def setup_all(args):
experiment, exp_id = setup_exp(args)
print(exp_id)
env = setup_env(args)
del env
experiment.log_parameters(args.__dict__)
data = setup_tr_val_test(args)
model = setup_model(args)
model_dir = setup_dir(basename=".models", args=args, exp_id=exp_id)
print("model save_dir: %s" % (str(model_dir)))
ims_dir = setup_dir(basename=".images", args=args, exp_id=exp_id)
return data, model, experiment, model_dir, ims_dir
from psychopy import visual, gui
import pathlib
import utils
import sys
#get experiment and monitor infos
expInfo, monInfo = utils.setup_exp()
#setup monitor
mon = utils.setup_mon(monInfo)
#get stimulation parameter from file
stim_params_path = './params_stim.txt'
params_stim = utils.get_stimulation_params(stim_params_path)
#start experiment
section = int(expInfo[3])
while section <= 10:
# setup window
win = visual.Window(size=mon.getSizePix(),
fullscr=True,
color='grey',
monitor=mon,
screen=0)
win.recordFrameIntervals = True
win.refreshThreshold = 1 / 60 + 0.004
# create folder to store data
data_path = "./data/%s/%s/%s/section_%s" % (expInfo[0], expInfo[1],
expInfo[2], section)
# json file with experiment config
CONFIG_FILE = './config_attack/bpda_eot_attack.json'
###############
# ## SETUP ## #
###############
# load experiment config
with open(CONFIG_FILE) as file:
config = json.load(file)
# directory for experiment results
exp_dir = config['exp_dir'] + '_' + datetime.datetime.now().strftime('%d-%m-%Y_%I-%M-%S_%p') + '_/'
# setup folders, save code, set seed and get device
setup_exp(exp_dir, config['seed'], ['log'], ['bpda_eot_attack.py', 'nets.py', 'utils.py', CONFIG_FILE])
print('Loading data and nets.')
# data loader
data, num_classes = import_data(config['data_type'], False, False)
attack_loader = DataLoader(data, batch_size=config['batch_size'], shuffle=config['subset_shuffle'], num_workers=0)
# get clf and ebm networks and load saved weights
clf = WideResNet(num_classes=num_classes).cuda()
clf.load_state_dict(t.load(config['clf_weight_path'], map_location=lambda storage, loc: storage.cuda()))
clf.eval()
if config['langevin_steps'] > 0:
ebm = EBM().cuda()
ebm.load_state_dict(t.load(config['ebm_weight_path'], map_location=lambda storage, loc: storage.cuda()))
ebm.eval()
# cross-entropy loss function to generate attack gradients
import copy
import torch
from functools import partial
from torch import nn
from torch.optim import Adam, RMSprop
import numpy as np
from pathlib import Path
import time
from data.setup import setup_data
import os
from utils import setup_args, setup_dir, setup_exp
if __name__ == "__main__":
args = setup_args()
data = setup_data(args)
experiment = setup_exp(args)
model = setup_model(args)
if args.task == "embed":
from training.embed_trainer import EmbedTrainer
trainer = EmbedTrainer(model, args, experiment)
elif args.task == "infer":
from training.inference_trainer import InferenceTrainer
trainer = InferenceTrainer(model, args, experiment)
elif args.task == "predict":
from training.prediction_trainer import PredictionTrainer
trainer = PredictionTrainer(model, args, experiment)
elif args.task == "control":
visited[origin_express] = origin_express
# change route_args current_position->origin_express embedding
embeddinged_route_args = env.get_embeddinged_route_args(
origin_express)
step_mix_s = np.append(embeddinged_route_args, now_traffic)
return path
# 如果选择的点有环路，则环境会返回来一个reward，reward的值，应该很小，表示不想出现环路
# 2：如果即将加入的originExpress的所有邻居，已经都在path中，则需要在path中删除后三个，重新设置 originExpress
# env setup #
setup_exp()
folder = setup_run()
env = ONOSEnv(folder)
# training #
s_dim = len(env.initial_route_args) - 1 + REPESENTATTION_SIZE + len(
env.env_loads.flatten())
a_dim = REPESENTATTION_SIZE
a_bound = 1
MAX_PATH_STEPS = env.active_nodes
MAX_EP_STEPS = 1000
ddpg = DDPG(a_dim, s_dim, a_bound)
# ddpg.train(routeTuple) # 路径元祖，是一个list：【（vector1，vector2）（vector1，vector2）】，vector是Embedding之后的表示
t1 = time.time()
for i in range(MAX_EPISODES):
# json file with experiment config
CONFIG_FILE = './config_train_clf/cifar10_nat.json'
###############
# ## SETUP ## #
###############
# load experiment config
with open(CONFIG_FILE) as file:
config = json.load(file)
# directory for experiment results
exp_dir = config['exp_dir'] + '_' + datetime.datetime.now().strftime('%d-%m-%Y_%I-%M-%S_%p') + '_/'
# setup folders, save code, set seed and get device
setup_exp(exp_dir, config['seed'], ['checkpoints'], ['train_clf.py', 'nets.py', 'utils.py', CONFIG_FILE])
print('Processing data...')
# import train and test datasets and set up data loaders
train_data, num_classes = import_data(config['data_type'], True, True)
train_data_loader = DataLoader(train_data, config['batch_size'], shuffle=True, num_workers=config['num_workers'])
test_data = import_data(config['data_type'], False, False)[0]
test_data_loader = DataLoader(test_data, config['batch_size'], shuffle=False, num_workers=config['num_workers'])
print('Setting up network and optimizer...')
# network structure and weight init
clf = WideResNet(num_classes=num_classes).cuda()
clf.apply(conv_init)
# initialize optim
assert len(config['lr_list']) == len(config['lr_schedule']), 'lr_list and lr_schedule must have the same length'
optim = t.optim.SGD(clf.parameters(), config['lr_list'][0], config['momentum'], weight_decay=config['weight_decay'])
# json file with experiment config
CONFIG_FILE = './config_train_ebm/cifar10_ebm.json'
#######################
# ## INITIAL SETUP ## #
#######################
# load experiment config
with open(CONFIG_FILE) as file:
config = json.load(file)
# directory for experiment results
exp_dir = config['exp_dir'] + '_' + datetime.datetime.now().strftime(
'%d-%m-%Y_%I-%M-%S_%p') + '_/'
# setup folders, save code, set seed and get device
setup_exp(exp_dir, config['seed'],
['checkpoints', 'shortrun', 'longrun', 'plots'],
['train_ebm.py', 'nets.py', 'utils.py', CONFIG_FILE])
print('Setting up network and optimizer...')
# set up network
ebm = EBM().cuda()
# set up adam optimizer for first learning phase
optim = t.optim.Adam(ebm.parameters(), lr=config['lr_adam'])
print('Processing data...')
# set up training data
data = import_data(config['data_type'], True, False)[0]
q = next(iter(DataLoader(data, len(data), num_workers=0)))[0].cuda()
if config['shortrun_init'] == 'persistent':
# initialize persistent images from noise (one persistent image for each data image)