strftime('%Y.%m.%d-%H-%M-%S', time.localtime(time.time()))
self.PCT_policy.train()
factor = args.normFactor # NormFactor controlls the maximum value of the original input of the network to less than 1.0, which helps the training of the network
obs = envs.reset()
all_nodes, leaf_nodes = tools.get_leaf_nodes(obs, args.internal_node_holder, args.leaf_node_holder)
all_nodes, leaf_nodes = all_nodes.to(device), leaf_nodes.to(device)
pct_rollout = PCTRolloutStorage(args.num_steps,
args.num_processes,
obs_shape=all_nodes.shape[1:],
gamma = args.gamma)
pct_rollout.to(device)
start = time.time()
ratio_recorder = 0
episode_rewards = deque(maxlen=10)
episode_ratio = deque(maxlen=10)
batchX = torch.arange(args.num_processes)
inside_counter = self.step_counter
num_steps, num_processes = args.num_steps, args.num_processes
pct_rollout.obs[0].copy_(all_nodes)
while True:
##############################################
####### Collect n-step training sample #######
##############################################
self.step_counter += 1
for step in range(num_steps):
with torch.no_grad():
selectedlogProb, selectedIdx, dist_entropy, _ = self.PCT_policy(all_nodes, normFactor = factor)
selected_leaf_node = leaf_nodes[batchX,selectedIdx.squeeze()]
obs, reward, done, infos = envs.step(selected_leaf_node.cpu().numpy())
all_nodes, leaf_nodes = tools.get_leaf_nodes(obs, args.internal_node_holder, args.leaf_node_holder)
all_nodes, leaf_nodes = all_nodes.to(device), leaf_nodes.to(device)
pct_rollout.insert(all_nodes, selectedIdx, selectedlogProb, reward, torch.tensor(1-done).unsqueeze(1))
for _ in range(len(infos)):
if done[_]:
if 'reward' in infos[_].keys():
episode_rewards.append(infos[_]['reward'])
else:
episode_rewards.append(infos[_]['episode']['r'])
if 'ratio' in infos[_].keys():
episode_ratio.append(infos[_]['ratio'])
with torch.no_grad():
_, _, _, next_value = self.PCT_policy(pct_rollout.obs[-1], normFactor = factor)
pct_rollout.compute_returns(next_value)
##############################################
########### PCT policy optimzation ###########
##############################################
obs_shape = pct_rollout.obs.size()[2:]
action_shape = pct_rollout.actions.size()[-1]
leaf_node_value, selectedlogProb, dist_entropy = self.PCT_policy.evaluate_actions(pct_rollout.obs[:-1].view(-1, *obs_shape),
pct_rollout.actions.view(-1, action_shape),
normFactor=factor)
leaf_node_value = leaf_node_value.view(num_steps, num_processes, 1)
selectedlogProb = selectedlogProb.view(num_steps, num_processes, 1)
advantages = pct_rollout.returns[:-1] - leaf_node_value
critic_loss = advantages.pow(2).mean()
actor_loss = -(advantages.detach() * selectedlogProb).mean()
if self.use_acktr and self.policy_optim.steps % self.policy_optim.Ts == 0:
# Sampled fisher, see Martens 2014d
self.PCT_policy.zero_grad()
pg_fisher_loss = - selectedlogProb.mean()
value_noise = torch.randn(leaf_node_value.size())
if leaf_node_value.is_cuda:
value_noise = value_noise.to(device)
sample_values = leaf_node_value + value_noise
vf_fisher_loss = -(leaf_node_value - sample_values.detach()).pow(2).mean()
fisher_loss = pg_fisher_loss + vf_fisher_loss
self.policy_optim.acc_stats = True
fisher_loss.backward(retain_graph=True)
self.policy_optim.acc_stats = False
self.policy_optim.zero_grad()
(args.actor_loss_coef * actor_loss
+ args.critic_loss_coef * critic_loss).backward()
torch.nn.utils.clip_grad_norm_(self.PCT_policy.parameters(), args.max_grad_norm)
self.policy_optim.step()
##############################################
############ After optimzation ###############
##############################################
pct_rollout.after_update()
# Save the trained policy model
if (self.step_counter % args.model_save_interval == 0) and args.model_save_path != "":
if self.step_counter % args.model_update_interval == 0:
sub_time_str = time.strftime('%Y.%m.%d-%H-%M-%S', time.localtime(time.time()))
if not os.path.exists(model_save_path):
os.makedirs(model_save_path)
torch.save(self.PCT_policy.state_dict(), os.path.join(model_save_path, 'PCT-' + self.timeStr + '_' + sub_time_str + ".pt"))
# Write tensorboard logs.
if self.step_counter % args.print_log_interval == 0 and len(episode_rewards)>1:
total_num_steps = (self.step_counter + 1 - inside_counter) * num_processes * num_steps
end = time.time()
if len(episode_ratio) != 0:
ratio_recorder = max(ratio_recorder, np.max(episode_ratio))
print("Time version: {} is training\n"
"Updates {}, num timesteps {}, FPS {}\n"
"Last {} training episodes: mean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}\n"
"The dist entropy {:.5f}, the value loss {:.5f}, the action loss {:.5f}\n"
"The mean space ratio is {}, the ratio threshold is{}\n"
.format(self.timeStr,
self.step_counter, total_num_steps, int(total_num_steps / (end - start)),
len(episode_rewards), np.mean(episode_rewards), np.median(episode_rewards), np.min(episode_rewards), np.max(episode_rewards),
dist_entropy.item(), critic_loss.item(), actor_loss.item(),
np.mean(episode_ratio), ratio_recorder))
self.writer.add_scalar('Rewards/Mean rewards', np.mean(episode_rewards), self.step_counter)
self.writer.add_scalar("Rewards/Max rewards", np.max(episode_rewards), self.step_counter)
self.writer.add_scalar('Rewards/Min rewards', np.min(episode_rewards), self.step_counter)
self.writer.add_scalar("Ratio/The max ratio", np.max(episode_ratio), self.step_counter)
self.writer.add_scalar("Ratio/The mean ratio", np.mean(episode_ratio), self.step_counter)
self.writer.add_scalar("Ratio/The max ratio in history", ratio_recorder, self.step_counter)
self.writer.add_scalar("Training/Value loss", critic_loss.item(), self.step_counter)
self.writer.add_scalar("Training/Action loss", actor_loss.item(), self.step_counter)
self.writer.add_scalar('Training/Distribution entropy', dist_entropy.item(), self.step_counter)