class RandomMutator:
class Mutator:
def __init__(self, name, total=0):
self.name = name
self.total = total
self.delta_bigger_than_zero = 0
def __init__(self, mutate_ops=None):
self.mylogger = Logger()
self.mylogger.info(f"Using {self.__class__.__name__} as selection strategy!")
self._mutators = [self.Mutator(name=op) for op in mutate_ops]
@property
def mutators(self):
mus = {}
for mu in self._mutators:
mus[mu.name] = mu
return mus
def choose_mutator(self, mu1=None):
return self._mutators[np.random.randint(0, len(self._mutators))].name
def index(self, mutator_name):
for i, mu in enumerate(self._mutators):
if mu.name == mutator_name:
return i
return -1
def __init__(self, mutate_ops=None):
self.mylogger = Logger()
self.mylogger.info(f"Using {self.__class__.__name__} as selection strategy!")
if mutate_ops is None:
from scripts.mutation.model_mutation_generators import all_mutate_ops
mutate_ops = all_mutate_ops()
self.p = 1 / len(mutate_ops)
self._mutators = [self.Mutator(name=op) for op in mutate_ops]
class MCMC:
class Mutator:
def __init__(self, name, total=0, delta_bigger_than_zero=0, epsilon=1e-7):
self.name = name
self.total = total
self.delta_bigger_than_zero = delta_bigger_than_zero
self.epsilon = epsilon
@property
def score(self, epsilon=1e-7):
mylogger = Logger()
rate = self.delta_bigger_than_zero / (self.total + epsilon)
mylogger.info("Name:{}, rate:{}".format(self.name, rate))
return rate
def __init__(self, mutate_ops=None):
self.mylogger = Logger()
self.mylogger.info(f"Using {self.__class__.__name__} as selection strategy!")
if mutate_ops is None:
from scripts.mutation.model_mutation_generators import all_mutate_ops
mutate_ops = all_mutate_ops()
self.p = 1 / len(mutate_ops)
self._mutators = [self.Mutator(name=op) for op in mutate_ops]
@property
def mutators(self):
mus = {}
for mu in self._mutators:
mus[mu.name] = mu
return mus
def choose_mutator(self, mu1=None):
if mu1 is None:
# which means it's the first mutation
return self._mutators[np.random.randint(0, len(self._mutators))].name
else:
self.sort_mutators()
k1 = self.index(mu1)
k2 = -1
prob = 0
while np.random.rand() >= prob:
k2 = np.random.randint(0, len(self._mutators))
prob = (1 - self.p) ** (k2 - k1)
mu2 = self._mutators[k2]
return mu2.name
def sort_mutators(self):
import random
random.shuffle(self._mutators)
self._mutators.sort(key=lambda mutator: mutator.score, reverse=True)
def index(self, mutator_name):
for i, mu in enumerate(self._mutators):
if mu.name == mutator_name:
return i
return -1
# -*-coding:UTF-8-*-
"""get prediction for each backend
"""
import sys
import os
import redis
import pickle
import argparse
import configparser
from scripts.tools.utils import DataUtils
from scripts.logger.lemon_logger import Logger
import warnings
main_logger = Logger()
def custom_objects():
def no_activation(x):
return x
def leakyrelu(x):
import keras.backend as K
return K.relu(x, alpha=0.01)
objects = {}
objects['no_activation'] = no_activation
objects['leakyrelu'] = leakyrelu
return objects
help="list of backends")
parse.add_argument(
"--mutate_num",
type=int,
help="number of variant models generated by each mutation operator")
parse.add_argument("--mutate_ratio", type=float, help="ratio of mutation")
parse.add_argument("--exp", type=str, help="experiments identifiers")
parse.add_argument("--test_size", type=int, help="amount of testing image")
parse.add_argument("--config_name", type=str, help="config name")
flags, unparsed = parse.parse_known_args(sys.argv[1:])
warnings.filterwarnings("ignore")
lemon_cfg = configparser.ConfigParser()
lemon_cfg.read(f"./config/{flags.config_name}")
mutate_logger = Logger()
pool = redis.ConnectionPool(host=lemon_cfg['redis']['host'],
port=lemon_cfg['redis']['port'],
db=lemon_cfg['redis'].getint('redis_db'))
redis_conn = redis.Redis(connection_pool=pool)
for k in redis_conn.keys():
