-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtransformer_block.py
142 lines (126 loc) · 5.78 KB
/
transformer_block.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
# Copyright (c) [2012]-[2021] Shanghai Yitu Technology Co., Ltd.
#
# This source code is licensed under the Clear BSD License
# LICENSE file in the root directory of this file
# All rights reserved.
"""
Borrow from timm(https://github.com/rwightman/pytorch-image-models)
"""
import torch
import torch.nn as nn
import numpy as np
import os
from timm.models.layers import DropPath
class Mlp(nn.Module):
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class Attention(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.count = 0
def forward(self, x):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
attn = torch.matmul(q,k.transpose(-2, -1)) * self.scale
# if self.count == 0:
# for param in self.qkv.parameters():
# print(param)
# print('-----------------------------------------')
# print('the number of 0 of q',torch.sum(torch.sum(torch.eq(q,torch.zeros_like(q)))))
# print('the number of 0 of k',torch.sum(torch.sum(torch.eq(k,torch.zeros_like(k)))))
# print('the number of 0 of attn',torch.sum(torch.sum(torch.eq(attn,torch.zeros_like(attn)))))
# if(os.path.exists('/home/chenguangyan/data/q.pth') == False):
# torch.save(x,'/home/chenguangyan/data/attnx.pth')
# torch.save(q,'/home/chenguangyan/data/q.pth')
# torch.save(k,'/home/chenguangyan/data/k.pth')
# count = 0
# for param in self.qkv.parameters():
# torch.save(param,'/home/chenguangyan/data/param-%d.pth'%count)
# count += 1
# print('q',q)
# print('k.shape',q.shape)
# print('k',k)
# print('attn',attn)
# self.count += 1
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = torch.matmul(attn,v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
def forward(self, x):
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 0',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 0',torch.all(torch.eq(judge_1,judge)))
x = self.norm1(x)
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 1',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 1',torch.all(torch.eq(judge_1,judge)))
x = self.attn(x)
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 2',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 2',torch.all(torch.eq(judge_1,judge)))
x = x + self.drop_path(x)
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 3',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 3',torch.all(torch.eq(judge_1,judge)))
x = self.norm2(x)
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 4',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 4',torch.all(torch.eq(judge_1,judge)))
x = self.mlp(x)
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 5',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 5',torch.all(torch.eq(judge_1,judge)))
x = x + self.drop_path(x)
# judge = torch.isnan(x)
# judge_1 = torch.zeros_like(x)
# print('[-3] - 6',torch.all(torch.eq(judge_1,x))==False)
# print('[-3] - 6',torch.all(torch.eq(judge_1,judge)))
return x
def get_sinusoid_encoding(n_position, d_hid):
''' Sinusoid position encoding table '''
def get_position_angle_vec(position):
return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)]
sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1
return torch.FloatTensor(sinusoid_table).unsqueeze(0)