sys.path.append(MODULE_PATH)
import tensorflow as tf
import iat.nets
ResNetDenoiseModel = iat.nets.ResNetDenoiseModel
from tensorpack import TowerContext
from tensorpack.tfutils import get_model_loader
from ares import ClassifierWithLogits
from ares.utils import get_res_path, download_res
import argparse
MODEL_PATH = get_res_path('./imagenet/R152-Denoise.npz')
def load(session):
model = ResNet152_Denoising()
model.load(session, MODEL_PATH)
return model
def download(model_path):
url = 'https://github.com/facebookresearch/ImageNet-Adversarial-Training/releases/download/v0.1/R152-Denoise.npz'
if not os.path.exists(model_path):
if not os.path.exists(os.path.dirname(model_path)):
os.makedirs(os.path.dirname(model_path), exist_ok=True)
download_res(url, model_path)
''' This file provides a wrapper class for AT_HE (https://github.com/ShawnXYang/AT_HE) model for CIFAR-10 dataset. '''
import sys
import os
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path, download_res
MODEL_PATH = get_res_path('./cifar10/model-wideres-pgdHE-wide10.pt')
def load(_):
model = AT_HE()
model.load()
return model
def download(model_path):
if not os.path.exists(model_path):
if not os.path.exists(os.path.dirname(model_path)):
os.makedirs(os.path.dirname(model_path), exist_ok=True)
url = 'http://ml.cs.tsinghua.edu.cn/~xiaoyang/downloads/weights/model-wideres-pgdHE-wide10.pt'
download_res(url, model_path)
''' This file provides a wrapper class for AWP (https://github.com/csdongxian/AWP) model for CIFAR-10 dataset. '''
import sys
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/RST-AWP_cifar10_linf_wrn28-10.pt')
def load(_):
model = AWP()
model.load()
return model
def filter_state_dict(state_dict):
from collections import OrderedDict
if 'state_dict' in state_dict.keys():
state_dict = state_dict['state_dict']
new_state_dict = OrderedDict()
for k, v in state_dict.items():
if 'sub_block' in k:
continue
import sys
import os
THIRD_PARTY_PATH = os.path.abspath(os.path.join(os.path.dirname(os.path.abspath(__file__)), '../../third_party'))
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
from bayesian_defense.models.vgg_rse import VGG
import torch
import tensorflow as tf
from ares.utils import get_res_path, download_res
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
MODEL_PATH = get_res_path('./cifar10/rse')
def load(_):
model = VGG_RSE()
model.load(MODEL_PATH)
return model
def download(model_path):
pth_path = os.path.join(model_path, 'cifar10_vgg_rse.pth')
if not os.path.exists(os.path.dirname(pth_path)):
os.makedirs(os.path.dirname(pth_path), exist_ok=True)
if not os.path.exists(pth_path):
print('Please download "cifar10_vgg_rse.pth" from ' +
'"https://drive.google.com/drive/folders/1uAP6q-fSADhlkx-YNBipPzv4KCvC_3_J?usp=sharing", ' +
''' This file provides a wrapper class for Hydra (https://github.com/inspire-group/hydra) model for CIFAR-10 dataset. '''
import sys
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
import math
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/model_best_dense.pth.tar')
def load(_):
model = Hydra()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10,
x_min=0.0,
x_max=1.0,
x_shape=(32, 32, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class Hydra(torch.nn.Module):
def __init__(self):
sys.path.append(THIRD_PARTY_PATH)
MODULE_PATH = os.path.join(THIRD_PARTY_PATH, 'models/research/slim')
if MODULE_PATH not in sys.path:
sys.path.append(MODULE_PATH)
import tensorflow as tf
