-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_train_gcn.py
207 lines (169 loc) · 6.49 KB
/
main_train_gcn.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
import torch
import gym
import sys
import time
from buffer.data_memory_gcn import Memory
from model.ppo_gcn import PPO
import numpy as np
import networkx as nx
#import matplotlib.pyplot as plt
from scipy.linalg import fractional_matrix_power
###### generate circuits graph ######
G = nx.Graph(name='twostageamp')
#Create nodes, considering there are 7 unique nodes in twostageamp. Note that the differential structure.
G.add_node(0, name='mp1')
G.add_node(1, name='mn1')
G.add_node(2, name='mp3')
G.add_node(3, name='mn3')
G.add_node(4, name='mn4')
G.add_node(5, name='mn5')
G.add_node(6, name='cc')
G.add_node(7, name='vdd')
G.add_node(8, name='gnd')
#Define the edges and the edges to the graph
edges = [(0, 1), (0, 2), (0, 6), (0, 7), (1, 2), (1, 3), (1, 6), (2, 5),
(2, 6), (2, 7), (3, 4), (3, 5), (3, 8), (4, 5), (4, 8), (5, 6), (5, 8)]
G.add_edges_from(edges)
#See graph info
print('Graph Info:\n', nx.info(G))
#Inspect the node features
print('\nGraph Nodes: ', G.nodes.data())
#Plot the graph
# nx.draw(G, with_labels=True, font_weight='bold')
# plt.show()
#Get the Adjacency Matrix (A) as numpy array
A = np.array(nx.attr_matrix(G)[0])
#Add Self Loops
G_self_loops = G.copy()
self_loops = []
for i in range(G.number_of_nodes()):
self_loops.append((i,i))
G_self_loops.add_edges_from(self_loops)
print('Edges of G with self-loops:\n', G_self_loops.edges)
#Get the Adjacency Matrix (A) of added self-lopps graph
A_hat = np.array(nx.attr_matrix(G_self_loops)[0])
#I = np.identity(A.shape[0]) # create Identity Matrix of A
#A_hat_1 = A + I
print('Adjacency Matrix of added self-loops G (A_hat):\n', A_hat)
#Get the Degree Matrix of the added self-loops graph
#Deg_Mat = G_self_loops.degree()
#print('Degree Matrix of added self-loops G (D): ', Deg_Mat)
#Convert the Degree Matrix to a N x N matrix where N is the number of nodes
#D = np.diag([deg for (n,deg) in list(Deg_Mat)])
D = np.diag(np.sum(A_hat, axis=0))
print('Degree Matrix of added self-loops G as numpy array (D):\n', D)
#Find the inverse of Degree Matrix (D)
#D_inv = np.linalg.inv(D)
#print('Inverse of D:\n', D_inv)
#Symmetrically-normalization
D_half_norm = fractional_matrix_power(D, -0.5)
DAD = D_half_norm.dot(A_hat).dot(D_half_norm)
print('DAD:\n', DAD)
#"""
wtdir = "/homes/wcao/Documents/ICML_21_AMP_EGCN_FC/train_log"
log = open(wtdir + '/train_gcn_s_2.log', 'a')
def mprint(s):
sys.stdout.write(time.strftime("%Y-%m-%d %H:%M:%S ") + s + "\n")
log.write(time.strftime("%Y-%m-%d %H:%M:%S ") + s + "\n")
sys.stdout.flush()
log.flush()
############## Hyperparameters ##############
env_name = "gym_twostageamp:twostageamp-v0" # creating environment
env = gym.make(env_name)
state_dim = 6 #env.observation_space.shape[0]
action_dim = 3
render = False
solved_reward = 1000 # stop training if avg_reward > solved_reward
log_interval = 25 # print avg reward in the interval
max_episodes = 50000 # max training episodes
max_timesteps = 50 # max timesteps in one episode
n_latent_var = 64 # number of variables in hidden layer
n_updata_episode = 30
update_timestep = n_updata_episode * max_timesteps # update policy every n timesteps
#lr = 0.002
lr = 0.0001
betas = (0.9, 0.9)
gamma = 0.95 # discount factor
K_epochs = 10 # update policy for K epochs
eps_clip = 0.3 # clip parameter for PPO
dropout = 0
#############################################
#
# def getlr(iter): ## get learning rate
# if iter < 2e3:
# return lr
# elif iter < 5e3:
# return 0.5*lr
# elif iter < 1e4:
# return 0.25*lr
# elif iter < 1.5e4:
# return 0.1*lr
# elif iter < 2e4:
# return 0.05*lr
# else:
# return 0.025*lr
#
memory = Memory()
ppo = PPO(state_dim, action_dim, n_latent_var, lr, betas, gamma, K_epochs, eps_clip)
#print(lr, betas)
# logging variables
running_reward = 0
avg_length = 0
timestep = 0
niter = 0
# training loop
mprint("Starting from Iteration %d" % niter)
for i_episode in range(1, max_episodes + 1):
#ppo = PPO(state_dim, action_dim, n_latent_var, getlr(i_episode), betas, gamma, K_epochs, eps_clip)
#ppo = PPO(lr =getlr(i_episode))
state = env.reset()
node_feature_type = np.array([[0, 0, 0, 0, 1], [0, 0, 0, 1, 0], [0, 0, 0, 0, 1], [0, 0, 0, 1, 0],
[0, 0, 0, 1, 0], [0, 0, 0, 1, 0], [0, 0, 1, 0, 0], [0, 1, 0, 0, 0], [1, 0, 0, 0, 0]])
device_node_feature_para = np.reshape(state[4:11], (len(state[4:11]), 1))
power_node_feature_para = np.array([[1.0], [0]])
node_feature_para = np.concatenate((device_node_feature_para, power_node_feature_para), axis=0)
state_gcn = np.concatenate((node_feature_type, node_feature_para), axis=1)
#state = np.reshape(state, (len(state), 1))
for t in range(max_timesteps):
timestep += 1
# Running policy_old:
action = ppo.policy_old.act(state_gcn, DAD, memory)
states, reward, done, _ = env.step(action)
device_node_feature_para = np.reshape(states[4:11] / 100, (len(states[4:11]), 1))
node_feature_para = np.concatenate((device_node_feature_para, power_node_feature_para), axis=0)
state_gcn = np.concatenate((node_feature_type, node_feature_para), axis=1)
# Saving reward and is_terminal:
memory.rewards.append(reward)
memory.is_terminals.append(done)
# update if its time
if timestep % update_timestep == 0:
ppo.update(DAD, memory)
memory.clear_memory()
timestep = 0
running_reward += reward
if render:
env.render()
if done:
break
avg_length += t
# stop training if avg_reward > solved_reward
target = log_interval * solved_reward
# logging
if i_episode % log_interval == 0:
avg_length = int(avg_length / log_interval)
running_reward = round(running_reward / log_interval, 3)
#running_reward = running_reward / log_interval
mprint("[%05d] lr = %.2e, Episode=[%05d], avg length=[%02d], reward = %.2e" % (niter, lr, i_episode, avg_length, running_reward))
# getlr(i_episode)
#print('Episode {} \t avg length: {} \t reward: {}'.format(i_episode, avg_length, running_reward))
torch.save(ppo.policy.state_dict(), './trained_model/PPO_{}.pth'.format(niter))
niter = niter + 1
if running_reward > target:
print("########## Solved! ##########")
torch.save(ppo.policy.state_dict(), './trained_model/PPO_{}.pth'.format(env_name))
break
running_reward = 0
avg_length = 0
mprint("Done!")
log.close()
#"""