def load_test_data(run_data, train, backdoor):
return load_data(
run_data["options"],
train=train,
shuffle=False,
backdoor=backdoor
)
def __init__(self, options, classes, controller):
super().__init__(options, classes)
self.shadow_data = load_data(
options, [options.adversaries['from']],
backdoor=options.adversaries['type'].find('backdoor') >= 0)
self.shadow_data['dataloader'].dataset.targets[:] = \
options.adversaries['to']
controller.add_sybil(self)
self.attacking = False
def __init__(self, options, classes):
super().__init__(options, classes)
self.shadow_data = load_data(
options, [options.adversaries['from']],
backdoor=options.adversaries['type'].find('backdoor') >= 0)
self.shadow_data['dataloader'].dataset.targets[:] = \
options.adversaries['to']
if (tt := self.options.adversaries['toggle_times']):
self.toggle_time = cycle(tt)
from models.wrapper_T import *
import tensorflow as tf
tfk = tf.keras
# Hyper Parameters
n_classes = 3
batch_size = 64
n_epochs = 50
# dataset
print('[Dataset]')
with time_counter():
x, y, _ = load_data(
'data/dataset/',
classes=['normal', 'nude', 'swimwear'],
size=128,
cache_path='data/data_cache/dataset_size128_autopad.pkl',
auto_pad_val=True)
x_train, x_test, y_train, y_test = train_test_split(x, y, train_rate=0.5)
n_train = len(x_train)
_random_cutout = gen_random_cutout(42)
@tf.function
def augment(image, label):
image, label = random_flip_left_right(image, label)
# image, label = random_flip_up_down(image, label)
image, label = _random_cutout(image, label)
image, label = random_rotate_90(image, label)
return image, label
trials = [int(el.split('_')[1]) for el in dirs]
trials.insert(0, -1)
trial = np.max(trials) + 1
os.mkdir('results/{}/trial_{}'.format(hyperparam_name, trial))
with tf.Graph().as_default():
model = models.gan.GAN(**hyperparameters)
for exploit in exploits:
data = []
for i in range(5):
trX, trY = ds.load_data(
(
'./data/three-step/{}/subset_{}/train_set.csv'
).format(exploit, i)
)
model.train(trX, trY)
for i in range(5):
teX, teY = ds.load_data(
(
'./data/three-step/{}/subset_{}/test_set.csv'
).format(exploit, i)
)
d = model.test(teX, teY)
data.append(d)
raw_data.append(d)
# tensorflow messageの抑制
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
tfk = tf.keras
# tfk.backend.set_floatx('float32')
# Params
batch_size = 64
n_epochs = 10
lr = 0.001
# Dataset
print('[Dataset]')
with time_counter():
x, y, _ = load_data('data/dataset/', classes=['normal', 'nude', 'swimwear'], size=256, cache_path='data/data_cache/dataset_size256_autopad.pkl')
x_train, x_test, y_train, y_test = train_test_split(x, y, train_rate=0.5)
n_train = len(x_train)
train_ds = tf.data.Dataset.from_tensor_slices((x_train, y_train)).shuffle(n_train).batch(batch_size)
test_ds = tf.data.Dataset.from_tensor_slices((x_test, y_test)).batch(batch_size)
print('x_train: {}'.format(x_train.shape))
print('y_train: {}'.format(y_train.shape))
print('x_test: {}'.format(x_test.shape))
print('y_test: {}'.format(y_test.shape))
in_shape = x_train.shape[1:]
del x_train, y_train, x_test, y_test
# Model
# model = SimpleCNN(in_shape, n_out=3)
# model.load_weights('simplecnn/best_param')
import models.wrapper_T
tfk = tf.keras
tfk.backend.set_floatx('float32')
# Params
n_classes = 2
batch_size = 64
n_epochs = 100
# lr = 0.001
# Dataset
print('[Dataset]')
with time_counter():
x, y, _ = load_data('data/sample/',
classes=['normal', 'nude', 'swimwear'],
size=128,
auto_pad_val=True)
ds = tf.data.Dataset.from_tensor_slices((x, y)).batch(batch_size)
print(_)
print('x: {}'.format(x.shape))
print('y: {}'.format(y.shape))
in_shape = x.shape[1:]
del x, y
model = models.wrapper.DenseNet121(weights=None,
classes=n_classes,
input_shape=in_shape)
model_name = 'densenet121'
def __init__(self, options, classes):
self.data = load_data(options, train=True, classes=classes)
params = options.model_params
self.net = load_model(params)
self.options = options
'orzhttpd_rootdir', 'orzhttpd_restore'
]
summaries = {'hyperparameters': hyperparameters}
raw_data = []
net_data = {exploit: [] for exploit in exploits}
model = models.gan.GAN(**hyperparameters)
for exploit in exploits:
data = []
for i in range(5):
trX, trY = ds.load_data(
(
'./data/ndss/ts/{}/subset_{}/train_set.csv'
).format(exploit, i)
)
model.train(trX, trY)
for j in range(5):
teX, teY = ds.load_data(
(
'./data/ndss/ts/{}/subset_{}/test_set.csv'
).format(exploit, j)
)
d = model.test(teX, teY)
data.append(d)
raw_data.append(d)
from keras.layers.convolutional import Conv2D
from keras.layers.pooling import MaxPooling2D, GlobalMaxPooling2D
from keras.models import Model
from keras.layers import Input
from keras.layers.core import Dropout, Flatten, Dense
from keras.optimizers import Adam
from utils import datasets
img_size = 128
nb_epochs = 15
#############################################################
# Load data
#############################################################
X_train, Y_train_masks, Y_train_classes, X_val, Y_val_masks, Y_val_classes, Y_train_bbox, Y_val_bbox = datasets.load_data(
)
print(X_train.shape)
print(Y_train_classes.shape)
print(X_val.shape)
print(Y_val_classes.shape)
#############################################################
# Data generators
#############################################################
train_datagen = ImageDataGenerator(rescale=1 / 255,
zoom_range=0.25,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