import numpy as np
import models.research.slim.nets.resnet_v2 as resnet_v2
from ares import ClassifierWithLogits
from ares.utils import get_res_path, download_res
slim = tf.contrib.slim
MODEL_PATH = get_res_path('./imagenet/imagenet64_alp025_2018_06_26.ckpt')
def load(session):
model = ResnetV2ALP()
model.load(session, MODEL_PATH)
return model
def download(model_path):
url = 'http://download.tensorflow.org/models/adversarial_logit_pairing/imagenet64_alp025_2018_06_26.ckpt.tar.gz'
if not os.path.exists(model_path + '.meta'):
if not os.path.exists(os.path.dirname(model_path)):
os.makedirs(os.path.dirname(model_path), exist_ok=True)
import tarfile
download_res(url, model_path + '.tar.gz')
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
MODULE_PATH = os.path.join(THIRD_PARTY_PATH, 'models/research/slim')
if MODULE_PATH not in sys.path:
sys.path.append(MODULE_PATH)
import tensorflow as tf
import models.research.slim.nets.inception_v3 as inception_v3
from ares import ClassifierWithLogits
from ares.utils import get_res_path, download_res
slim = tf.contrib.slim
MODEL_PATH = get_res_path('./imagenet/inception_v3.ckpt')
def load(session):
model = InceptionV3()
model.load(session, MODEL_PATH)
return model
def download(model_path):
if not os.path.exists(model_path):
if not os.path.exists(os.path.dirname(model_path)):
os.makedirs(os.path.dirname(model_path), exist_ok=True)
gz_path = os.path.join(os.path.dirname(model_path),
'inception_v3_2016_08_28.tar.gz')
if not os.path.exists(gz_path):
(https://github.com/tensorflow/models/blob/master/research/adv_imagenet_models/README.md) for ImageNet dataset.
'''
import sys
import os
import tensorflow as tf
from inception_v3 import inception_v3
from ares import ClassifierWithLogits
from ares.utils import get_res_path, download_res
slim = tf.contrib.slim
MODEL_PATH = get_res_path('./imagenet/ens4_adv_inception_v3')
def load(session):
model = Ens4AdvInceptionV3()
model.load(session, MODEL_PATH)
return model
def download(model_path):
if not os.path.exists(model_path):
import tarfile
os.makedirs(model_path, exist_ok=True)
download_res(
'http://download.tensorflow.org/models/ens4_adv_inception_v3_2017_08_18.tar.gz',
''' ImageNet dataset (ILSVRC 2012). '''
import os
import tensorflow as tf
import numpy as np
from PIL import Image
from ares.utils import get_res_path, download_res
PATH_IMGS = get_res_path('./imagenet/ILSVRC2012_img_val')
PATH_VAL_TXT = get_res_path('./imagenet/val.txt')
PATH_TARGET_TXT = get_res_path('./imagenet/target.txt')
def load_dataset_for_classifier(classifier,
offset=0,
load_target=False,
target_label=None,
clip=True):
''' Get an ImageNet dataset in tf.data.Dataset format.
The first element of the dataset is the filename, the second one is the image tensor with shape of the classifier's
``x_shape`` in the classifier's ``x_dtype``, the third one is the label in the classifier's ``y_dtype``. If
``load_target`` is true, the target label would be returned as the fourth element of the dataset. It would
automatically handle ``n_class == 1000`` and ``n_class == 1001`` case (assume the empty class is labeled 0).
:param offset: Ignore the first ``offset`` images.
:param load_target: Whether to load the target label.
:param target_label: If it is a integer, the returned dataset would only include data points with this label.
:param clip: If it is true, the images would be clipped towards center.
:return: A ``tf.data.Dataset`` instance.
