-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathattribute.py
193 lines (146 loc) · 7.55 KB
/
attribute.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import torch
import numpy as np
from torch.utils.data import DataLoader
import pandas as pd
import captum
import argparse
from captum.attr import IntegratedGradients, DeepLift, Saliency, FeatureAblation, GradientShap, KernelShap, DeepLiftShap
from captum.metrics import infidelity, sensitivity_max
from METAFormer.models import METAFormer
from METAFormer.dataloader import MultiAtlas
torch.manual_seed(1337)
np.random.seed(1337)
def perturb_func(input):
aal, cc200, dos160 = input
noise_aal = torch.tensor(np.random.normal(0, 0.003, aal.shape)).float()
noise_cc200 = torch.tensor(np.random.normal(0, 0.003, cc200.shape)).float()
noise_dos160 = torch.tensor(np.random.normal(0,0.003, dos160.shape)).float()
return (noise_aal, noise_cc200, noise_dos160), (aal - noise_aal, cc200 - noise_cc200, dos160 - noise_dos160)
def get_deep_lift_sens(model, x_list, targets, baselines):
dl = DeepLift(model)
sens = sensitivity_max(dl.attribute, x_list, target=targets, baselines=baselines)
return sens
def get_integrated_gradients_sens(model, x_list, targets, baselines):
ig = IntegratedGradients(model)
sens = sensitivity_max(ig.attribute, x_list, target=targets, baselines=baselines)
return sens
def get_feature_ablation_sens(model, x_list, targets, baselines):
fa = FeatureAblation(model)
sens = sensitivity_max(fa.attribute, x_list, target=targets, baselines=baselines)
return sens
def get_shap_sens(model, x_list,baselines, targets):
gs = GradientShap(model)
sens = sensitivity_max(gs.attribute, x_list, baselines=baselines, target=targets)
return sens
def get_saliency_sens(model, x_list, targets, baselines):
sal = Saliency(model)
sens = sensitivity_max(sal.attribute, x_list, target=targets, baselines=baselines)
return sens
def get_kernel_shap_sens(model, x_list, baselines, targets):
ks = KernelShap(model)
sens = sensitivity_max(ks.attribute, x_list, baselines=baselines, target=targets)
return sens
def get_deep_lift_shap(model, x_list, baselines, targets):
dls = DeepLiftShap(model)
sens = sensitivity_max(dls.attribute, x_list, baselines=baselines, target=targets)
return sens
def main(args):
model = METAFormer()
state_dict = torch.load(args.checkpoint)
model.load_state_dict(state_dict)
model.eval()
test_df = pd.read_csv(args.test_csv)
ds = MultiAtlas(test_df)
dl = DataLoader(ds, batch_size=256, shuffle=False)
# get batch, note batch_size is so large that all data is in one batch
x_batch, y_batch = next(iter(dl))
(x_aal, x_cc200, x_dos160) = x_batch[0], x_batch[1], x_batch[2]
y_batch = y_batch
target = y_batch.reshape(-1).to(torch.int64)
model.cpu()
model.eval()
inf_df = calc_infidelity(model, x_aal, x_cc200, x_dos160, target)
inf_df.to_csv("infidelity.csv")
sens_df = calc_sensitivity(model, x_aal, x_cc200, x_dos160, target)
sens_df.to_csv("sensitivity.csv")
sal = Saliency(model)
sal_sens = sensitivity_max(sal.attribute, (x_aal, x_cc200, x_dos160), target=target)
print("Saliency Sensitivity: ", sal_sens)
sal_attr = sal.attribute((x_aal, x_cc200, x_dos160), target=target)
sal_attr = sal_attr.detach().numpy()
sal_inf = infidelity(model, perturb_func, (x_aal, x_cc200, x_dos160),(sal_attr[0], sal_attr[1], sal_attr[2]), target=target)
print("Saliency Infidelity: ", sal_inf.item())
def calc_sensitivity(model, x_aal, x_cc200, x_dos160, target):
results = []
for i in range(-10, 11, 1):
print("Baseline: ", i/10)
