Skip to content

Commit eba1a0d

Browse files
author
Saswat Das
committed
results add
1 parent bfbe8d4 commit eba1a0d

15 files changed

+337
-0
lines changed

‎cmnist.py

Lines changed: 169 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,169 @@
1+
#Experiment Objectives:
2+
#1. Generate Attack from model j and test accuracies, L2 Distances and Reconstructions on model i (!=j)
3+
4+
import torch
5+
import numpy as np
6+
import torch
7+
import torch.nn as nn
8+
import torch.nn.functional as F
9+
from tqdm import tqdm
10+
import os
11+
import matplotlib.pyplot as plt
12+
from advertorch.attacks import LinfPGDAttack, GradientSignAttack, CarliniWagnerL2Attack, LinfBasicIterativeAttack
13+
# import seaborn as sns
14+
15+
import sys
16+
sys.path.insert(0,'/raid/sdas_ma/Adversarial_CapsNet_Pytorch/')
17+
from model.net import *
18+
from model.cnn_net import *
19+
from utils.training import *
20+
from data.data import *
21+
22+
23+
base_path = '/raid/sdas_ma/Adversarial_CapsNet_Pytorch/'
24+
model_path = "/raid/sdas_ma/Adversarial_CapsNet_Pytorch/weights/"#os.path.join(os.getcwd(), "weights")
25+
26+
Caps_args = {
27+
'DATASET_NAME':'mnist',
28+
'num_classes':10,
29+
30+
'USE_CUDA': True if torch.cuda.is_available() else False,
31+
'BATCH_SIZE': 512,
32+
33+
##For Decoder
34+
'num_features':160,
35+
'LReLU_negative_slope':0.1,
36+
'input_height':28,
37+
'input_width':28,
38+
'input_channel':1,
39+
}
40+
41+
CNN_args = {
42+
'DATASET_NAME':'mnist',
43+
'num_classes':10,
44+
45+
'USE_CUDA': True if torch.cuda.is_available() else False,
46+
'BATCH_SIZE': 256,
47+
#For Decoder
48+
'num_features':160,
49+
'LReLU_negative_slope':0.1,
50+
'input_height':28,
51+
'input_width':28,
52+
'input_channel':1,
53+
'type':'plusCR',
54+
}
55+
56+
class Caps_Config:
57+
def __init__(self, dataset='mnist'):
58+
# CNN (cnn)
59+
self.cnn_in_channels = 1
60+
self.cnn_out_channels = 12
61+
self.cnn_kernel_size = 15
62+
63+
# Primary Capsule (pc)
64+
self.pc_num_capsules = 1
65+
self.pc_in_channels = 12
66+
self.pc_out_channels = 16
67+
self.pc_kernel_size = 8
68+
self.pc_num_routes = 7 * 7
69+
70+
# Digit Capsule 1 (dc)
71+
self.dc_num_capsules = 49
72+
self.dc_num_routes = 7 * 7
73+
self.dc_in_channels = 16
74+
self.dc_out_channels = 16 #1
75+
76+
# Digit Capsule 2 (dc)
77+
self.dc_2_num_capsules = 10
78+
self.dc_2_num_routes = 7 * 7
79+
self.dc_2_in_channels = 16 #1
80+
self.dc_2_out_channels = 16
81+
82+
# Decoder
83+
self.input_width = 28
84+
self.input_height = 28
85+
86+
class CNN_Config:
87+
def __init__(self, dataset='mnist'):
88+
# CONV1
89+
self.conv1_in = 1
90+
self.conv1_out = 12
91+
self.conv1_kernel_size = 15
92+
93+
# CONV2
94+
self.conv2_in = 12
95+
self.conv2_out = 16
96+
self.conv2_kernel_size = 8
97+
98+
# FC1
99+
self.fc1_in = 7 * 7 * 16
100+
self.fc1_out = 784
101+
102+
# FC1
103+
self.fc2_in = 784
104+
self.fc2_out = 160
105+
106+
torch.manual_seed(1)
107+
108+
