File size: 6,145 Bytes
2e1a3f8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import numpy as np
import torch
import glob

from captum._utils.common import _get_module_from_name

# compute rollout between attention layers
def compute_rollout_attention(all_layer_matrices, start_layer=0):
    # adding residual consideration- code adapted from https://github.com/samiraabnar/attention_flow
    num_tokens = all_layer_matrices[0].shape[1]
    batch_size = all_layer_matrices[0].shape[0]
    eye = torch.eye(num_tokens).expand(batch_size, num_tokens, num_tokens).to(all_layer_matrices[0].device)
    all_layer_matrices = [all_layer_matrices[i] + eye for i in range(len(all_layer_matrices))]
    matrices_aug = [all_layer_matrices[i] / all_layer_matrices[i].sum(dim=-1, keepdim=True)
                          for i in range(len(all_layer_matrices))]
    joint_attention = matrices_aug[start_layer]
    for i in range(start_layer+1, len(matrices_aug)):
        joint_attention = matrices_aug[i].bmm(joint_attention)
    return joint_attention

class Generator:
    def __init__(self, model, key="bert.encoder.layer"):
        self.model = model
        self.key = key
        self.model.eval()

    def forward(self, input_ids, attention_mask):
        return self.model(input_ids, attention_mask)

    def _build_one_hot(self, output, index):
        if index == None:
            index = np.argmax(output.cpu().data.numpy(), axis=-1)

        one_hot = np.zeros((1, output.size()[-1]), dtype=np.float32)
        one_hot[0, index] = 1
        one_hot_vector = one_hot
        one_hot = torch.from_numpy(one_hot).requires_grad_(True).to(output.device)
        one_hot = torch.sum(one_hot * output)

        return one_hot, one_hot_vector

    def generate_LRP(self, input_ids, attention_mask,

                     index=None, start_layer=11):
        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        kwargs = {"alpha": 1}

        one_hot, one_hot_vector = self._build_one_hot(output, index)
        self.model.zero_grad()
        one_hot.backward(retain_graph=True)

        self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)

        cams = []
        blocks = _get_module_from_name(self.model, self.key)
        for blk in blocks:
            grad = blk.attention.self.get_attn_gradients()
            cam = blk.attention.self.get_attn_cam()
            cam = cam[0].reshape(-1, cam.shape[-1], cam.shape[-1])
            grad = grad[0].reshape(-1, grad.shape[-1], grad.shape[-1])
            cam = grad * cam
            cam = cam.clamp(min=0).mean(dim=0)
            cams.append(cam.unsqueeze(0))
        rollout = compute_rollout_attention(cams, start_layer=start_layer)
        rollout[:, 0, 0] = rollout[:, 0].min()
        return rollout[:, 0]

    def generate_LRP_last_layer(self, input_ids, attention_mask,

                     index=None):
        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        kwargs = {"alpha": 1}

        one_hot, one_hot_vector = self._build_one_hot(output, index)

        self.model.zero_grad()
        one_hot.backward(retain_graph=True)

        self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)

        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn_cam()[0]
        cam = cam.clamp(min=0).mean(dim=0).unsqueeze(0)
        cam[:, 0, 0] = 0
        return cam[:, 0]

    def generate_full_lrp(self, input_ids, attention_mask,

                     index=None):
        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        kwargs = {"alpha": 1}

        one_hot, one_hot_vector = self._build_one_hot(output, index)

        self.model.zero_grad()
        one_hot.backward(retain_graph=True)

        cam = self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)
        cam = cam.sum(dim=2)
        cam[:, 0] = 0
        return cam

    def generate_attn_last_layer(self, input_ids, attention_mask,

                     index=None):
        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn()[0]
        cam = cam.mean(dim=0).unsqueeze(0)
        cam[:, 0, 0] = 0
        return cam[:, 0]

    def generate_rollout(self, input_ids, attention_mask, start_layer=0, index=None):
        self.model.zero_grad()
        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        blocks = _get_module_from_name(self.model, self.key)
        all_layer_attentions = []
        for blk in blocks:
            attn_heads = blk.attention.self.get_attn()
            avg_heads = (attn_heads.sum(dim=1) / attn_heads.shape[1]).detach()
            all_layer_attentions.append(avg_heads)
        rollout = compute_rollout_attention(all_layer_attentions, start_layer=start_layer)
        rollout[:, 0, 0] = 0
        return rollout[:, 0]

    def generate_attn_gradcam(self, input_ids, attention_mask, index=None):
        output = self.model(input_ids=input_ids, attention_mask=attention_mask)[0]
        kwargs = {"alpha": 1}

        if index == None:
            index = np.argmax(output.cpu().data.numpy(), axis=-1)

        one_hot, one_hot_vector = self._build_one_hot(output, index)

        self.model.zero_grad()
        one_hot.backward(retain_graph=True)

        self.model.relprop(torch.tensor(one_hot_vector).to(input_ids.device), **kwargs)

        cam = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn()
        grad = _get_module_from_name(self.model, self.key)[-1].attention.self.get_attn_gradients()

        cam = cam[0].reshape(-1, cam.shape[-1], cam.shape[-1])
        grad = grad[0].reshape(-1, grad.shape[-1], grad.shape[-1])
        grad = grad.mean(dim=[1, 2], keepdim=True)
        cam = (cam * grad).mean(0).clamp(min=0).unsqueeze(0)
        cam = (cam - cam.min()) / (cam.max() - cam.min())
        cam[:, 0, 0] = 0
        return cam[:, 0]