BoxEL/box_model.py at main · Box-EL/BoxEL · GitHub

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
#!/usr/bin/env python
# coding: utf-8

# In[59]:


import time
import torch
import math
import torch.nn as nn
from basic_box import Box
import torch.nn.functional as F
from torch.distributions import uniform


eps = 1e-8

def l2_side_regularizer(box, log_scale: bool = True):
    """Applies l2 regularization on all sides of all boxes and returns the sum.
    """
    min_x = box.min_embed
    delta_x = box.delta_embed

    if not log_scale:
        return torch.mean(delta_x ** 2)
    else:
        return torch.mean(F.relu(min_x + delta_x - 1 + eps )) +  torch.mean(F.relu(-min_x - eps)) #+ F.relu(torch.norm(min_x, p=2)-1)

class BoxEL(nn.Module):

    def __init__(self, vocab_size, relation_size,embed_dim, min_init_value, delta_init_value, relation_init_value, scaling_init_value, args):
        super(BoxEL, self).__init__()
        min_embedding = self.init_concept_embedding(vocab_size, embed_dim, min_init_value)
        delta_embedding = self.init_concept_embedding(vocab_size, embed_dim, delta_init_value)
        relation_embedding = self.init_concept_embedding(relation_size, embed_dim, relation_init_value)
        scaling_embedding = self.init_concept_embedding(relation_size, embed_dim, scaling_init_value)
        self.temperature = args.softplus_temp
        self.min_embedding = nn.Parameter(min_embedding)
        self.delta_embedding = nn.Parameter(delta_embedding)
        self.relation_embedding = nn.Parameter(relation_embedding)
        self.scaling_embedding = nn.Parameter(scaling_embedding)

        self.gumbel_beta = args.gumbel_beta
        self.scale = args.scale

    def forward(self, data):
        nf1_min = self.min_embedding[data[0][:,[0,2]]]
        nf1_delta = self.delta_embedding[data[0][:,[0,2]]]
        nf1_max = nf1_min+torch.exp(nf1_delta)
        boxes1 = Box(nf1_min[:, 0, :], nf1_max[:, 0, :])
        boxes2 = Box(nf1_min[:, 1, :], nf1_max[:, 1, :])
        nf1_loss, nf1_reg_loss = self.nf1_loss(boxes1, boxes2)

        nf2_min = self.min_embedding[data[1]]
        nf2_delta = self.delta_embedding[data[1]]
        nf2_max = nf2_min+torch.exp(nf2_delta)
        boxes1 = Box(nf2_min[:, 0, :], nf2_max[:, 0, :])
        boxes2 = Box(nf2_min[:, 1, :], nf2_max[:, 1, :])
        boxes3 = Box(nf2_min[:, 2, :], nf2_max[:, 2, :])
        nf2_loss,nf2_reg_loss = self.nf2_loss(boxes1, boxes2, boxes3)

        nf3_min = self.min_embedding[data[2][:,[0,2]]]
        nf3_delta = self.delta_embedding[data[2][:,[0,2]]]
        nf3_max = nf3_min+torch.exp(nf3_delta)
        relation = self.relation_embedding[data[2][:,1]]
        scaling = self.scaling_embedding[data[2][:,1]]
        boxes1 = Box(nf3_min[:, 0, :], nf3_max[:, 0, :])
        boxes2 = Box(nf3_min[:, 1, :], nf3_max[:, 1, :])
        nf3_loss,nf3_reg_loss = self.nf3_loss(boxes1, relation, scaling, boxes2)

        nf4_min = self.min_embedding[data[3][:,1:]]
        nf4_delta = self.delta_embedding[data[3][:,1:]]
        nf4_max = nf4_min+torch.exp(nf4_delta)
        relation = self.relation_embedding[data[3][:,0]]
        scaling = self.scaling_embedding[data[3][:,0]]
        boxes1 = Box(nf4_min[:, 0, :], nf4_max[:, 0, :])
        boxes2 = Box(nf4_min[:, 1, :], nf4_max[:, 1, :])
        nf4_loss,nf4_reg_loss = self.nf4_loss(relation, scaling, boxes1, boxes2)

        disjoint_min = self.min_embedding[data[4]]
        disjoint_delta = self.delta_embedding[data[4]]
        disjoint_max = disjoint_min+torch.exp(disjoint_delta)
        boxes1 = Box(disjoint_min[:, 0, :], disjoint_max[:, 0, :])
        boxes2 = Box(disjoint_min[:, 1, :], disjoint_max[:, 1, :])
        disjoint_loss,disjoint_reg_loss = self.disjoint_loss(boxes1, boxes2)

        nf1_neg_min = self.min_embedding[data[5][:,[0,2]]]
        nf1_neg_delta = self.delta_embedding[data[5][:,[0,2]]]
        nf1_neg_max = nf1_neg_min+torch.exp(nf1_neg_delta)
        boxes1 = Box(nf1_neg_min[:, 0, :], nf1_neg_max[:, 0, :])
        boxes2 = Box(nf1_neg_min[:, 1, :], nf1_neg_max[:, 1, :])
        nf1_neg_loss, nf1_neg_reg_loss = self.nf1_loss(boxes1, boxes2)
        nf1_neg_loss = 1 - nf1_neg_loss