''' This file provides a wrapper class for MMC (https://github.com/P2333/Max-Mahalanobis-Training) model for CIFAR-10 dataset. '''
import os
import numpy as np
import tensorflow as tf
import keras
from keras.layers import Dense, Conv2D, BatchNormalization, Activation, GlobalAveragePooling2D
from keras.layers import Input, Lambda
from keras.regularizers import l2
from keras.models import Model
from scipy.io import loadmat
from ares.model import ClassifierWithLogits
from ares.utils import get_res_path, download_res
MODEL_PATH = get_res_path('./cifar10/mmc')
def load(session):
model = MMC()
model.load(session, MODEL_PATH)
return model
def download(model_path):
if not os.path.exists(model_path):
os.makedirs(model_path, exist_ok=True)
h5_name = 'model.180.h5'
h5_path = os.path.join(model_path, h5_name)
if not os.path.exists(h5_path):
url = 'http://ml.cs.tsinghua.edu.cn/~tianyu/MMC/pretrained_models/MMC_mom_advtrain_cifar10/' + h5_name
''' This file provides a wrapper class for Fast_AT (https://github.com/locuslab/fast_adversarial) model for CIFAR-10 dataset. '''
import sys
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/cifar_model_weights_30_epochs.pth')
def load(_):
model = Fast_AT()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10, x_min=0.0, x_max=1.0,
x_shape=(32, 32, 3), x_dtype=tf.float32, y_dtype=tf.int32)
class Fast_AT(torch.nn.Module):
def __init__(self):
torch.nn.Module.__init__(self)
self.model = PreActResNet18().cuda()
self._mean_torch = torch.tensor((0.4914, 0.4822, 0.4465)).view(3,1,1).cuda()
self._std_torch = torch.tensor((0.2471, 0.2435, 0.2616)).view(3,1,1).cuda()
''' This file provides a wrapper class for Free_AT (https://github.com/mahyarnajibi/FreeAdversarialTraining) model for ImageNet dataset. '''
import sys
import os
import torch
import torchvision.models as models
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path(
'./imagenet/free_at/resnet50_free_adv_step4_eps4_repeat4_bs256/model_best.pth.tar'
)
def load(_):
model = Free_AT()
model.load()
return model
@pytorch_classifier_with_logits(n_class=1000,
x_min=0.0,
x_max=1.0,
x_shape=(288, 288, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class Free_AT(torch.nn.Module):
def __init__(self):
''' This file provides a wrapper class for Robust_Overfitting (https://github.com/locuslab/robust_overfitting) model for CIFAR-10 dataset. '''
import sys
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/cifar10_wide10_linf_eps8.pth')
def load(_):
model = Robust_Overfitting()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10,
x_min=0.0,
x_max=1.0,
x_shape=(32, 32, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class Robust_Overfitting(torch.nn.Module):
def __init__(self):
torch.nn.Module.__init__(self)
THIRD_PARTY_PATH = os.path.abspath(
os.path.join(os.path.dirname(os.path.abspath(__file__)),
'../../third_party'))
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
import torch
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
from trades.models.wideresnet import WideResNet
MODEL_PATH = get_res_path('./cifar10/wrn.pt')
def load(_):
model = WideResNet_TRADES()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10,
x_min=0.0,
x_max=1.0,
x_shape=(32, 32, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class WideResNet_TRADES(torch.nn.Module):
THIRD_PARTY_PATH = os.path.abspath(
os.path.join(os.path.dirname(os.path.abspath(__file__)),
'../../third_party'))
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
import torch
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
from trades.models.wideresnet import WideResNet
MODEL_PATH = get_res_path('./cifar10/label_smoothing/model_best.pth')
def load(_):
model = Label_Smoothing()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10,