################Loading Dataset####################
109+
cmnistlist = ["brightness", "canny_edges", "dotted_line", "fog" , "glass_blur" , "identity", "impulse_noise", "motion_blur", "rotate", "scale", "shear" , "shot_noise" , "spatter", "stripe", "translate", "zigzag"]
110+
cmnistpath = os.path.join(base_path,"data/mnist_c/")
111+
112+
def test_n_l2(model, model_name):
113+
net.eval()
114+
Und_l2 = {key:torch.tensor([],dtype=torch.bool) for key in cmnistlist}
115+
l2_distances_all = {key:torch.tensor([],dtype=torch.int16) for key in cmnistlist}
116+
117+
for corruption in cmnistlist:
118+
datapath = os.path.join(cmnistpath, corruption)
119+
unnorm_data = torch.tensor(np.load(os.path.join(datapath, "test_images.npy"))/255).permute((0,3,1,2))
120+
labels = torch.tensor(np.load(os.path.join(datapath, "test_labels.npy")))
121+
if(labels.min()==1):
122+
labels -= 1
123+
unnorm_data, labels = unnorm_data.cuda().type(torch.cuda.FloatTensor), labels.cuda()
124+
#normalizing
125+
data = (unnorm_data - 0.1307)/0.3081
126+
with torch.no_grad():
127+
output, reconstructions, max_length_indices = net(data)
128+
l2_distances = ((reconstructions.view(unnorm_data.size(0),-1)-unnorm_data.view(unnorm_data.size(0), -1))**2).sum(1).squeeze().detach()
129+
Und_l2[corruption] = torch.cat((Und_l2[corruption], (max_length_indices != labels).detach().cpu()))
130+
l2_distances_all[corruption] = torch.cat((l2_distances_all[corruption], l2_distances.detach().cpu()))
131+
Und_l2[corruption] = Und_l2[corruption].numpy()
132+
l2_distances_all[corruption] = l2_distances_all[corruption].numpy()
133+
print(model_name," : " ,corruption," : ",np.sum(Und_l2[corruption])/100," // ",np.sum(l2_distances_all[corruption][Und_l2[corruption]]<45)/100)
134+
np.save(os.path.join(base_path , "results", str("cmnist"+model_name+"Undl2.npy")), Und_l2)
135+
np.save(os.path.join(base_path , "results", str("cmnist"+model_name+"l2_dist.npy")), l2_distances_all)
136+
137+
def capsnet():
138+
config = Caps_Config()
139+
net = CapsNet(Caps_args, config)
140+
# capsule_net = torch.nn.DataParallel(capsule_net)
141+
if Caps_args['USE_CUDA']:
142+
net = net.cuda()
143+
net.load_state_dict(torch.load(os.path.join(model_path, 'CapsNet_mnist.pth'), map_location='cpu'))
144+
return net
145+
146+
def CNN(model_type):
147+
CNN_args['type'] = model_type
148+
config = CNN_Config()
149+
net = CNNnet(CNN_args, config)
150+
net.load_state_dict(torch.load(os.path.join(model_path, 'CNN'+model_type+'_mnist.pth'), map_location='cpu'))
151+
if CNN_args['USE_CUDA']:
152+
net = net.cuda()
153+
return net
154+
155+
def load_model(model_name):
156+
if(model_name=="capsnet"):
157+
net = capsnet()
158+
return net
159+
else:
160+
net = CNN(model_name)
161+
return net
162+
163+
model_name_list = ["capsnet", "plusCR", "plusR"]
164+
for model_name in model_name_list:
165+
net = load_model(model_name)
166+
test_n_l2(net, model_name)
167+
168+
169+

