-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathevaluation.py
257 lines (205 loc) · 6.95 KB
/
evaluation.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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
import numpy as np
import torch
from a2c_ppo_acktr import utils
from a2c_ppo_acktr.envs import make_vec_envs
def evaluate(
actor_critic,
obs_rms,
env_name,
seed,
num_processes,
eval_log_dir,
device,
antagonist=None,
):
eval_envs = make_vec_envs(
env_name, seed + num_processes, num_processes, None, eval_log_dir, device, True
)
vec_norm = utils.get_vec_normalize(eval_envs)
if vec_norm is not None:
vec_norm.eval()
vec_norm.obs_rms = obs_rms
eval_episode_rewards = []
obs = eval_envs.reset()
eval_recurrent_hidden_states = torch.zeros(
num_processes, actor_critic.recurrent_hidden_state_size, device=device
)
eval_masks = torch.zeros(num_processes, 1, device=device)
while len(eval_episode_rewards) < 200:
with torch.no_grad():
_, action, _, eval_recurrent_hidden_states = actor_critic.act(
obs, eval_recurrent_hidden_states, eval_masks, deterministic=True
)
print(eval_episode_rewards)
# Obser reward and next obs
obs, _, done, infos = eval_envs.step(action)
eval_masks = torch.tensor(
[[0.0] if done_ else [1.0] for done_ in done],
dtype=torch.float32,
device=device,
)
for info in infos:
if "episode" in info.keys():
eval_episode_rewards.append(info["episode"]["r"])
eval_envs.close()
print(
" Evaluation using {} episodes: mean reward {:.5f}, stddev reward {:.5f}\n".format(
len(eval_episode_rewards), np.mean(eval_episode_rewards), np.std(eval_episode_rewards)
)
)
def evaluate_antagonist(
trained_agent,
antagonist,
num_attacks,
attack_threshold,
obs_rms,
env_name,
seed,
num_processes,
eval_log_dir,
device,
):
eval_envs = make_vec_envs(
env_name, seed + num_processes, num_processes, None, eval_log_dir, device, True
)
trained_agent.eval()
antagonist.eval()
vec_norm = utils.get_vec_normalize(eval_envs)
if vec_norm is not None:
vec_norm.eval()
vec_norm.obs_rms = obs_rms
eval_episode_rewards = []
obs = eval_envs.reset()
eval_recurrent_hidden_states = torch.zeros(
num_processes, antagonist.recurrent_hidden_state_size, device=device
)
eval_masks = torch.zeros(num_processes, 1, device=device)
attacks_left = (
torch.tensor([num_attacks] * num_processes, dtype=torch.float)
.reshape(num_processes, 1)
.to(utils.get_device())
)
total_actions = 0
total_attacks = 0
while len(eval_episode_rewards) < 32:
with torch.no_grad():
(
_,
ant_action,
ant_action_log_prob,
eval_recurrent_hidden_states,
) = antagonist.act(
obs, eval_recurrent_hidden_states, eval_masks, deterministic=True
)
_, tr_action, _, _, = trained_agent.act(
obs, eval_recurrent_hidden_states, eval_masks, deterministic=True
)
ant_action_mask = (ant_action_log_prob > np.log(attack_threshold)) & (
attacks_left > 0
)
total_actions += len(ant_action_mask)
total_attacks += sum(ant_action_mask).item()
print(total_attacks, total_actions)
attacks_left -= (
torch.ones(num_processes, 1).to(utils.get_device()) * ant_action_mask
)
action = ant_action_mask * ant_action + ~ant_action_mask * tr_action
# action = ant_action
print(eval_episode_rewards)
# Obser reward and next obs
obs, reward, done, infos = eval_envs.step(action)
# reset num attacks
scored_mask = reward != 0
attacks_left[scored_mask] = num_attacks
eval_masks = torch.tensor(
[[0.0] if done_ else [1.0] for done_ in done],
dtype=torch.float32,
device=device,
)
for info in infos:
if "episode" in info.keys():
eval_episode_rewards.append(info["episode"]["r"])
eval_envs.close()
print(
" Evaluation using {} episodes: mean reward {:.5f}\n".format(
len(eval_episode_rewards), np.mean(eval_episode_rewards)
)
)
def evaluate_antagonist_defence(
trained_agent,
antagonist,
num_attacks,
attack_threshold,
obs_rms,
env_name,
seed,
num_processes,
eval_log_dir,
device,
):
eval_envs = make_vec_envs(
env_name, seed + num_processes, num_processes, None, eval_log_dir, device, True
)
trained_agent.eval()
antagonist.eval()
vec_norm = utils.get_vec_normalize(eval_envs)
if vec_norm is not None:
vec_norm.eval()
vec_norm.obs_rms = obs_rms
eval_episode_rewards = []
obs = eval_envs.reset()
eval_recurrent_hidden_states = torch.zeros(
num_processes, antagonist.recurrent_hidden_state_size, device=device
)
eval_masks = torch.zeros(num_processes, 1, device=device)
attacks_left = (
torch.tensor([num_attacks] * num_processes, dtype=torch.float)
.reshape(num_processes, 1)
.to(utils.get_device())
)
total_actions = 0
total_attacks = 0
while len(eval_episode_rewards) < 32:
with torch.no_grad():
(
_,
ant_action,
ant_action_log_prob,
eval_recurrent_hidden_states,
) = antagonist.act(
obs, eval_recurrent_hidden_states, eval_masks, deterministic=True
)
_, tr_action, _, _, = trained_agent.act(
obs, eval_recurrent_hidden_states, eval_masks, deterministic=True
)
ant_action_mask = (ant_action_log_prob > np.log(attack_threshold)) & (
attacks_left > 0
)
total_actions += len(ant_action_mask)
total_attacks += sum(ant_action_mask).item()
print(total_attacks, total_actions)
attacks_left -= (
torch.ones(num_processes, 1).to(utils.get_device()) * ant_action_mask
)
action = ant_action_mask * ant_action + ~ant_action_mask * tr_action
# action = ant_action
print(eval_episode_rewards)
# Obser reward and next obs
obs, reward, done, infos = eval_envs.step(action)
# reset num attacks
scored_mask = reward != 0
attacks_left[scored_mask] = num_attacks
eval_masks = torch.tensor(
[[0.0] if done_ else [1.0] for done_ in done],
dtype=torch.float32,
device=device,
)
for info in infos:
if "episode" in info.keys():
eval_episode_rewards.append(info["episode"]["r"])
eval_envs.close()
print(
" Evaluation using {} episodes: mean reward {:.5f}, stddev reward {:.5f}\n".format(
len(eval_episode_rewards), np.mean(eval_episode_rewards), np.std(eval_episode_rewards)
)
)