-
Notifications
You must be signed in to change notification settings - Fork 0
/
vit_grad_rollout.py
163 lines (134 loc) · 6.12 KB
/
vit_grad_rollout.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
import torch
from PIL import Image
import numpy
import sys
from torchvision import transforms
import numpy as np
import cv2
import pdb
import torch.nn.functional as F
def grad_rollout(attentions, gradients, discard_ratio):
result = torch.eye(attentions[0].size(-1))
with torch.no_grad():
for attention, grad in zip(attentions, gradients):
weights = grad
attention_heads_fused = (attention*weights).mean(axis=1)
attention_heads_fused[attention_heads_fused < 0] = 0
# Drop the lowest attentions, but
# don't drop the class token
flat = attention_heads_fused.view(attention_heads_fused.size(0), -1)
_, indices = flat.topk(int(flat.size(-1)*discard_ratio), -1, False)
#indices = indices[indices != 0]
flat[0, indices] = 0
I = torch.eye(attention_heads_fused.size(-1))
a = (attention_heads_fused + 1.0*I)/2
a = a / a.sum(dim=-1)
result = torch.matmul(a, result)
# Look at the total attention between the class token,
# and the image patches
mask = result[0, 0 , 1 :]
# In case of 224x224 image, this brings us from 196 to 14
width = int(mask.size(-1)**0.5)
mask = mask.reshape(width, width).numpy()
mask = mask / np.max(mask)
return mask
def grad_rollout_batch(attentions, gradients, discard_ratio):
result = torch.eye(attentions[0].size(-1)).unsqueeze(0).repeat(attentions[0].size(0),1,1).to(attentions[0].device)
# result = torch.eye(attentions[0].size(-1))
with torch.no_grad():
for attention, grad in zip(attentions, gradients):
weights = grad
attention_heads_fused = (attention*weights).mean(axis=1)
attention_heads_fused[attention_heads_fused < 0] = 0
# Drop the lowest attentions, but
# don't drop the class token
flat = attention_heads_fused.view(attention_heads_fused.size(0), -1)
_, indices = flat.topk(int(flat.size(-1)*discard_ratio), -1, False)
#indices = indices[indices != 0]
flat.scatter_(-1, indices, torch.zeros(indices.shape).cuda())
I = torch.eye(attention_heads_fused.size(-1)).unsqueeze(0).repeat(attention_heads_fused.size(0),1,1).to(attention_heads_fused[0].device)
a = (attention_heads_fused + 1.0*I)/2
# a = a / a.sum(dim=-1)
a = a / (a.sum(dim=-1)[:,np.newaxis])
result = torch.matmul(a, result)
# Look at the total attention between the class token,
# and the image patches
mask = result[:, 0 , 1 :]
# In case of 224x224 image, this brings us from 196 to 14
width = int(mask.size(-1)**0.5)
mask = mask.reshape(mask.size(0),width, width).cpu().numpy()
# mask = mask / np.max(mask)
breakpoint()
max_div = np.max(mask,axis=(1,2))
max_div = np.repeat(np.repeat(max_div[:,np.newaxis],mask.shape[1],axis=1)[:,:,np.newaxis],mask.shape[2],axis=2)
mask = mask/(max_div+1e-8)
return mask
class VITAttentionGradRollout:
def __init__(self, model, attention_layer_name='attn_drop',
discard_ratio=0.9):
self.model = model
self.discard_ratio = discard_ratio
for name, module in self.model.named_modules():
if attention_layer_name in name:
module.register_forward_hook(self.get_attention)
module.register_backward_hook(self.get_attention_gradient)
self.attentions = []
self.attention_gradients = []
def clear_cache(self):
self.attentions = []
self.attention_gradients = []
def get_attention(self, module, input, output):
self.attentions.append(output.cpu())
def get_attention_gradient(self, module, grad_input, grad_output):
self.attention_gradients.append(grad_input[0].cpu())
def __call__(self, input_tensor, category_index):
self.model.zero_grad()
output = self.model(input_tensor)
category_mask = torch.zeros(output.size())
category_mask[:, category_index] = 1
category_mask = category_mask.to(output.device)
loss = (output*category_mask).sum()
loss.backward()
return grad_rollout(self.attentions, self.attention_gradients,
self.discard_ratio)
class VITAttentionGradRollout_Batch:
def __init__(self, model, attention_layer_name='attn_drop',
discard_ratio=0.9):
self.model = model
self.discard_ratio = discard_ratio
self.hook_handles = []
for name, module in self.model.named_modules():
if attention_layer_name in name:
self.hook_handles.append(module.register_forward_hook(self.get_attention))
self.hook_handles.append(module.register_backward_hook(self.get_attention_gradient))
self.attentions = []
self.attention_gradients = []
def remove_hooks(self):
for i in range(len(self.hook_handles)):
self.hook_handles[i].remove()
def clear_cache(self):
self.attentions = []
self.attention_gradients = []
def get_attention(self, module, input, output):
# self.attentions.append(output.cpu())
self.attentions.append(output)
def get_attention_gradient(self, module, grad_input, grad_output):
# self.attention_gradients.append(grad_input[0].cpu())
self.attention_gradients.append(grad_input[0])
# def __call__(self, input_tensor, category_indices):
def __call__(self, input_tensor, top=True):
self.model.zero_grad()
output = self.model(input_tensor)
class_idx = output.data.topk(1, dim=1)[1][0]
# category_mask = torch.zeros(output.size())
# category_mask[:, category_index] = 1
# category_mask = F.one_hot(category_indices,num_classes =output.size(1))
if top:
category_mask = F.one_hot(class_idx,num_classes =output.size(1))
else:
raise NotImplementedError
category_mask = category_mask.to(output.device)
loss = (output*category_mask).sum()
loss.backward()
return grad_rollout_batch(self.attentions, self.attention_gradients,
self.discard_ratio)