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
|
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"%load_ext autoreload\n",
"%autoreload 2\n",
"\n",
"import cv2\n",
"%matplotlib inline\n",
"import matplotlib.pyplot as plt\n",
"import numpy as np\n",
"import torch\n",
"from torch import nn\n",
"from importlib.util import find_spec\n",
"if find_spec(\"text_recognizer\") is None:\n",
" import sys\n",
" sys.path.append('..')\n",
"\n",
"from text_recognizer.networks.transformer.attention import MultiHeadAttention"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"temp_mha = MultiHeadAttention(hidden_dim=512, num_heads=8)\n",
"def print_out(Q, K, V):\n",
" temp_out, temp_attn = temp_mha.scaled_dot_product_attention(Q, K, V)\n",
" print('Attention weights are:', temp_attn.squeeze())\n",
" print('Output is:', temp_out.squeeze())"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"test_K = torch.tensor(\n",
" [[10, 0, 0],\n",
" [ 0,10, 0],\n",
" [ 0, 0,10],\n",
" [ 0, 0,10]]\n",
").float()[None,None]\n",
"\n",
"test_V = torch.tensor(\n",
" [[ 1,0,0],\n",
" [ 10,0,0],\n",
" [ 100,5,0],\n",
" [1000,6,0]]\n",
").float()[None,None]\n",
"\n",
"test_Q = torch.tensor(\n",
" [[0, 10, 0]]\n",
").float()[None,None]\n"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Attention weights are: tensor([8.4333e-26, 1.0000e+00, 8.4333e-26, 8.4333e-26])\n",
"Output is: tensor([1.0000e+01, 9.2766e-25, 0.0000e+00])\n"
]
}
],
"source": [
"print_out(test_Q, test_K, test_V)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Attends to the second element, as it should!"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Attention weights are: tensor([4.2166e-26, 4.2166e-26, 5.0000e-01, 5.0000e-01])\n",
"Output is: tensor([550.0000, 5.5000, 0.0000])\n"
]
}
],
"source": [
"test_Q = torch.tensor([[0, 0, 10]]).float()[None,None]\n",
"print_out(test_Q, test_K, test_V)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Focuses equally on the third and fourth key."
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Attention weights are: tensor([[4.2166e-26, 4.2166e-26, 5.0000e-01, 5.0000e-01],\n",
" [8.4333e-26, 1.0000e+00, 8.4333e-26, 8.4333e-26],\n",
" [5.0000e-01, 5.0000e-01, 4.2166e-26, 4.2166e-26]])\n",
"Output is: tensor([[5.5000e+02, 5.5000e+00, 0.0000e+00],\n",
" [1.0000e+01, 9.2766e-25, 0.0000e+00],\n",
" [5.5000e+00, 4.6383e-25, 0.0000e+00]])\n"
]
}
],
"source": [
"test_Q = torch.tensor(\n",
" [[0, 0, 10], [0, 10, 0], [10, 10, 0]]\n",
").float()[None,None]\n",
"print_out(test_Q, test_K, test_V)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.9.7"
}
},
"nbformat": 4,
"nbformat_minor": 4
}
|
