-
Notifications
You must be signed in to change notification settings - Fork 63
/
PositionEmbedding.py
180 lines (149 loc) · 7.77 KB
/
PositionEmbedding.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
# 1D, 2D, 3D
import math
import torch
from torch import nn
from wama_modules.utils import tensor2array, mat2gray
def PositionalEncoding_1D_learnable(embedding_dim=1024, token_num=600):
"""
return: [1, token_num, token_dim]
"""
return nn.Parameter(torch.zeros(1, token_num, embedding_dim))
def PositionalEncoding_1D_sincos(embedding_dim=1024, token_num=600, temperature=10000.):
"""
from the original transformer "attention is all you need"
:param embedding_dim:
:param token_num:
:param temperature:
:return: [1, token_num, token_dim]
from wama_modules.utils import show2D
pe = tensor2array(torch.squeeze(PositionalEncoding_1D_sincos(1002, 200, temperature=10000)))
show2D(pe)
pe = tensor2array(torch.squeeze(PositionalEncoding_1D_sincos(1002, 200, temperature=1000)))
show2D(pe)
"""
pe = torch.zeros(token_num, embedding_dim)
position = torch.arange(0, token_num).unsqueeze(1)
# e**(a*b) = (e**a)**b
div_term = torch.exp(torch.arange(0, embedding_dim, 2) * -(math.log(temperature) / embedding_dim)) # term 10000**（2i/d_model)
pe[:, 0::2] = torch.sin(position * div_term) # even number pe
pe[:, 1::2] = torch.cos(position * div_term) # odd number pe
pe = pe.unsqueeze(0)
return pe # [1, token_num, token_dim]
def PositionalEncoding_2D_learnable(embedding_dim=1024, token_shape=[100, 200], return_flatten=False):
"""
return:
if return_flatten is True, [1, token_shape[0]*token_shape[1], token_dim]
else, [1, token_dim, *token_shape]
# demo
print(PositionalEncoding_2D_learnable(1024,[100,300]).shape)
print(PositionalEncoding_2D_learnable(1024,[100,300],return_flatten = True).shape)
"""
if return_flatten:
return nn.Parameter(torch.zeros(1, embedding_dim, token_shape[0]*token_shape[1]))
else:
return nn.Parameter(torch.zeros(1, token_shape[0], token_shape[1], embedding_dim))
def PositionalEncoding_2D_sincos(embedding_dim=1024, token_shape = [100,200], temperature = 10000., return_flatten=False):
"""
from C-tran:
https://github.com/QData/C-Tran/blob/9278a693872a00dd7bdf8f6cdf2a599b83f01d16/models/position_enc.py#L59
which is the same as MAE or moco and stuff:
moco: https://github.com/facebookresearch/moco-v3/blob/38aae447f5e8f5a6221dbff0d8db40012512fe32/vits.py#L53
mae: https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
:param d_model: dimension of the model
:param height: height of the positions
:param width: width of the positions
:return:
if return_flatten is True, [1, token_shape[0]*token_shape[1], token_dim]
else, [1, token_dim, *token_shape]
from wama_modules.utils import show3D
print(PositionalEncoding_2D_sincos(1024, [200, 100]).shape)
pe = tensor2array(torch.squeeze(PositionalEncoding_2D_sincos(1024, [200, 200])))
show3D(pe)
pe = tensor2array(torch.squeeze(PositionalEncoding_2D_sincos(1024, [200, 200], temperature=1000)))
show3D(pe)
pe = tensor2array(torch.squeeze(
PositionalEncoding_2D_sincos(1024, [200, 200], temperature=1000, return_flatten=True).reshape(1,1024,200,200)
))
show3D(pe)
"""
if embedding_dim % 4 != 0:
raise ValueError("Cannot use sin/cos positional encoding with "
"odd dimension (got dim={:d})".format(embedding_dim))
height, width = token_shape
pe = torch.zeros(embedding_dim, height, width)
# Each dimension use half of d_model
or_embedding_dim = embedding_dim
embedding_dim = int(embedding_dim / 2)
div_term = torch.exp(torch.arange(0., embedding_dim, 2) * -(math.log(temperature) / embedding_dim))
pos_w = torch.arange(0., width).unsqueeze(1)
pos_h = torch.arange(0., height).unsqueeze(1)
