def single_run(size, score_to_win, AgentClass, model, **kwargs):
game = Game(size, score_to_win)
agent = AgentClass(game, model, display=Display(), **kwargs)
agent.play(verbose=True)
return game.score
if __name__ == '__main__':
GAME_SIZE = 4
SCORE_TO_WIN = 2048
N_TESTS = 10
'''====================
Use your own agent here.'''
PATH = './2048.pth'
#PATH = './2048_CPU.pth'
model = Net()
model.load_state_dict(torch.load(PATH))
#这行语句在测试的时候一定要加，不然参数可能发生变化
model.eval()
TestAgent = MyAgent
'''===================='''
scores = []
for _ in range(N_TESTS):
score = single_run(GAME_SIZE,
SCORE_TO_WIN,
AgentClass=TestAgent,
model=model.cpu())
#score = single_run(GAME_SIZE, SCORE_TO_WIN,
# AgentClass=TestAgent, model = model)
scores.append(score)
from torch.autograd import Variable
import torch.optim as optim
from tools_for_model import generateTrainSet
from my_model import Net, weights_init
import torch
import time
#用于保存生成的数据集
import pickle
start_time = time.time()
PATH = './2048.pth'
PATH_CPU = './2048_CPU.pth'
model = Net()
#加载已经训练的模型
model.load_state_dict(torch.load(PATH))
#参数初始化
#model.apply(weights_init)
if torch.cuda.is_available():
print("gpu is available")
model.cuda() #将所有的模型参数移动到GPU上
print("Load Time {} minutes".format(int((time.time() - start_time) / 60)))
criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=5e-5)
#最外层循环次数
EPOCH_1 = 20
from torchvision.models.inception import inception_v3
from torchvision import datasets, transforms
from PIL import Image
#from scipy.misc import imsave
import imageio
import matplotlib.pyplot as plt
import os
import numpy as np
reverse_trans = lambda x: np.asarray(T.ToPILImage()(x))
loss = nn.CrossEntropyLoss()
net = Net()
net.load_state_dict(torch.load('./fashion_mnist_cnn.pt'))
net.cuda()
net.eval()
eps = 2 * 8 / 225.
steps = 40
norm = float('inf')
step_alpha = 0.01
classes = {
0:"T-shirt/top",
1:"Trouser",
2:"Pullover",
3:"Dress",
4:"Coat",
import torch
from white_box_attack import *
import pickle
import os
from torch import nn
from torchvision import transforms
from my_model import Net
from torch.autograd import Variable
from torch.autograd.gradcheck import zero_gradients
import matplotlib.pyplot as plt
import os
import numpy as np
"""------------ load models and data ---------------------"""
device = 'cuda: 0'
with open("./attack_data/correct_1k.pkl", "rb") as f:
data = pickle.load(f)
black_model = torch.load('./Res18_accuracy_0.9264.pkl').to(device)
black_model.conv5_x = nn.AvgPool2d(7, 7, 0)
white_model = Net().to(device)
white_model.load_state_dict(torch.load('./fashion_mnist_cnn.pt'))
imgs = torch.tensor(data[0]).to(device)
labels = torch.tensor(data[1]).to(device).squeeze(dim=1)
labels = labels.max(1)[1]
"""-------------- white box attack ---------------------"""
agent = AgentClass(game=game, myNet=model, **kwargs)
trace = []
for board in board_json:
game.board = np.array(board)
direction = agent.step()
trace.append(direction)
fingerprint = "".join(str(i) for i in trace)
return fingerprint
from collections import Counter
'''====================
Use your own agent here.'''
PATH = './2048.pth'
#PATH = './2048_CPU.pth'
model = Net()
model.load_state_dict(torch.load(PATH))
#这行语句在测试的时候一定要加，不然参数可能发生变化
model.eval()
TestAgent = MyAgent
'''===================='''
fingerprint = generate_fingerprint(AgentClass=TestAgent, model=model.cpu())
#fingerprint = generate_fingerprint(AgentClass=TestAgent, model=model)
with open("EE369_fingerprint.json", 'w') as f:
pack = dict()
pack['fingerprint'] = fingerprint
pack['statstics'] = dict(Counter(fingerprint))
f.write(json.dumps(pack, indent=4))