        # role inclusion
        translation_1 = self.relation_embedding[data[6][:,0]]
        translation_2 = self.relation_embedding[data[6][:,1]]
        scaling_1 = self.scaling_embedding[data[6][:,0]]
        scaling_2 = self.scaling_embedding[data[6][:,1]]
        role_inclusion_loss = self.role_inclusion_loss(translation_1,translation_2,scaling_1,scaling_2)
        # role chain
        translation_1 = self.relation_embedding[data[7][:,0]]
        translation_2 = self.relation_embedding[data[7][:,1]]
        translation_3 = self.relation_embedding[data[7][:,2]]
        scaling_1 = self.scaling_embedding[data[7][:,0]]
        scaling_2 = self.scaling_embedding[data[7][:,1]]
        scaling_3 = self.scaling_embedding[data[7][:,2]]
        role_chain_loss = self.role_chain_loss(translation_1,translation_2,translation_3,scaling_1,scaling_2,scaling_3)

        return nf1_loss.sum(), nf1_neg_loss.sum(), nf2_loss.sum(), nf3_loss.sum(), nf4_loss.sum(), disjoint_loss.sum(), role_inclusion_loss.sum(),role_chain_loss.sum(),nf1_reg_loss, nf1_neg_reg_loss, nf2_reg_loss , nf3_reg_loss , nf4_reg_loss, disjoint_reg_loss

    def get_cond_probs(self, data):
        nf3_min = self.min_embedding[data[:,[0,2]]]
        nf3_delta = self.delta_embedding[data[:,[0,2]]]
        nf3_max = nf3_min+torch.exp(nf3_delta)
        relation = self.relation_embedding[data[:,1]]
        boxes1 = Box(nf3_min[:, 0, :], nf3_max[:, 0, :])
        boxes2 = Box(nf3_min[:, 1, :], nf3_max[:, 1, :])
        log_intersection = torch.log(torch.clamp(self.volumes(self.intersection(boxes1, boxes2)), 1e-10, 1e4))
        log_box2 = torch.log(torch.clamp(self.volumes(boxes2), 1e-10, 1e4))
        return torch.exp(log_intersection-log_box2)

    def volumes(self, boxes):
        return F.softplus(boxes.delta_embed, beta=self.temperature).prod(1)

    def intersection(self, boxes1, boxes2):
        intersections_min = torch.max(boxes1.min_embed, boxes2.min_embed)
        intersections_max = torch.min(boxes1.max_embed, boxes2.max_embed)
        intersection_box = Box(intersections_min, intersections_max)
        return intersection_box

    def inclusion_loss(self, boxes1, boxes2):
        log_intersection = torch.log(torch.clamp(self.volumes(self.intersection(boxes1, boxes2)), 1e-10, 1e4))
        log_box1 = torch.log(torch.clamp(self.volumes(boxes1), 1e-10, 1e4))
        return 1-torch.exp(log_intersection-log_box1)

    def nf1_loss(self, boxes1, boxes2):
        return self.inclusion_loss(boxes1, boxes2), l2_side_regularizer(boxes1, log_scale=True) + l2_side_regularizer(boxes2, log_scale=True)

    def nf2_loss(self, boxes1, boxes2, boxes3):
        inter_box = self.intersection(boxes1, boxes2)
        return self.inclusion_loss(inter_box, boxes3), l2_side_regularizer(inter_box, log_scale=True) + l2_side_regularizer(boxes1, log_scale=True) + l2_side_regularizer(boxes2, log_scale=True) + l2_side_regularizer(boxes3, log_scale=True)

    def nf3_loss(self, boxes1, relation, scaling, boxes2):
        trans_min = boxes1.min_embed*(scaling + eps) + relation
        trans_max = boxes1.max_embed*(scaling + eps) + relation
        trans_boxes = Box(trans_min, trans_max)
        return self.inclusion_loss(trans_boxes, boxes2), l2_side_regularizer(trans_boxes, log_scale=True) + l2_side_regularizer(boxes1, log_scale=True) + l2_side_regularizer(boxes2, log_scale=True)

    def nf4_loss(self, relation, scaling, boxes1, boxes2):
        trans_min = (boxes1.min_embed - relation)/(scaling + eps)
        trans_max = (boxes1.max_embed - relation)/(scaling + eps)
        trans_boxes = Box(trans_min, trans_max)
        return self.inclusion_loss(trans_boxes, boxes2), l2_side_regularizer(trans_boxes, log_scale=True) + l2_side_regularizer(boxes1, log_scale=True) + l2_side_regularizer(boxes2, log_scale=True)

    def disjoint_loss(self, boxes1, boxes2):
        log_intersection = torch.log(torch.clamp(self.volumes(self.intersection(boxes1, boxes2)), 1e-10, 1e4))
        log_boxes1 = torch.log(torch.clamp(self.volumes(boxes1), 1e-10, 1e4))
        log_boxes2 = torch.log(torch.clamp(self.volumes(boxes2), 1e-10, 1e4))
        union = log_boxes1 + log_boxes2
        return torch.exp(log_intersection-union), l2_side_regularizer(boxes1, log_scale=True) + l2_side_regularizer(boxes2, log_scale=True)

    def role_inclusion_loss(self, translation_1,translation_2,scaling_1,scaling_2):
        loss_1 = torch.norm(translation_1-translation_2, p=2, dim=1,keepdim=True)
        loss_2 = torch.norm(F.relu(scaling_1/(scaling_2 +eps) -1), p=2, dim=1,keepdim=True)
        return loss_1+loss_2

    def role_chain_loss(self, translation_1,translation_2,translation_3,scaling_1,scaling_2,scaling_3):
        loss_1 = torch.norm(scaling_1*translation_1 + translation_2 -translation_3, p=2, dim=1,keepdim=True)
        loss_2 = torch.norm(F.relu(scaling_1*scaling_2/(scaling_3 +eps) -1), p=2, dim=1,keepdim=True)
        return loss_1+loss_2

    def init_concept_embedding(self, vocab_size, embed_dim, init_value):
        distribution = uniform.Uniform(init_value[0], init_value[1])
        box_embed = distribution.sample((vocab_size, embed_dim))
        return box_embed