-
Notifications
You must be signed in to change notification settings - Fork 0
/
PG_Transformer(fail,discretized).py
199 lines (170 loc) · 6.39 KB
/
PG_Transformer(fail,discretized).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
"""
This is the Transformer version, where the state vector is 9 propositions = 9 x 3 = 27-vector
Refer to net_config() below for the current network topology and total number of weights info.
For example: h = (3-9-9) x 9, g = (9-12-9) x 1
Total # weights = (3 * 9 + 9 * 9) * 9 + 9 * 9 + 9 * 9 = 1134
Duplicate weights are counted because they are updated multiple times.
============================================================
Policy Gradient, Reinforcement Learning. Adapted from:
Morvan Zhou's tutorial page: https://morvanzhou.github.io/tutorials/
Using:
PyTorch: 1.9.0+cpu
gym: 0.8.0
"""
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
from torch.distributions import Categorical
# reproducible (this may be an overkill... but it works)
seed=666
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.use_deterministic_algorithms(True)
class PolicyGradient(nn.Module):
def __init__(
self,
n_actions,
n_features,
learning_rate,
gamma, # only for discounting within a game (episode), seems useless
):
super(PolicyGradient, self).__init__()
self.n_actions = n_actions
self.n_features = n_features
self.lr = learning_rate
self.gamma = gamma
# Episode data: actions, rewards:
self.ep_actions = Variable(torch.Tensor())
self.ep_rewards = []
self._build_net()
self.optimizer = optim.Adam(self.parameters(), lr=self.lr)
def net_info(self):
# Total number of params:
total = 0 # **** TO-DO
return ("?x?x?", total)
def _build_net(self):
encoder_layer = nn.TransformerEncoderLayer(d_model=3, nhead=3)
self.trm = nn.TransformerEncoder(encoder_layer, num_layers=4)
self.softmax = nn.Softmax(dim=0)
def forward(self, x):
# input dim = n_features = 9 x 3 = 27
# there are 9 h-networks each taking a dim-3 vector input
# First we need to split the input into 9 parts:
xs = torch.stack(torch.split(x, 3))
# print("xs =", xs)
y = self.trm(xs)
return y
# What is the output here? Old output = probs over actions
# The most reasonable output is: probability distribution over actions.
# But there is a waste of 3 dimensions
# Perhaps an even better output format is: probability distribution over specific [x,y]'s.
# Then we need to deal with the problem of merging duplicated [x,y] values.
# The duplicated probabilities could be added or maxed.
# P(A or B) = P(A) + P(B) - P(A and B) but it's difficult to estimate P(A and B)
# Non-monotonic logic is actually more economical as knowledge representation!
# Another question is: if the rule-base suggests action X with probability P1,
# and it also alternatively suggests the same action X with probability P2.
# So it is actually up to us to interpret P1 and P2.
# Naturally, we can normalize all P[i]'s to 1.
# But then X would be chosen with probability P1 + P2.
def choose_action(self, state):
# Select an action (0-9) by running policy model and choosing based on the probabilities in state
state = torch.from_numpy(state).type(torch.FloatTensor)
y = self(Variable(state))
# print("y =", y)
# y = 27-dim vector = [probability, X, Y] x 9
probs = self.softmax(torch.select(y, 1, 0)) # get every 1st element = probability
# print("probs =", probs)
c = Categorical(probs)
i = c.sample()
# print("i =", i)
def discretize(z):
if z > 0.5:
return 2
elif z < -0.5:
return 0
else:
return 1
action = discretize(y[i][1]) + discretize(y[i][2]) * 3
print(action, end='')
# **** Is it OK for "action" to be different from "i" (for back-prop) ?
# Yes, as long as the reward (for action "i") is assigned correctly.
# **** Explanation for failure of convergence:
# "i" is real number, but "action" is discretized.
# When i changes infinitesimally, "action" may remain unchanged.
# Thus the reward fails to reflect changes in learned parameters.
# Add log probability of our chosen action to our history
# Unsqueeze(0): tensor (prob, grad_fn) ==> ([prob], grad_fn)
log_prob = c.log_prob(i).unsqueeze(0)
# print("log prob:", c.log_prob(action))
# print("log prob unsqueezed:", log_prob)
if self.ep_actions.dim() != 0:
self.ep_actions = torch.cat([self.ep_actions, log_prob])
else:
self.ep_actions = (log_prob)
return action
def play_random(self, state, action_space):
# Select an action (0-9) randomly
# NOTE: random player never chooses occupied squares
while True:
action = action_space.sample()
x = action % 3
y = action // 3
occupied = False
for i in range(0, 27, 3): # scan through all 9 propositions, each proposition is a 3-vector
# 'proposition' is a numpy array[3]
proposition = state[i : i + 3]
# print("proposition=",proposition)
if ([x,y,1] == proposition).all():
occupied = True
break
if ([x,y,-1] == proposition).all():
occupied = True
break
if not occupied:
break
return action
def store_transition(self, s, a, r): # state, action, reward
# s is not needed, a is stored during choose_action().
self.ep_rewards.append(r)
def learn(self):
R = 0
rewards = []
# Discount future rewards back to the present using gamma
# print("\nLength of reward episode:", len(self.ep_rewards))
for r in self.ep_rewards[::-1]: # [::-1] reverses a list
R = r + self.gamma * R
rewards.insert(0, R)
# Scale rewards
#if len(rewards) == 1:
#rewards = torch.FloatTensor([-1.0])
#else:
rewards = torch.FloatTensor(rewards)
#rewards = (rewards - rewards.mean()) / (rewards.std() + np.finfo(np.float32).eps)
# print(rewards)
# Calculate loss
# print("policy history:", self.ep_actions)
# print("rewards:", rewards)
# loss = torch.sum(torch.mul(self.ep_actions, Variable(rewards)).mul(-1), -1)
loss = sum(torch.mul(self.ep_actions, Variable(rewards)).mul(-1), -1)
# Update network weights
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# Empty episode data
self.ep_actions = Variable(torch.Tensor())
self.ep_rewards = []
return rewards # == discounted_ep_rewards_norm
def save_net(self, fname):
torch.save(self.state_dict(), "PyTorch_models/" + fname + ".dict")
print("Model saved.")
def load_net(self, fname):
torch.load(self.state_dict(), fname + ".dict")
print("Model loaded.")