x_min=0.0,
x_max=1.0,
x_shape=(32, 32, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class Label_Smoothing(torch.nn.Module):
'../../third_party'))
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
from deepdefense_pytorch.models.cifar10 import ConvNet
from ares.model import ClassifierWithLogits
from ares.utils import get_res_path, download_res
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
import torch
import scipy.io
import numpy as np
import tensorflow as tf
MODEL_PATH = get_res_path('./cifar10/deepdefense')
def load(_):
model = ConvNet_DeepDefense()
model.load(MODEL_PATH)
return model
def download(model_path):
mat_name = 'cifar10-convnet-15742544.mat'
mat_path = os.path.abspath(os.path.join(model_path, mat_name))
tar_path = os.path.abspath(
os.path.join(model_path, 'mnist-cifar10-data-model.tar'))
if not os.path.exists(mat_path):
if not os.path.exists(os.path.dirname(tar_path)):
import sys
import os
THIRD_PARTY_PATH = os.path.abspath(os.path.join(os.path.dirname(os.path.abspath(__file__)), '../../third_party'))
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
from functools import partial
import tensorflow as tf
from ares.model import ClassifierWithLogits
from ares.utils import get_res_path
from cifar10_challenge.model import Model
MODEL_PATH = get_res_path('./cifar10/free_at/c10seed1cifar10_m8_eps8.0_b128')
def load(session):
model = Free_AT()
model.load(session, MODEL_PATH)
return model
def convert_model(model_path):
session = tf.Session(config=tf.ConfigProto(device_count={'GPU': 0}))
reader = tf.train.load_checkpoint(os.path.join(model_path, 'checkpoint-79999'))
var_list = list(reader.get_variable_to_dtype_map().keys())
for var in var_list:
tf.Variable(reader.get_tensor(var), name=var.replace(var, 'free_at/'+var))
session.run(tf.global_variables_initializer())
from ares.utils import get_res_path
from ares.defense.jpeg_compression import jpeg_compression
import resnet56
MODEL_PATH = get_res_path('./cifar10/resnet56.ckpt')
def load(session):
model = ResNet56_JPEG()
model.load(MODEL_PATH, session)
return model
@jpeg_compression()
class ResNet56_JPEG(resnet56.ResNet56):
pass
if __name__ == '__main__':
resnet56.download(MODEL_PATH)
''' This file provides a wrapper class for Feature_Scatter (https://github.com/Haichao-Zhang/FeatureScatter) model for CIFAR-10 dataset. '''
import sys
import os
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/checkpoint-199-ipot')
def load(_):
model = Feature_Scatter()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10, x_min=0.0, x_max=1.0,
x_shape=(32, 32, 3), x_dtype=tf.float32, y_dtype=tf.int32)
class Feature_Scatter(torch.nn.Module):
def __init__(self):
torch.nn.Module.__init__(self)
self.model = WideResNet(depth=28,
num_classes=10,
widen_factor=10).cuda()
''' CIFAR-10 dataset. '''
import tensorflow as tf
import numpy as np
from keras.datasets.cifar10 import load_data
from ares.utils import get_res_path
PATH_TARGET = get_res_path('cifar10/target.npy')
def load_dataset_for_classifier(classifier,
offset=0,
load_target=False,
target_label=None):
''' Get an CIFAR-10 dataset in tf.data.Dataset format.
The first element of the dataset is the index, the second one is the image tensor with shape of the classifier's
``x_shape`` in the classifier's ``x_dtype``, the third one is the label in the classifier's ``y_dtype``. If
``load_target`` is true, the target label would be returned as the fourth element of the dataset.
:param offset: Ignore the first ``offset`` images.
:param load_target: Whether to load the target label.
:param target_label: If it is a integer, the returned dataset would only include data points with this label.
:return: A ``tf.data.Dataset`` instance.
