-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_gridworld_sadqn.py
235 lines (205 loc) · 9.61 KB
/
test_gridworld_sadqn.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
from gridworld import *
import sys
sys.path.append("./common")
sys.path.append("./auto_LiRPA")
from auto_LiRPA import BoundedModule, BoundedTensor
from auto_LiRPA.perturbations import PerturbationLpNorm
from argparser import argparser
import numpy as np
from read_config import load_config
from attacks import *
from common.wrappers import make_atari, wrap_deepmind, wrap_pytorch, make_atari_cart
from models import QNetwork, model_setup
import torch.optim as optim
import torch
import torch.autograd as autograd
import time
import os
import argparse
import random
from datetime import datetime
from utils import Logger, get_acrobot_eps, test_plot
from async_env import AsyncEnv
from train import get_logits_lower_bound
UINTS=[np.uint8, np.uint16, np.uint32, np.uint64]
USE_CUDA = torch.cuda.is_available()
Variable = lambda *args, **kwargs: autograd.Variable(*args, **kwargs).cuda() if USE_CUDA else autograd.Variable(*args, **kwargs)
def get_ball(state):
result = []
result.append([state[0]-1, state[1]])
result.append([state[0], state[1]])
result.append([state[0]+1, state[1]])
result.append([state[0], state[1]-1])
result.append([state[0], state[1]+1])
true_result = []
for i in result:
# print((i[0]<0 or i[0]>self.environment.height-1))
# print((i[1]<0 or i[1]>self.environment.width-1))
if not ((i[0]<0 or i[0]>4) or (i[1]<0 or i[1]>4)):
true_result.append(i)
#print(true_result)
return true_result
def get_att_state(true_state, model):
states = get_ball(true_state)
values = dict()
true_state_tensor = torch.from_numpy(np.ascontiguousarray(true_state)).unsqueeze(0).cuda().to(torch.float32)
for i in states:
tensor_i = torch.from_numpy(np.ascontiguousarray(i)).unsqueeze(0).cuda().to(torch.float32)
action = model.act(tensor_i)[0]
#print(model.forward(true_state_tensor))
values[tuple(i)] = model.forward(true_state_tensor)[0][action].data.cpu()
att_state = min(values, key = values.get)
return att_state
def main(args):
config = load_config(args)
prefix = config['env_id']
training_config = config['training_config']
test_config = config['test_config']
attack_config = test_config["attack_config"]
if config['name_suffix']:
prefix += config['name_suffix']
if config['path_prefix']:
prefix = os.path.join(config['path_prefix'], prefix)
if 'load_model_path' in test_config and os.path.isfile(test_config['load_model_path']):
if not os.path.exists(prefix):
os.makedirs(prefix)
test_log = os.path.join(prefix, test_config['log_name'])
else:
if os.path.exists(prefix):
test_log = os.path.join(prefix, test_config['log_name'])
else:
raise ValueError('Path {} not exists, please specify test model path.')
logger = Logger(open(test_log, "w"))
logger.log('Command line:', " ".join(sys.argv[:]))
logger.log(args)
logger.log(config)
certify = test_config.get('certify', False)
env_params = training_config['env_params']
env_params['clip_rewards'] = False
env_params['episode_life'] = False
env_id = config['env_id']
# if "NoFrameskip" not in env_id:
# env = make_atari_cart(env_id)
# else:
# env = make_atari(env_id)
# env = wrap_deepmind(env, **env_params)
# env = wrap_pytorch(env)
env = GridWorld()
state = env.reset()
dtype = state.dtype
logger.log("env_shape: {}, num of actions: {}".format(env.observation_space.shape, env.action_space.n))
model_width = training_config['model_width']
robust_model = certify
dueling = training_config.get('dueling', True)
model = model_setup(env_id, env, robust_model, logger, USE_CUDA, dueling, model_width)
if 'load_model_path' in test_config and os.path.isfile(test_config['load_model_path']):
model_path = test_config['load_model_path']
else:
logger.log("choosing the best model from " + prefix)
all_idx = [int(f[6:-4]) for f in os.listdir(prefix) if os.path.isfile(os.path.join(prefix, f)) and os.path.splitext(f)[1]=='.pth' and 'best' not in f]
all_best_idx = [int(f[11:-4]) for f in os.listdir(prefix) if os.path.isfile(os.path.join(prefix, f)) and os.path.splitext(f)[1]=='.pth' and 'best' in f]
if all_best_idx:
model_frame_idx = max(all_best_idx)