‎cmnist_leftovers.py

Lines changed: 168 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,168 @@
1+
#Experiment Objectives:
2+
#1. Generate Attack from model j and test accuracies, L2 Distances and Reconstructions on model i (!=j)
3+
4+
import torch
5+
import numpy as np
6+
import torch
7+
import torch.nn as nn
8+
import torch.nn.functional as F
9+
from tqdm import tqdm
10+
import os
11+
import matplotlib.pyplot as plt
12+
from advertorch.attacks import LinfPGDAttack, GradientSignAttack, CarliniWagnerL2Attack, LinfBasicIterativeAttack
13+
# import seaborn as sns
14+
15+
import sys
16+
sys.path.insert(0,'/raid/sdas_ma/Adversarial_CapsNet_Pytorch/')
17+
from model.net import *
18+
from model.cnn_net import *
19+
from utils.training import *
20+
from data.data import *
21+
22+
23+
base_path = '/raid/sdas_ma/Adversarial_CapsNet_Pytorch/'
24+
model_path = "/raid/sdas_ma/Adversarial_CapsNet_Pytorch/weights/"#os.path.join(os.getcwd(), "weights")
25+
26+
Caps_args = {
27+
'DATASET_NAME':'mnist',
28+
'num_classes':10,
29+
30+
'USE_CUDA': True if torch.cuda.is_available() else False,
31+
'BATCH_SIZE': 512,
32+
33+
##For Decoder
34+
'num_features':160,
35+
'LReLU_negative_slope':0.1,
36+
'input_height':28,
37+
'input_width':28,
38+
'input_channel':1,
39+
}
40+
41+
CNN_args = {
42+
'DATASET_NAME':'mnist',
43+
'num_classes':10,
44+
45+
'USE_CUDA': True if torch.cuda.is_available() else False,
46+
'BATCH_SIZE': 256,
47+
#For Decoder
48+
'num_features':160,
49+
'LReLU_negative_slope':0.1,
50+
'input_height':28,
51+
'input_width':28,
52+
'input_channel':1,
53+
'type':'plusCR',
54+
}
55+
56+
class Caps_Config:
57+
def __init__(self, dataset='mnist'):
58+
# CNN (cnn)
59+
self.cnn_in_channels = 1
60+
self.cnn_out_channels = 12
61+
self.cnn_kernel_size = 15
62+
63+
# Primary Capsule (pc)
64+
self.pc_num_capsules = 1
65+
self.pc_in_channels = 12
66+
self.pc_out_channels = 16
67+
self.pc_kernel_size = 8
68+
self.pc_num_routes = 7 * 7
69+
70+
# Digit Capsule 1 (dc)
71+
self.dc_num_capsules = 49
72+
self.dc_num_routes = 7 * 7
73+
self.dc_in_channels = 16
74+
self.dc_out_channels = 16 #1
75+
76+
# Digit Capsule 2 (dc)
77+
self.dc_2_num_capsules = 10
78+
self.dc_2_num_routes = 7 * 7
79+
self.dc_2_in_channels = 16 #1
80+
self.dc_2_out_channels = 16
81+
82+
# Decoder
83+
self.input_width = 28
84+
self.input_height = 28
85+
86+
class CNN_Config:
87+
def __init__(self, dataset='mnist'):
88+
# CONV1
89+
self.conv1_in = 1
90+
self.conv1_out = 12
91+
self.conv1_kernel_size = 15
92+
93+
# CONV2
94+
self.conv2_in = 12
95+
self.conv2_out = 16
96+
self.conv2_kernel_size = 8
97+
98+
# FC1
99+
self.fc1_in = 7 * 7 * 16
100+
self.fc1_out = 784
101+
102+
# FC1
103+
self.fc2_in = 784
104+
self.fc2_out = 160
105+
106+
torch.manual_seed(1)
107+
108+
################Loading Dataset####################
109+
cmnistlist = ["contrast","elastic_transform" ,"gaussian_blur" ,"inverse" ,"line" ,"pixelate" ,"saturate" ,"speckle_noise", "defocus_blur" ,"frost" ,"gaussian_noise" ,"jpeg_compression" ,"pessimal_noise" ,"quantize" ,"snow" ,"zoom_blur"]
110+
cmnistpath = os.path.join(base_path,"data/mnist_c_leftovers/")
