class DefenseTorch(Defense):
def __init__(self):
import torch
self.convnet = AllConvModelTorch(num_classes=10,
num_filters=64,
input_shape=[3, 32, 32])
self.convnet.load_state_dict(
torch.load(get_checkpoint_abs_path(MODEL_PATH) + ".torchmodel"))
class DefenseTorch(Defense):
def __init__(self):
import torch
self.convnet = AllConvModelTorch(num_classes=10,
num_filters=64,
input_shape=[3, 32, 32])
self.convnet.load_state_dict(
torch.load(get_checkpoint_abs_path(MODEL_PATH) + ".torchmodel"))
def augment(self, x):
return data.augment_strong_np(x.transpose((0, 2, 3, 1))).transpose(
(0, 3, 1, 2))
class DefenseTorch(Defense):
def __init__(self):
import torch
self.convnet = AllConvModelTorch(num_classes=128,
num_filters=64,
input_shape=[3, 32, 32])
self.convnet.load_state_dict(
torch.load(get_checkpoint_abs_path(MODEL_PATH) + ".torchmodel"))
self.convnet.layers = self.convnet.layers[:-1]
self.features = np.load(get_checkpoint_abs_path(FEATURE_PATH)) # 50000 x 128
self.labels = np.load(get_checkpoint_abs_path(LABEL_PATH)) # 50000 x 1
self.NEAREST = 8
class DefenseTorch(Defense):
def __init__(self):
import torch
self.convnet = AllConvModelTorch(num_classes=10,
num_filters=64,
input_shape=[3, 32, 32])
self.convnet.load_state_dict(
torch.load(get_checkpoint_abs_path(MODEL_PATH) + ".torchmodel"))
self.signature = np.load(get_checkpoint_abs_path(SIGNATURE_PATH))
self.backdoor = np.load(get_checkpoint_abs_path(BACKDOOR_PATH))
self.backdoor = self.backdoor.transpose((0, 3, 1, 2))
self.hidden = get_hidden_layer(self.convnet, 3, (2, 3))
self.to_tensor = torch.tensor
class DefenseTorch(Defense):
def __init__(self):
import torch
self.convnet = AllConvModelTorch(num_classes=10,
num_filters=64,
input_shape=[3, 32, 32])
self.convnet.load_state_dict(
torch.load(get_checkpoint_abs_path(MODEL_PATH) + ".torchmodel"))
def blur(self, x):
x_pad = np.pad(x, [(0, 0), (0, 0), (1, 1), (1, 1)])
x_pad = (x_pad[:, :, :1] + x_pad[:, :, :-1]) / 2
x_pad = (x_pad[:, :, :, :1] + x_pad[:, :, :, :-1]) / 2
return x_pad
class DefenseTorch(Defense):
def __init__(self):
import torch
class Jump(torch.nn.Module):
def forward(self, x):
activation = torch.nn.LeakyReLU(0.2)
x = activation(x)
x += (x > 0).float() * 5
return x
self.convnet = AllConvModelTorch(num_classes=10,
num_filters=64,
input_shape=[3, 32, 32],
activation=Jump())
self.convnet.load_state_dict(torch.load(MODEL_PATH + ".torchmodel"))
class DefenseTorch(Defense):
def __init__(self):
import torch
self.convnet = AllConvModelTorch(num_classes=10,
num_filters=64,
input_shape=[3 * 20, 32, 32])
self.convnet.load_state_dict(
torch.load(get_checkpoint_abs_path(MODEL_PATH) + ".torchmodel"))
def encode(self, xs):
thresholds = np.arange(0, 1, .05) + .05
shape = xs.shape
less_than_threshold = xs[:, :, None, :, :] < thresholds[None, None, :,
None, None]
xs = np.array(less_than_threshold, dtype=np.float32)
xs = np.reshape(xs,
[-1, shape[1] * len(thresholds), shape[2], shape[3]])
return xs