-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_atari_step_aba.py
293 lines (267 loc) · 11.9 KB
/
test_atari_step_aba.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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
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
from eps_scheduler import EpsilonScheduler
from read_config import load_config
import numpy as np
import cpprb
import re
from attacks import attack
import gym
import random
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
from torch.nn import CrossEntropyLoss
import torch.autograd as autograd
import math
import time
import os
import argparse
import copy
from datetime import datetime
from utils import CudaTensorManager, ActEpsilonScheduler, BufferBetaScheduler, Logger, update_target, get_acrobot_eps, plot
from my_replay_buffer import ReplayBuffer, NaivePrioritizedBuffer
from common.replay_buffer import PrioritizedReplayBuffer
from async_env import AsyncEnv
from async_rb import AsyncReplayBuffer
from gridworld import *
from ibp import *
from data_utils import *
from pfrnns.model import *
from pfrnns.evaluate import *
from train_atari import *
from autoencoder_atari import *
import cv2
import csv
from paad_rl.a2c_ppo_acktr import algo, utils
from paad_rl.a2c_ppo_acktr.algo import gail
from paad_rl.a2c_ppo_acktr.arguments import get_args
from paad_rl.a2c_ppo_acktr.envs import make_vec_envs
from paad_rl.a2c_ppo_acktr.model import Policy
from paad_rl.a2c_ppo_acktr.storage import RolloutStorage
# from evaluation import evaluate
from paad_rl.attacker.attacker import *
from paad_rl.utils.dqn_core import DQN_Agent, Q_Atari,model_get
from paad_rl.utils.param import Param
from paad_rl.a2c_ppo_acktr.algo.kfac import KFACOptimizer
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)
from denoising_diffusion_pytorch import Unet, GaussianDiffusion, Trainer
# def normalize_to_neg_one_to_one(img):
# return img * 2 - 1
#
# def unnormalize_to_zero_to_one(t):
# return (t + 1) * 0.5
#
# def sample_with_prior(diffusion, prior, steps = 10):
# img = torch.from_numpy(np.asarray(prior)).unsqueeze(0).cuda().to(torch.float32)
# tensor_steps = torch.linspace(1., 0., steps + 1).cuda().to(torch.float32)
# for i in range(steps):
# times = tensor_steps[i].cuda()
# times_next = tensor_steps[i+1].cuda()
# #print(*img.shape)
# img = diffusion.p_sample(img, times, times_next)
# img.clamp_(-1., 1.)
# img = unnormalize_to_zero_to_one(img)
# return img
def mini_test(env, model, diffusion, config, dtype, max_frames_per_episode=1000, seed = 1000, steps = 1):
begin_time = time.time()
device = torch.device('cuda')
state = env.reset()
all_rewards = []
episode_reward = 0
env.seed(seed)
np.random.seed(seed)
random.seed(seed)
torch.manual_seed(seed)
state = env.reset()
episode_idx = 1
this_episode_frame = 1
obs=[]
pos=[]
actions=[]
# #import belief model
# belief_args = parse_args()
# #belief_model = get_model(belief_args).cuda()
# belief_model = Belief_model(belief_args).cuda()
# model_checkpoint, _ = get_checkpoint(belief_args)
# # print(model_checkpoint)
# belief_model.load_state_dict(model_checkpoint)
# belief = []
action_space = Box(-1.0, 1.0, (env.action_space.n-1,))
pa_attacker = Policy(
env.observation_space.shape,
action_space,
beta=False,
epsilon=15/255,
base_kwargs={'recurrent': False})
if True:
KFACOptimizer(pa_attacker) # the model structure for the acktr attacker is different
old_steps, pa_attacker_state, _ = \
torch.load('sa-dqn_models/paad/dqn_obspol_attacker_PongNoFrameskip-v4', map_location=device)
pa_attacker.load_state_dict(pa_attacker_state)
pa_attacker.cuda()
attacker = ObsPol_Attack(pa_attacker, det=True, cont=False)
recurrent = torch.zeros(
1, pa_attacker.recurrent_hidden_state_size).cuda()
masks = torch.ones(1, 1).cuda()
#print("training steps for this model:", old_steps)
for frame_idx in range(1, max_frames_per_episode + 1):
state_tensor = torch.from_numpy(np.ascontiguousarray(state)).unsqueeze(0).cuda().to(torch.float32)/255
# att_state = get_att_state(model, state/255, max_diff_norm = 1/255, sample_time = 10)
# att_state_tensor = torch.from_numpy(np.ascontiguousarray(att_state)).unsqueeze(0).cuda().to(torch.float32)
# print("_______________att________________")
# print("dis", np.linalg.norm(att_stat))
# print("max",np.max(att_stat))
# print("mean", np.mean(att_stat))
# dif = restore_img[0].cpu().numpy() - restore_img_1[0].cpu().numpy()
# print("_______________restore____________")
# print("dis", np.linalg.norm(dif))
# print("max",np.max(dif))
# print("mean", np.mean(dif))
att_state_tensor = state_tensor
#att_state_tensor = pgd(model, state_tensor, model.act(state_tensor), env_id = "Pong")
att_state = att_state_tensor.cpu().numpy()[0]
# att_state_tensor = min_best(model, state_tensor, 15/255, pgd_steps=10, lr=1e-1, fgsm=False,
# norm=np.inf, rand_init=False, momentum=False, env_id = "Freeway")
# att_state_tensor, recurrent = attacker.attack_torch(model, torch.from_numpy(np.ascontiguousarray(state)).cuda().to(torch.float32)/255, recurrent, masks, epsilon=15/255,
# fgsm=True, lr=0.1, pgd_steps=10, device=device)
# #print(att_state_tensor.shape)
#att_state = att_state_tensor.cpu().numpy()
