-
Notifications
You must be signed in to change notification settings - Fork 0
/
agent.py
186 lines (147 loc) · 6.37 KB
/
agent.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
import numpy as np
import torch
import torch.nn as nn
import pickle
import os
from NN import DQNetwork
from Memory import Memory
from settings import DATA_DIRECTORY
from rewards import *
np.set_printoptions(precision=2)
np.set_printoptions(suppress=True)
class Agent:
def __init__(self, name, training= False, input_dim = False, gamma = False, burnin = False, epsilon = False, learning_rate = False,
action_dim = False, batch_size = False, eps_min = False, eps_dec = False, replace = False, memory_size = False):
self.save_dir = DATA_DIRECTORY
self.name = name
self.is_training = training
if self.is_training:
self.input_dim = input_dim
self.gamma = gamma
self.burnin = burnin
self.learning_rate = learning_rate
self.action_dim = action_dim
self.batch_size = batch_size
self.eps_min = eps_min
self.eps_dec = eps_dec
self.replace = replace
self.save_dir = DATA_DIRECTORY
self.save_step = 0
self.memory_size = memory_size
self.action_space = list(range(self.action_dim))
self.learn_step_counter = 0
self.memory = Memory(self.memory_size, input_dim)
self.q_online = DQNetwork(self.input_dim, self.action_dim,
self.learning_rate, training=self.is_training)
self.q_target = DQNetwork(self.input_dim, self.action_dim,
self.learning_rate)
self.epsilon = epsilon
directory = f'{self.save_dir}/{self.name}'
if os.path.exists(directory):
self.load()
if learning_rate:
self.learning_rate = learning_rate
if epsilon:
self.epsilon = epsilon
if not self.is_training:
self.epsilon = 0
@torch.no_grad()
def choose_action(self, observation, output=False, force_exploit=False):
if np.random.random() > self.epsilon or force_exploit:
state = torch.tensor([observation], dtype=torch.float32, requires_grad=False)
q_values = self.q_online(state).detach().numpy().flatten()
action = np.argmax(q_values)
if output:
print('KI sieht {}\nund sagt {}\nq_values:{}\n'.format(observation, action, q_values))
return action, True
else:
action = np.random.randint(self.action_dim)
return action, False
@torch.no_grad()
def store_transition(self, state, action, reward, new_state, done):
self.memory.store(state, action, reward, new_state, done)
@torch.no_grad()
def sample_memory(self, batch_size):
old_states, actions, rewards, new_states, dones = self.memory.sample(batch_size)
return old_states, actions, rewards, new_states, dones
@torch.no_grad()
def replace_target_network(self):
if self.learn_step_counter % self.replace == 0:
self.q_target.load_state_dict(self.q_online.state_dict())
self.q_target.zero_grad()
for p in self.q_target.parameters():
p.requires_grad = False
@torch.no_grad()
def decrement_epsilon(self):
self.epsilon = max(self.epsilon - self.eps_dec, self.eps_min)
@torch.no_grad()
def save(self):
directory = f'{self.save_dir}/{self.name}'
save_state = {
'q_online': self.q_online.state_dict(),
'q_target': self.q_target.state_dict(),
'epsilon': self.epsilon,
'save_step': self.save_step,
'input_dim': self.input_dim,
'gamma': self.gamma,
'burnin': self.burnin,
'learning_rate': self.learning_rate,
'action_dim': self.action_dim,
}
if not os.path.exists(directory):
os.makedirs(directory)
with open(f'{directory}/{self.save_step}.step', 'wb') as f:
pickle.dump(save_state, f)
print(f'saved: {directory}/{self.save_step}.step')
self.save_step += 1
@torch.no_grad()
def load(self, save_step=False, save_file=None):
directory = f'{self.save_dir}/{self.name}'
if type(save_step) == int:
self.save_step = save_step
else:
self.save_step = 0
for f in os.listdir(directory):
value, ext = os.path.splitext(f)
if ext == '.step':
if int(value) > self.save_step:
self.save_step = int(value)
if save_file is None:
with open(f'{directory}/{self.save_step}.step', 'rb') as f:
save_state = pickle.load(f)
print(f'loaded: {directory}/{self.save_step}.step')
else:
with open(save_file, 'rb') as f:
save_state = pickle.load(f)
self.save_step
self.input_dim = save_state['input_dim']
self.gamma = save_state['gamma']
self.burnin = save_state['burnin']
self.learning_rate = save_state['learning_rate']
self.action_dim = save_state['action_dim']
self.q_online = DQNetwork(self.input_dim, self.action_dim,
self.learning_rate, training=self.is_training)
self.q_target = DQNetwork(self.input_dim, self.action_dim,
self.learning_rate)
self.q_online.load_state_dict(save_state['q_online'])
self.q_target.load_state_dict(save_state['q_target'])
self.save_step = save_state['save_step']
def learn(self):
if len(self.memory) < max(self.batch_size, self.burnin):
return None
self.q_online.optimizer.zero_grad()
self.replace_target_network()
old_states, actions, rewards, new_states, dones = self.sample_memory(self.batch_size)
indices = torch.arange(self.batch_size, dtype=torch.int64)
q_pred = self.q_online(old_states)[indices, actions]
q_actual = self.q_target(new_states).max(dim=1)[0]
q_actual[dones] = 0.0
q_actual = rewards + self.gamma*q_actual
loss = self.q_online.loss(q_actual, q_pred)
loss.backward()
self.q_online.optimizer.step()
self.learn_step_counter += 1
self.decrement_epsilon()
return loss
def get_print_data(self):
return (self.name, self.gamma, self.learning_rate)