if flags.exp in k.decode("utf-8"):
redis_conn.delete(k)
experiment_dir = os.path.join(flags.output_dir,
flags.exp) # exp : like lenet5-mnist
mut_dir = os.path.join(experiment_dir, "mut_model")
crash_dir = os.path.join(experiment_dir, "crash")
nan_dir = os.path.join(experiment_dir, "nan")
inner_output_dir = os.path.join(experiment_dir, "inner_output")
import math
from typing import *
from scripts.mutation.mutation_utils import *
from scripts.mutation.layer_matching import LayerMatching
import random
import os
import warnings
from scripts.logger.lemon_logger import Logger
import datetime
warnings.filterwarnings("ignore")
os.environ["TF_CPP_MIN_LOG_LEVEL"] = '2' # 只显示 warning 和 Error
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = ""
mylogger = Logger()
def _assert_indices(mutated_layer_indices: List[int], depth_layer: int):
assert max(mutated_layer_indices
) < depth_layer, "Max index should be less than layer depth"
assert min(mutated_layer_indices
) >= 0, "Min index should be greater than or equal to zero"
def _shuffle_conv2d(weights, mutate_ratio):
new_weights = []
for val in weights:
# val is bias if len(val.shape) == 1
if len(val.shape) > 1:
pickle.dump(layers_output, fw)
if __name__ == "__main__":
"""Parser of command args"""
parse = argparse.ArgumentParser()
parse.add_argument("--backend", type=str, help="name of backends")
parse.add_argument("--exp", type=str, help="experiments identifiers")
parse.add_argument("--output_dir",
type=str,
help="relative path of output dir(from root dir)")
parse.add_argument("--data_index", type=int, help="redis db port")
parse.add_argument("--config_name", type=str, help="config name")
parse.add_argument("--model_idntfr", type=str, help="redis db port")
flags, unparsed = parse.parse_known_args(sys.argv[1:])
mylogger = Logger()
"""Load Configuration"""
warnings.filterwarnings("ignore")
lemon_cfg = configparser.ConfigParser()
lemon_cfg.read(f"./config/{flags.config_name}")
parameters = lemon_cfg['parameters']
gpu_ids = parameters['gpu_ids']
gpu_list = parameters['gpu_ids'].split(",")
"""Init cuda"""
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_ids
warnings.filterwarnings("ignore")
batch_size = 64
"""Switch backend"""
bk_list = ['tensorflow', 'theano', 'cntk', 'mxnet']
def score(self, epsilon=1e-7):
mylogger = Logger()
rate = self.delta_bigger_than_zero / (self.total + epsilon)
mylogger.info("Name:{}, rate:{}".format(self.name, rate))
return rate
mutate_ops=parameters['mutate_ops'],
exps=parameters['exps'].lstrip().rstrip().split(" "),
origin_model_dir=parameters['origin_model_dir'],
output_dir=parameters['output_dir'],
backend=parameters['backend'],
mutate_num=parameters.getint('mutate_num'),
mutate_ratio=parameters.getfloat('mutate_ratio'),
test_size=parameters.getint('test_size'),
threshold=parameters.getfloat('threshold'),
redis_db=lemon_cfg['parameters'].getint('redis_db'),
python_prefix=parameters['python_prefix'].rstrip("/"))
if not os.path.exists(flags.output_dir):
os.makedirs(flags.output_dir)
main_log = Logger()
"""Lemon process"""
main_log.logger.info("Success: Lemon start successfully!")
start_time = datetime.datetime.now()
for exp_identifier in flags.exps:
"""Make directory"""
experiment_dir = os.path.join(
flags.output_dir, exp_identifier) # exp : like lenet5-mnist
mut_dir = os.path.join(experiment_dir, "mut_model")
crash_dir = os.path.join(experiment_dir, "crash")
nan_dir = os.path.join(experiment_dir, "nan")
inner_output_dir = os.path.join(experiment_dir, "inner_output")
metrics_result_dir = os.path.join(experiment_dir, "metrics_result")
if not os.path.exists(experiment_dir):
os.makedirs(experiment_dir)
def __init__(self, mutate_ops=None):
self.mylogger = Logger()
self.mylogger.info(f"Using {self.__class__.__name__} as selection strategy!")
self._mutators = [self.Mutator(name=op) for op in mutate_ops]