#diffusion
att_state_diff = diffusion.p_sample_loop_with_prior((att_state.transpose(0,2,1)), samples = 1, steps = steps).cpu().numpy().transpose(0,1,3,2)
#att_state = att_state_tensor.squeeze(0).cpu().numpy()
# att_state_tensor = torch.from_numpy(np.ascontiguousarray(att_state)).unsqueeze(0).cuda().to(torch.float32)/255
att_state_tensor = torch.from_numpy(np.ascontiguousarray(att_state_diff[0])).unsqueeze(0).cuda().to(torch.float32)
# #
#cv2.imwrite('./test_pic/'+str(frame_idx)+'att.png', (att_state*255).transpose(2,1,0))
# # # cv2.imwrite('./test_pic/'+str(frame_idx)+'ori.png', state.transpose(2,1,0))
#cv2.imwrite('./test_pic/'+str(frame_idx)+'restore.png', (att_state_diff[0]*255).transpose(2,1,0))
#restore_img = diffusion.p_sample_loop_with_prior((att_state.transpose(0,2 ,1)), steps = 250)
#restore_img_1 = diffusion.p_sample_loop_with_prior((state.transpose(0,2,1)/255), samples = 1, steps = 12).cpu().numpy().transpose(0,1,3,2)
#cv2.imwrite('./test_pic/'+str(frame_idx)+'restore_1.png', (restore_img_1[0]*255).transpose(2,1,0))
#att_stat = att_state - state/255
#print(np.linalg.norm(att_state - state/255))
#print(np.linalg.norm(att_state_diff[0] - restore_img_1[0]))
#print(att_state)
#print(state_tensor, att_state_tensor)
#belief = random_belief(att_state)
#print(model.forward(att_state_tensor))
#action = model.act(att_state_tensor, 0)[0]
#print(action)
action = get_action_withbelief(env, model, att_state_diff, 0)
#print(action)
# if len(belief) == 0:
# action = model.act(att_state_tensor, 0)[0]
# else:
# action = model.act(att_state_tensor, 0)[0]
#action = get_action_withbelief(env, model, belief, 0)
# obs.append(att_state)
# temp_state = copy.deepcopy(state)
# temp_state = list(temp_state)
# temp_state = np.asarray(temp_state)
# pos.append(temp_state)
# actions.append([action])
# obs_tensor = torch.Tensor(np.asarray(obs)).unsqueeze(0).cuda()
# pos_tensor = torch.Tensor(np.asarray(pos)).unsqueeze(0).cuda()
# actions_tensor = torch.Tensor(np.asarray(actions)).unsqueeze(0).cuda()
#belief = get_belief(belief_model, belief_args, env_map, obs_tensor, pos_tensor, actions_tensor)
#belief = []
#action = model.act(state_tensor)[0]
next_state, reward, done, _ = env.step(action)
# logger.log(action)
state = next_state
episode_reward += reward
if frame_idx%50 == 0:
print(frame_idx, episode_reward)
if this_episode_frame == max_frames_per_episode:
done = True
if done:
break
else:
this_episode_frame += 1
fps = frame_idx/(time.time()-begin_time)
return episode_reward, fps
def main(args):
model = Unet(
dim = 64,
dim_mults = (1, 2),
channels = 1
)
diffusion = GaussianDiffusion(
model,
image_size = 84,
# channels = 1,
timesteps = 1000, # number of steps
sampling_timesteps = 250, # number of sampling timesteps (using ddim for faster inference [see citation for ddim paper])
loss_type = 'l1' # L1 or L2
)
#torch.load(diffusion, "./results/model-150.pt")
diffusion.load_state_dict(torch.load("./results_pong/model-150.pt")['model'])
diffusion = diffusion.cuda()
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 not os.path.exists(prefix):
os.makedirs(prefix)
env_params = training_config['env_params']
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)
state = env.reset()
dtype = state.dtype
model_width = training_config['model_width']
certify = test_config.get('certify', False)
robust_model = certify
dueling = training_config.get('dueling', True)
model = model_setup(env_id, env, robust_model, None, USE_CUDA, dueling, model_width)
#model_path = "Grid_Continous/att_frame_890000.pth"
model_path = "PongNoFrameskip-v4_very_good/att_frame_370000.pth"
#model_path = "PongNoFrameskip-v4/frame_3200000.pth"
#model_path = "PongNoFrameskip-v4_good/att_frame_170000.pth"
#model_path = "sa-dqn_models/models/Pong-convex-3.pth"
#model_path = "FreewayNoFrameskip-v4/att_frame_350000.pth"
#model_path = "FreewayNoFrameskip-v4_good/att_frame_535000.pth"
#model_path = "PongNoFrameskip-v4/frame_3000000.pth"
#model_path = "sa-dqn_models/models/Freeway-natural.model"
#model_path = "FreewayNoFrameskip-v4/frame_2400000.pth"
model.features.load_state_dict(torch.load(model_path))
# seed = training_config['seed']
# env.seed(seed)
# np.random.seed(seed)
# random.seed(seed)
# torch.manual_seed(seed)
for steps in [1,5,10,15,20,25,30]:
seed = 5004
#seed = 1000
reward = []
fps = []
for i in range(5):
tmp, tmp1 = mini_test(env, model, diffusion, config, dtype, max_frames_per_episode=5000, seed = seed+i, steps = steps)
print(steps, seed+i, tmp)
reward.append(tmp)
fps.append(tmp1)
print("average fps is ", np.mean(fps))
file = open("pong-vanila-aba-steps-"+str(steps)+"fps.csv", 'w')
writer = csv.writer(file)
writer.writerow(fps)
file.close()
if __name__ == "__main__":
args= argparser()
main(args)