model_name = 'best_frame_{}.pth'.format(model_frame_idx)
else:
model_frame_idx = max(all_idx)
model_name = 'frame_{}.pth'.format(model_frame_idx)
model_path = os.path.join(prefix, model_name)
logger.log('model loaded from ' + model_path)
model.features.load_state_dict(torch.load(model_path))
num_episodes = test_config['num_episodes']
max_frames_per_episode = test_config['max_frames_per_episode']
all_rewards = []
episode_reward = 0
seed = random.randint(0, 10000)
logger.log('reseting env with seed', seed)
env.seed(seed)
state = env.reset()
start_time = time.time()
if training_config['use_async_env']:
# Create an environment in a separate process, run asychronously
async_env = AsyncEnv(env_id, result_path=prefix, draw=training_config['show_game'], record=training_config['record_game'], save_frames=test_config['save_frames'], env_params=env_params, seed=args.seed)
episode_idx = 1
this_episode_frame = 1
# if certify:
# certified = 0
#
# if dtype in UINTS:
# state_max = 1.0
# state_min = 0.0
# else:
# state_max = float('inf')
# state_min = float('-inf')
for frame_idx in range(1, num_episodes * max_frames_per_episode + 1):
state_tensor = torch.from_numpy(np.ascontiguousarray(state)).unsqueeze(0).cuda().to(torch.float32)
# Normalize input pixel to 0-1
# if dtype in UINTS:
# state_tensor /= 255
att_state = get_att_state(state, model)
#state_tensor = attack(model, state_tensor, attack_config)
att_state_tensor = pgd(model, state_tensor, model.act(state_tensor), env_id = "Grid")
# print("ori",state_tensor)
# print("att",att_state_tensor)
#attack(model, state_tensor, attack_config)
att_state_tensor = torch.from_numpy(np.ascontiguousarray(att_state)).unsqueeze(0).cuda().to(torch.float32)
if certify:
beta = training_config.get('convex_final_beta',0)
eps = attack_config['params']['epsilon']
if env_id == 'Acrobot-v1':
eps_v = get_acrobot_eps(eps)
if USE_CUDA:
eps_v = eps_v.cuda()
else:
eps_v = eps
state_ub = torch.clamp(state_tensor + eps_v, max=state_max)
state_lb = torch.clamp(state_tensor - eps_v, min=state_min)
action = model.act(att_state_tensor)[0]
#print(action)
if certify:
max_logit = torch.tensor([action])
c = torch.eye(model.num_actions).type_as(state_tensor)[max_logit].unsqueeze(1) - torch.eye(model.num_actions).type_as(state_tensor).unsqueeze(0)
I = (~(max_logit.data.unsqueeze(1) == torch.arange(model.num_actions).type_as(max_logit.data).unsqueeze(0)))
c = (c[I].view(state_tensor.size(0), model.num_actions-1, model.num_actions))
logits_diff_lb = get_logits_lower_bound(model, state_tensor, state_ub, state_lb, eps_v, c, beta)
if torch.min(logits_diff_lb[0], 0)[0].data.cpu().numpy() > 0:
certified += 1
if training_config['use_async_env']:
async_env.async_step(action)
next_state, reward, done, _ = async_env.wait_step()
else:
next_state, reward, done, _ = env.step(action)
state = next_state
episode_reward += reward
if frame_idx % test_config['print_frame']==0:
logger.log('\ntotal frame {}/{}, episode {}/{}, episode frame{}/{}, latest episode reward: {:.6g}, avg 10 episode reward: {:.6g}'.format(frame_idx, num_episodes*max_frames_per_episode, episode_idx, num_episodes, this_episode_frame, max_frames_per_episode,
all_rewards[-1] if all_rewards else np.nan,
np.average(all_rewards[:-11:-1]) if all_rewards else np.nan))
if certify:
logger.log('certified action: {}, certified action ratio: {:.6g}'.format(certified, certified*1.0/frame_idx))
if this_episode_frame == max_frames_per_episode:
logger.log('maximum number of frames reached in this episode, reset environment!')
done = True
if training_config['use_async_env']:
async_env.epi_reward = 0
if done:
logger.log('reseting env with seed', seed)
if training_config['use_async_env']:
state = async_env.reset()
print("init state", state)
else:
state = env.reset()
all_rewards.append(episode_reward)
episode_reward = 0
this_episode_frame = 1
episode_idx += 1
if episode_idx > num_episodes:
break
else:
this_episode_frame += 1
logger.log('\navg reward' + (' and avg certify:' if certify else ':'))
logger.log(np.mean(all_rewards),'+-',np.std(all_rewards))
if certify:
logger.log(certified*1.0/frame_idx)
if __name__ == "__main__":
args= argparser()
main(args)