'''
label_dtype = classifier.y_dtype
x_shape, x_dtype, x_min, x_max = classifier.x_shape, classifier.x_dtype, classifier.x_min, classifier.x_max
dataset = load_dataset(offset=offset,
label_dtype=label_dtype,
''' This file provides a wrapper class for Pre_Training (https://github.com/hendrycks/pre-training) model for CIFAR-10 dataset. '''
import sys
import os
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/cifar10wrn_baseline_epoch_4.pt')
def load(_):
model = Pre_Training()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10,
x_min=0.0,
x_max=1.0,
x_shape=(32, 32, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class Pre_Training(torch.nn.Module):
def __init__(self):
''' This file provides a wrapper class for Fast_AT (https://github.com/locuslab/fast_adversarial) model for ImageNet dataset. '''
import sys
import os
import torch
import torchvision.models as models
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./imagenet/imagenet_model_weights_4px.pth')
def load(_):
model = Fast_AT()
model.load()
return model
@pytorch_classifier_with_logits(n_class=1000,
x_min=0.0,
x_max=1.0,
x_shape=(288, 288, 3),
x_dtype=tf.float32,
y_dtype=tf.int32)
class Fast_AT(torch.nn.Module):
def __init__(self):
torch.nn.Module.__init__(self)
self.model = models.__dict__['resnet50']()
import sys
import os
THIRD_PARTY_PATH = os.path.abspath(
os.path.join(os.path.dirname(os.path.abspath(__file__)),
'../../third_party'))
if THIRD_PARTY_PATH not in sys.path:
sys.path.append(THIRD_PARTY_PATH)
from functools import partial
import tensorflow as tf
from ares.model import ClassifierWithLogits
from ares.utils import get_res_path, download_res
MODEL_PATH = get_res_path('./cifar10/pgd_at')
from cifar10_challenge.model import Model
def load(session):
model = PgdAT()
model.load(session, MODEL_PATH)
return model
def download(model_path):
if not os.path.exists(
os.path.join(model_path, 'models/adv_trained/checkpoint')):
zip_path = os.path.join(model_path, 'adv_trained.zip')
if not os.path.exists(zip_path):
import os
import numpy as np
import tensorflow as tf
import keras
from keras.layers import Dense, Conv2D, BatchNormalization, Activation
from keras.layers import AveragePooling2D, Input, Flatten
from keras.regularizers import l2
from keras.models import Model
from ares.model import ClassifierWithLogits
from ares.utils import get_res_path, download_res
MODEL_PATH = get_res_path('./cifar10/adp')
def load(session):
model = ADP()
model.load(session, MODEL_PATH)
return model
def download(model_path):
if not os.path.exists(model_path):
os.makedirs(model_path, exist_ok=True)
h5_name = 'cifar10_ResNet110v2_model.200.h5'
h5_path = os.path.join(model_path, h5_name)
if not os.path.exists(h5_path):
url = 'http://ml.cs.tsinghua.edu.cn/~tianyu/ADP/pretrained_models/ADP_standard_3networks/' + h5_name
download_res(url, h5_path)
npy_path = os.path.join(model_path, 'mean.npy')
if not os.path.exists(npy_path):
''' This file provides a wrapper class for RST (https://github.com/yaircarmon/semisup-adv) model for CIFAR-10 dataset. '''
import sys
import os
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import tensorflow as tf
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
from ares.utils import get_res_path
MODEL_PATH = get_res_path('./cifar10/rst_adv.pt.ckpt')
def load(_):
model = RST()
model.load()
return model
@pytorch_classifier_with_logits(n_class=10, x_min=0.0, x_max=1.0,
x_shape=(32, 32, 3), x_dtype=tf.float32, y_dtype=tf.int32)
class RST(torch.nn.Module):
def __init__(self):
torch.nn.Module.__init__(self)
self.model = WideResNet(depth=28,
num_classes=10,
widen_factor=10)
if MODULE_PATH not in sys.path:
sys.path.append(MODULE_PATH)
from convex_adversarial_pytorch.examples.problems import cifar_model_resnet, Flatten
from convex_adversarial.utils import Dense, DenseSequential
sys.path.remove(MODULE_PATH)
import torch
import numpy as np
import tensorflow as tf
from ares.utils import get_res_path, download_res
from ares.model.pytorch_wrapper import pytorch_classifier_with_logits
MODEL_PATH = get_res_path('./cifar10/convex')
def load(_):
model = ResNet_Convex()
model.load(MODEL_PATH)
return model
def download(model_path):
pth_path = os.path.join(model_path, 'cifar_resnet_2px.pth')
if not os.path.exists(pth_path):
if not os.path.exists(os.path.dirname(pth_path)):
os.makedirs(os.path.dirname(pth_path), exist_ok=True)
url = 'https://github.com/locuslab/convex_adversarial/blob/master/models_scaled/cifar_resnet_2px.pth?raw=true'
download_res(url, pth_path)
