-
Notifications
You must be signed in to change notification settings - Fork 0
/
convNext_code.py
130 lines (93 loc) · 3.38 KB
/
convNext_code.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
import torch
import torch.nn as nn
import math
class InvertedResidualBlock(nn.Module):
def __init__(self,
in_channels,
out_channels,
):
super().__init__() # Just have to do this for all nn.Module classes
#Depthwise Convolution
self.spatial_mixing = nn.Sequential(
nn.Conv2d(in_channels, in_channels, kernel_size=7, padding =3,
stride = 1, groups = in_channels, bias=False),
nn.BatchNorm2d(in_channels),
)
# Expand Ratio is like 4, so hidden_dim >> in_channels
hidden_dim = in_channels * 4
#Pointwise Convolution
self.feature_mixing = nn.Sequential(
nn.Conv2d(in_channels, hidden_dim, kernel_size=1, padding =0,
stride = 1, bias=False),
nn.GELU(),
)
self.bottleneck_channels = nn.Sequential(
nn.Conv2d(hidden_dim,out_channels,kernel_size=1,stride=1,padding=0,bias=False),
)
def forward(self, x):
out = self.spatial_mixing(x)
out = self.feature_mixing(out)
out = self.bottleneck_channels(out)
return x + out
class ConvNext(nn.Module):
def __init__(self, num_classes= 7000):
super().__init__()
self.num_classes = num_classes
"""
First couple of layers are special, just do them here.
This is called the "stem". Usually, methods use it to downsample or twice.
"""
self.stem = nn.Sequential(
nn.Conv2d(3, 96, kernel_size=4, stride=4),
nn.BatchNorm2d(96),
)
self.stage_cfgs = [
# expand_ratio, channels, # blocks, stride of first block
[4, 96, 3, 1],
[4, 192, 3, 1],
[4, 384, 9, 1],
[4, 768, 3, 1],
]
in_channels = 96
layers = []
#BLOCK TYPE 1 - 3 TIMES
for i in range(3):
layers.append(InvertedResidualBlock(
in_channels=96,
out_channels=96))
layers.append(nn.BatchNorm2d(96))
layers.append(nn.Conv2d(96,192,kernel_size=2,stride=2))
#BLOCK TYPE 2 - 3 TIMES
for i in range(3):
layers.append(InvertedResidualBlock(
in_channels=192,
out_channels=192))
layers.append(nn.BatchNorm2d(192))
layers.append(nn.Conv2d(192,384,kernel_size=2,stride=2))
#BLOCK TYPE 3 - 9 TIMES
for i in range(9):
layers.append(InvertedResidualBlock(
in_channels=384,
out_channels=384))
layers.append(nn.BatchNorm2d(384))
layers.append(nn.Conv2d(384,768,kernel_size=2,stride=2))
#BLOCK TYPE 4 - 3 TIMES
for i in range(3):
layers.append(InvertedResidualBlock(
in_channels=768,
out_channels=768))
self.layers = nn.Sequential(*layers)
self.mid_cls_layer = nn.Sequential(
nn.AdaptiveAvgPool2d((1,1)),
nn.Flatten(),
)
self.final_cls_layer = nn.Sequential(nn.Linear(768,num_classes),)
def forward(self, x,return_feats=False):
out = self.stem(x)
out = self.layers(out)
feats = self.mid_cls_layer(out)
out = self.final_cls_layer(feats)
if return_feats:
return feats
else:
return out