baseline_aal = torch.ones_like(x_aal)*(i/10)
baseline_cc200 = torch.ones_like(x_cc200)*(i/10)
baseline_dos160 = torch.ones_like(x_dos160)*(i/10)
ig_sens = get_integrated_gradients_sens(model, (x_aal, x_cc200, x_dos160), target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
dl_sens = get_deep_lift_sens(model, (x_aal, x_cc200, x_dos160), target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
fa_sens = get_feature_ablation_sens(model, (x_aal, x_cc200, x_dos160), target, baselines=i/10)
gs_sens = get_shap_sens(model, (x_aal, x_cc200, x_dos160), (baseline_aal, baseline_cc200, baseline_dos160), target)
dls_sens = get_deep_lift_sens(model, (x_aal, x_cc200, x_dos160), target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
# save mean and std for each method
results_dict = {
"Baseline": i/10,
"ig_mean": ig_sens[0].item(),
"ig_std": ig_sens[1].item(),
"dl_mean": dl_sens[0].item(),
"dl_std": dl_sens[1].item(),
"fa_mean": fa_sens[0].item(),
"fa_std": fa_sens[1].item(),
"gs_mean": gs_sens[0].item(),
"gs_std": gs_sens[1].item(),
"dls_mean": dls_sens[0].item(),
"dls_std": dls_sens[1].item()
}
results.append(results_dict)
results_df = pd.DataFrame(results)
return results_df
def calc_infidelity(model, x_aal, x_cc200, x_dos160, target):
results = []
saliency = Saliency(model)
dl = DeepLift(model)
ig = IntegratedGradients(model)
fa = FeatureAblation(model)
shap = GradientShap(model)
dls = DeepLiftShap(model)
for i in range(-10, 11, 1):
print("Baseline: ", i/10)
baseline_aal = torch.ones_like(x_aal)*(i/10)
baseline_cc200 = torch.ones_like(x_cc200)*(i/10)
baseline_dos160 = torch.ones_like(x_dos160)*(i/10)
dl_attr = dl.attribute((x_aal, x_cc200, x_dos160), target=target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
ig_attr = ig.attribute((x_aal, x_cc200, x_dos160), target=target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
fa_attr = fa.attribute((x_aal, x_cc200, x_dos160), target=target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
shap_attr = shap.attribute((x_aal, x_cc200, x_dos160), target=target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
dls_attr = dls.attribute((x_aal, x_cc200, x_dos160), target=target, baselines=(baseline_aal, baseline_cc200, baseline_dos160))
fid_dl = infidelity(model, perturb_func, (x_aal, x_cc200, x_dos160),(dl_attr[0], dl_attr[1], dl_attr[2]), target=target)
fid_ig = infidelity(model, perturb_func, (x_aal, x_cc200, x_dos160),(ig_attr[0], ig_attr[1], ig_attr[2]), target=target)
fid_fa = infidelity(model, perturb_func, (x_aal, x_cc200, x_dos160), (fa_attr[0], fa_attr[1], fa_attr[2]), target=target)
fid_shap = infidelity(model, perturb_func, (x_aal, x_cc200, x_dos160), (shap_attr[0], shap_attr[1], shap_attr[2]), target=target)
fid_dls = infidelity(model, perturb_func, (x_aal, x_cc200, x_dos160), (dls_attr[0], dls_attr[1], dls_attr[2]), target=target)
results_dict = {
"baseline": i/10,
"dl_mean": fid_dl.mean().item(),
"dl_std": fid_dl.std().item(),
"ig_mean": fid_ig.mean().item(),
"ig_std": fid_ig.std().item(),
"dls_mean": fid_dls.mean().item(),
"dls_std": fid_dls.std().item(),
"shap_mean": fid_shap.mean().item(),
"shap_std": fid_shap.std().item(),
"fa_mean": fid_fa.mean().item(),
"fa_std": fid_fa.std().item()
}
results.append(results_dict)
results_df = pd.DataFrame(results)
return results_df
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint", type=str, required=True)
parser.add_argument("--data", type=str, required=True)
args = parser.parse_args()
main(args)