111+
112+
def test_n_l2(model, model_name):
113+
net.eval()
114+
Und_l2 = {key:torch.tensor([],dtype=torch.bool) for key in cmnistlist}
115+
l2_distances_all = {key:torch.tensor([],dtype=torch.int16) for key in cmnistlist}
116+
117+
for corruption in cmnistlist:
118+
datapath = os.path.join(cmnistpath, corruption)
119+
unnorm_data = torch.tensor(np.load(os.path.join(datapath, "test_images.npy"))/255).permute((0,3,1,2))
120+
labels = torch.tensor(np.load(os.path.join(datapath, "test_labels.npy"))).reshape(-1)
121+
if(labels.min()==1):
122+
labels -= 1
123+
unnorm_data, labels = unnorm_data.cuda().type(torch.cuda.FloatTensor), labels.cuda()
124+
#normalizing
125+
data = (unnorm_data - 0.1307)/0.3081
126+
with torch.no_grad():
127+
output, reconstructions, max_length_indices = net(data)
128+
l2_distances = ((reconstructions.view(unnorm_data.size(0),-1)-unnorm_data.view(unnorm_data.size(0), -1))**2).sum(1).squeeze().detach()
129+
Und_l2[corruption] = (max_length_indices != labels).detach().cpu()
130+
l2_distances_all[corruption] = l2_distances.detach().cpu()
131+
Und_l2[corruption] = Und_l2[corruption].numpy()
132+
print(model_name," : " ,corruption," : ",np.sum(Und_l2[corruption])/100," // ",np.count_nonzero(l2_distances_all[corruption][Und_l2[corruption]]<45)/100)
133+
np.save(os.path.join(base_path , "results", str("cmnist_leftovers"+model_name+"Undl2.npy")), Und_l2)
134+
np.save(os.path.join(base_path , "results", str("cmnist_leftovers"+model_name+"l2_dist.npy")), l2_distances_all)
135+
136+
def capsnet():
137+
config = Caps_Config()
138+
net = CapsNet(Caps_args, config)
139+
# capsule_net = torch.nn.DataParallel(capsule_net)
140+
if Caps_args['USE_CUDA']:
141+
net = net.cuda()
142+
net.load_state_dict(torch.load(os.path.join(model_path, 'CapsNet_mnist.pth'), map_location='cpu'))
143+
return net
144+
145+
def CNN(model_type):
146+
CNN_args['type'] = model_type
147+
config = CNN_Config()
148+
net = CNNnet(CNN_args, config)
149+
net.load_state_dict(torch.load(os.path.join(model_path, 'CNN'+model_type+'_mnist.pth'), map_location='cpu'))
150+
if CNN_args['USE_CUDA']:
151+
net = net.cuda()
152+
return net
153+
154+
def load_model(model_name):
155+
if(model_name=="capsnet"):
156+
net = capsnet()
157+
return net
158+
else:
159+
net = CNN(model_name)
160+
return net
161+
162+
model_name_list = ["capsnet", "plusCR", "plusR"]
163+
for model_name in model_name_list:
164+
net = load_model(model_name)
165+
test_n_l2(net, model_name)
166+
167+
168+

‎results/cmnist_leftoverscapsnetUndl2.npy

157 KB
Binary file not shown.

‎results/cmnist_leftoverscapsnetl2_dist.npy

630 KB
Binary file not shown.

‎results/cmnist_leftoversplusCRUndl2.npy

157 KB
Binary file not shown.

‎results/cmnist_leftoversplusCRl2_dist.npy

630 KB
Binary file not shown.

‎results/cmnist_leftoversplusRUndl2.npy

157 KB
Binary file not shown.

‎results/cmnist_leftoversplusRl2_dist.npy

630 KB
Binary file not shown.

‎results/cmnistcapsnetUndl2.npy

157 KB
Binary file not shown.

‎results/cmnistcapsnetl2_dist.npy

626 KB
Binary file not shown.

0 commit comments

Comments
 (0)