pe[0:embedding_dim:2, :, :] = torch.sin(pos_w * div_term).transpose(0, 1).unsqueeze(1).repeat(1, height, 1)
pe[1:embedding_dim:2, :, :] = torch.cos(pos_w * div_term).transpose(0, 1).unsqueeze(1).repeat(1, height, 1)
pe[embedding_dim::2, :, :] = torch.sin(pos_h * div_term).transpose(0, 1).unsqueeze(2).repeat(1, 1, width)
pe[embedding_dim + 1::2, :, :] = torch.cos(pos_h * div_term).transpose(0, 1).unsqueeze(2).repeat(1, 1, width)
pe = pe.unsqueeze(0)
if return_flatten:
return pe.reshape(1, token_shape[0]*token_shape[1], or_embedding_dim)
else:
return pe
def PositionalEncoding_3D_learnable(embedding_dim=1024, token_shape=[20, 30, 40], return_flatten=False):
"""
return:
if return_flatten is True, [1, token_shape[0]*token_shape[1]*token_shape[2], token_dim]
else, [1, token_dim, *token_shape]
# demo
print(PositionalEncoding_3D_learnable(1024,[20, 30, 40]).shape)
print(PositionalEncoding_3D_learnable(1024,[20, 30, 40],return_flatten = True).shape)
"""
if return_flatten:
return nn.Parameter(torch.zeros(1, embedding_dim, token_shape[0]*token_shape[1]*token_shape[2]))
else:
return nn.Parameter(torch.zeros(1, token_shape[0], token_shape[1], token_shape[2], embedding_dim))
def PositionalEncoding_3D_sincos(embedding_dim=6*128, token_shape = [20,30,40], temperature = 10000., return_flatten=False):
"""
:param d_model: dimension of the model
:param height: height of the positions
:param width: width of the positions
:return:
if return_flatten is True, [1, token_shape[0]*token_shape[1]*token_shape[2], token_dim]
else, [1, token_dim, *token_shape]
# demo:
from wama_modules.utils import show3D
print(PositionalEncoding_3D_sincos(6*64, [50, 60, 70]).shape)
pe = tensor2array(torch.squeeze(PositionalEncoding_3D_sincos(6*64, [50, 60, 70])))
show3D(pe[:,:,:,35])
show3D(pe[:,:,30,:])
show3D(pe[:,25,:,:])
pe = tensor2array(torch.squeeze(PositionalEncoding_3D_sincos(6*64, [50, 60, 70], temperature=100)))
show3D(pe[:,:,:,35])
show3D(pe[:,:,30,:])
show3D(pe[:,25,:,:])
pe = tensor2array(torch.squeeze(PositionalEncoding_3D_sincos(6*64, [50, 60, 70], temperature=100, return_flatten=True).reshape(1,6*64,50,60,70)
))
show3D(pe[:,:,:,35])
show3D(pe[:,:,30,:])
show3D(pe[:,25,:,:])
"""
if embedding_dim % 6 != 0:
raise ValueError("Cannot use sin/cos positional encoding with "
"odd dimension (got dim={:d})".format(embedding_dim))
height, width, length = token_shape
pe = torch.zeros(embedding_dim, height, width, length)
# Each dimension use 1/3 of d_model
or_embedding_dim = embedding_dim
embedding_dim = int(embedding_dim / 3)
div_term = torch.exp(torch.arange(0., embedding_dim, 2) * -(math.log(temperature) / embedding_dim))
pos_w = torch.arange(0., width).unsqueeze(1)
pos_h = torch.arange(0., height).unsqueeze(1)
pos_l = torch.arange(0., length).unsqueeze(1)
pe[0:embedding_dim:2, :, :, :] = torch.sin(pos_w * div_term).transpose(0, 1).unsqueeze(1).unsqueeze(-1).repeat(1, height, 1,length)
pe[1:embedding_dim:2, :, :, :] = torch.cos(pos_w * div_term).transpose(0, 1).unsqueeze(1).unsqueeze(-1).repeat(1, height, 1,length)
pe[embedding_dim:embedding_dim*2:2, :, :, :] = torch.sin(pos_h * div_term).transpose(0, 1).unsqueeze(-1).unsqueeze(-1).repeat(1, 1, width,length)
pe[embedding_dim + 1:embedding_dim*2+1:2, :, :, :] = torch.cos(pos_h * div_term).transpose(0, 1).unsqueeze(-1).unsqueeze(-1).repeat(1, 1, width,length)
pe[embedding_dim*2::2, :, :, :] = torch.sin(pos_l * div_term).transpose(0, 1).unsqueeze(-2).unsqueeze(-2).repeat(1, height, width, 1)
pe[embedding_dim*2+1::2, :, :, :] = torch.cos(pos_l * div_term).transpose(0, 1).unsqueeze(-2).unsqueeze(-2).repeat(1, height, width, 1)
pe = pe.unsqueeze(0)
if return_flatten:
return pe.reshape(1, token_shape[0]*token_shape[1]*token_shape[2], or_embedding_dim)
else:
return pe