-
Notifications
You must be signed in to change notification settings - Fork 10
/
test.py
236 lines (207 loc) · 9.95 KB
/
test.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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import os
import cv2
from datetime import datetime
import argparse
import warnings
warnings.filterwarnings('ignore')
import numpy as np
import random
import tensorflow as tf
# from xvfbwrapper import Xvfb
from models.missVAE import missVAE
from models.dqn_pol import DoubleDQNAgent
from utils.policy.extract import extract_theta, encode_obs
from utils.policy.min_n_suff_set import min_n_suff_set_state, min_n_suff_set_theta
from utils.misc.data_handler import DataHandler
from utils.misc.hyper_params import default_hps
from utils.misc.env_init import env_init
def infer_theta(n_epoch, n_domain, hps, model, dh, save_p):
N_data = 1000 * n_domain
N_batches = int(np.floor(N_data / hps.batch_size))
sign = 1
for epoch in range(n_epoch):
for idx in range(N_batches):
step = model.sess.run(model.global_step)
curr_learning_rate = \
(hps.learning_rate - hps.min_learning_rate) * hps.decay_rate ** step + hps.min_learning_rate
batch_obs, batch_action, batch_reward, batch_domain_index = dh.next_batch()
action_init = np.zeros((hps.batch_size, 1, hps.action_size))
batch_action_prev = np.concatenate((action_init, batch_action[:, :-1, :]), axis=1)
reward_init = np.zeros((hps.batch_size, 1, hps.reward_size))
batch_reward_prev = np.concatenate((reward_init, batch_reward[:, :-1, :]), axis=1)
batch_obs = np.reshape(batch_obs, (hps.batch_size, hps.max_seq_len, 128, 128, 1))
batch_domain_index = np.reshape(batch_domain_index, [hps.batch_size])
feed = {model.input_x: batch_obs,
model.input_a_prev: batch_action_prev,
model.input_a: batch_action,
model.input_r_prev: batch_reward_prev,
model.input_r: batch_reward,
model.input_domain_index: batch_domain_index,
model.lr: curr_learning_rate,
model.seq_length: hps.max_seq_len,
model.input_sign: sign}
(total_loss,
vae_loss, vae_r_obs_loss, vae_r_next_obs_loss, vae_r_reward_loss, vae_kl_loss, vae_causal_filter_loss,
transition_loss, causal_filter_loss, state, _) \
= model.sess.run([model.total_loss,
model.vae_loss,
model.r_obs_loss,
model.r_next_obs_loss,
model.r_reward_loss,
model.kl_loss,
model.vae_causal_filter_loss,
model.transition_loss,
model.causal_filter_loss,
model.final_state,
model.train_op], feed)
if vae_loss < 50:
sign = 0
if step % 100 == 0:
SSL_A, SSL_B, SSL_C, SSL_D, SSL_E, SSL_F, theta_o, theta_s, theta_r = model.sess.run(
[model.SSL_A, model.SSL_B, model.SSL_C,
model.SSL_D, model.SSL_E, model.SSL_F,
model.theta_o, model.theta_s, model.theta_r])
print("theta_o:", theta_o)
print("theta_r:", theta_r)
print("theta_s:", theta_s)
output_log = "step: %d (Epoch: %d idx: %d), " \
"total_loss: %.4f, " \
"vae_loss: %.4f, " \
"vae_r_obs_loss: %.4f, " \
"vae_r_next_obs_loss: %.4f, " \
"vae_r_reward_loss: %.4f, " \
"vae_kl_loss: %.4f, " \
"vae_causal_filter_loss: %.4f, " \
"transition_loss: %.4f, " \
"causal_filter_loss: %.4f," \
% (step,
epoch,
idx,
total_loss,
vae_loss, vae_r_obs_loss, vae_r_next_obs_loss,
vae_r_reward_loss, vae_kl_loss, vae_causal_filter_loss,
transition_loss,
causal_filter_loss)
print(output_log)
f = open(os.path.join(save_p, 'output.txt'), 'a')
f.write(output_log + '\n')
f.close()
save_p_e = os.path.join(save_p, 'epochs')
if not os.path.exists(save_p_e):
os.makedirs(save_p_e)
if step % 500 == 0:
model.save_json(os.path.join(save_p_e, str(epoch) + '_' + str(step) + 'test.json'))
model.save_json(os.path.join(save_p, 'test.json'))
def predict(env, agent, model, theta_all, k):
score = 0
# initialization
state_ori = env[k].reset() # the ground-truth state
# generate observational image
obs = env[k].render(mode='rgb_array')
obs = cv2.cvtColor(obs, cv2.COLOR_RGB2GRAY)
obs = cv2.resize(obs, (128, 128), interpolation=cv2.INTER_CUBIC)
obs = cv2.normalize(obs, None, alpha=-1, beta=1, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
a_prev, r_prev, c_prev = mvae.reset()
state, c = encode_obs(hps, model, obs, a_prev, r_prev, c_prev, k)
state = np.reshape(state, [1, hps.z_size])
while 1:
# get action for the current observation and go one step in environment
action = agent.get_action(state, theta_all[k])
next_state_ori, reward, done, info = env[k].step(action)
next_obs = env[k].render(mode='rgb_array')
next_obs = cv2.cvtColor(next_obs, cv2.COLOR_RGB2GRAY)
next_obs = cv2.resize(next_obs, (128, 128), interpolation=cv2.INTER_CUBIC)
next_obs = cv2.normalize(next_obs, None, alpha=-1, beta=1, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)
# infer the states
next_state, next_c = encode_obs(hps, model, next_obs, action, reward, c, k)
next_state = np.reshape(next_state, [1, hps.z_size])
state = next_state
c = next_c
score += 1
if done:
break
return score
def testing(game, src_domain_index, hps, model, reduction_set_s, reduction_set_theta, k):
# gpu setup
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
gpu_config = tf.ConfigProto(device_count={'GPU': 0},
allow_soft_placement=False,
log_device_placement=False)
gpu_config.gpu_options.allow_growth = True
with tf.Session(config=gpu_config) as sess:
scores = []
env = env_init(game, src_domain_index)
tf.set_random_seed(int(100))
theta_all = extract_theta(hps, model)
for i in range(30):
# set the random seed
random.seed(int(i * 100))
np.random.seed(int(i * 100))
for j in range(n_domain):
env[j].seed(int(i * 100))
state_size = hps.z_size
action_size = env[0].action_space.n
theta_size = hps.theta_s_size + hps.theta_r_size
if reduction_set_theta == None:
agent = DoubleDQNAgent(len(reduction_set_s[0]), action_size, theta_size,
is_load=True, load_p=policy_save_path)
else:
agent = DoubleDQNAgent(len(reduction_set_s[0]), action_size,
len(reduction_set_theta) + hps.theta_r_size,
is_load=True, load_p=policy_save_path)
score = predict(env, agent, model, theta_all, k)
scores.append(score)
print(scores)
parser = argparse.ArgumentParser()
parser.add_argument('-name', type=str, required=True, help='data path')
parser.add_argument('-source', type=str, required=True, help='data path')
parser.add_argument('-dest', type=str, required=True, help='data path')
parser.add_argument('-domain', type=str, nargs='+', required=True, help='full domain index')
parser.add_argument('-mvae_p', type=str, required=True, help='data path')
parser.add_argument('-k', type=int, required=False, help='test domain index')
parser.add_argument('-step', type=int, required=True, help='test domain index')
args = parser.parse_args()
game = args.name
source_p = args.source
dest_p = args.dest
if not os.path.exists(dest_p):
os.makedirs(dest_p)
k = args.k
step = args.step
src_domain_index = args.domain
n_domain = len(src_domain_index)
mvae_p = args.mvae_p
date_format = '%A_%d_%B_%Y_%Hh_%Mm_%Ss'
time_now = datetime.now().strftime(date_format)
data_indicator = args.source.split('/', 2)[-1]
model_save_path = './results/' + data_indicator + '/test/' + time_now
policy_save_path = './results/' + data_indicator + '/policy_opt/' + time_now + '/policy.h5'
if not os.path.exists(model_save_path):
os.makedirs(model_save_path)
n_epoch = 1000
hps = default_hps(game, 'test2', n_domain)
mvae = missVAE(hps, step=3, scope='all')
mvae.load_json(os.path.join(mvae_p), is_tesing=True)
dh = DataHandler(hps, 3, source_p, dest_p, src_domain_index, 1000 * n_domain, is_test=True)
(theta_s, theta_r,
SSL_A, SSL_B, SSL_C,
SSL_D, SSL_E, SSL_F) = mvae.sess.run([mvae.theta_s, mvae.theta_r,
tf.math.erf(mvae.SSL_A), tf.math.erf(mvae.SSL_B), tf.math.erf(mvae.SSL_C),
tf.math.erf(mvae.SSL_D), tf.math.erf(mvae.SSL_E), tf.math.erf(mvae.SSL_F)])
# minimal & sufficient set identification
reduction_set_s = min_n_suff_set_state(SSL_B, SSL_D)
reduction_set_theta = min_n_suff_set_theta(reduction_set_s, SSL_F)
# Step 1: inference theta
if step == 1:
infer_theta(n_epoch, n_domain, hps, mvae, dh, model_save_path)
elif step == 2:
# Step 2: testing
mvae_pretrain_p = os.path.join(model_save_path, 'test.json')
mvae.load_json(mvae_pretrain_p, is_dyn=False)
SCREEN_X = 128
SCREEN_Y = 128
# vdisplay = Xvfb(width=SCREEN_X, height=SCREEN_Y)
# vdisplay.start()
testing(game, src_domain_index, hps, mvae, reduction_set_s, reduction_set_theta, k=0)
else:
Exception('Wrong step number !')