/
HAT_Net.py
230 lines (192 loc) · 8.58 KB
/
HAT_Net.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
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.layers import DropPath, trunc_normal_
from timm.models.registry import register_model
from timm.models.vision_transformer import _cfg
class InvertedResidual(nn.Module):
def __init__(self, in_dim, hidden_dim=None, out_dim=None, kernel_size=3,
drop=0., act_layer=nn.SiLU):
super().__init__()
hidden_dim = hidden_dim or in_dim
out_dim = out_dim or in_dim
pad = (kernel_size - 1) // 2
self.conv1 = nn.Sequential(
nn.GroupNorm(1, in_dim, eps=1e-6),
nn.Conv2d(in_dim, hidden_dim, 1, bias=False),
act_layer(inplace=True)
)
self.conv2 = nn.Sequential(
nn.Conv2d(hidden_dim, hidden_dim, kernel_size, padding=pad, groups=hidden_dim, bias=False),
act_layer(inplace=True)
)
self.conv3 = nn.Sequential(
nn.Conv2d(hidden_dim, out_dim, 1, bias=False),
nn.GroupNorm(1, out_dim, eps=1e-6)
)
self.drop = nn.Dropout2d(drop, inplace=True)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.drop(x)
x = self.conv3(x)
x = self.drop(x)
return x
class Attention(nn.Module):
def __init__(self, dim, head_dim, grid_size=1, ds_ratio=1, drop=0.):
super().__init__()
assert dim % head_dim == 0
self.num_heads = dim // head_dim
self.head_dim = head_dim
self.scale = self.head_dim ** -0.5
self.grid_size = grid_size
self.norm = nn.GroupNorm(1, dim, eps=1e-6)
self.qkv = nn.Conv2d(dim, dim * 3, 1)
self.proj = nn.Conv2d(dim, dim, 1)
self.proj_norm = nn.GroupNorm(1, dim, eps=1e-6)
self.drop = nn.Dropout2d(drop, inplace=True)
if grid_size > 1:
self.grid_norm = nn.GroupNorm(1, dim, eps=1e-6)
self.avg_pool = nn.AvgPool2d(ds_ratio, stride=ds_ratio)
self.ds_norm = nn.GroupNorm(1, dim, eps=1e-6)
self.q = nn.Conv2d(dim, dim, 1)
self.kv = nn.Conv2d(dim, dim * 2, 1)
def forward(self, x):
B, C, H, W = x.shape
qkv = self.qkv(self.norm(x))
if self.grid_size > 1:
grid_h, grid_w = H // self.grid_size, W // self.grid_size
qkv = qkv.reshape(B, 3, self.num_heads, self.head_dim, grid_h,
self.grid_size, grid_w, self.grid_size)
qkv = qkv.permute(1, 0, 2, 4, 6, 5, 7, 3)
qkv = qkv.reshape(3, -1, self.grid_size * self.grid_size, self.head_dim)
q, k, v = qkv[0], qkv[1], qkv[2]
attn = (q * self.scale) @ k.transpose(-2, -1)
attn = attn.softmax(dim=-1)
grid_x = (attn @ v).reshape(B, self.num_heads, grid_h, grid_w,
self.grid_size, self.grid_size, self.head_dim)
grid_x = grid_x.permute(0, 1, 6, 2, 4, 3, 5).reshape(B, C, H, W)
grid_x = self.grid_norm(x + grid_x)
q = self.q(grid_x).reshape(B, self.num_heads, self.head_dim, -1)
q = q.transpose(-2, -1)
kv = self.kv(self.ds_norm(self.avg_pool(grid_x)))
kv = kv.reshape(B, 2, self.num_heads, self.head_dim, -1)
kv = kv.permute(1, 0, 2, 4, 3)
k, v = kv[0], kv[1]
else:
qkv = qkv.reshape(B, 3, self.num_heads, self.head_dim, -1)
qkv = qkv.permute(1, 0, 2, 4, 3)
q, k, v = qkv[0], qkv[1], qkv[2]
attn = (q * self.scale) @ k.transpose(-2, -1)
attn = attn.softmax(dim=-1)
global_x = (attn @ v).transpose(-2, -1).reshape(B, C, H, W)
if self.grid_size > 1:
global_x = global_x + grid_x
x = self.drop(self.proj(global_x))
return x
class Block(nn.Module):
def __init__(self, dim, head_dim, grid_size=1, ds_ratio=1, expansion=4,
drop=0., drop_path=0., kernel_size=3, act_layer=nn.SiLU):
super().__init__()
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.attn = Attention(dim, head_dim, grid_size=grid_size, ds_ratio=ds_ratio, drop=drop)
self.conv = InvertedResidual(dim, hidden_dim=dim * expansion, out_dim=dim,
kernel_size=kernel_size, drop=drop, act_layer=act_layer)
def forward(self, x):
x = x + self.drop_path(self.attn(x))
x = x + self.drop_path(self.conv(x))
return x
class Downsample(nn.Module):
def __init__(self, in_dim, out_dim, kernel_size=3):
super().__init__()
self.conv = nn.Conv2d(in_dim, out_dim, kernel_size, padding=1, stride=2)
self.norm = nn.GroupNorm(1, out_dim, eps=1e-6)
def forward(self, x):
x = self.norm(self.conv(x))
return x
class HAT_Net(nn.Module):
def __init__(self, img_size=224, in_chans=3, num_classes=1000, dims=[64, 128, 256, 512],
head_dim=64, expansions=[4, 4, 6, 6], grid_sizes=[1, 1, 1, 1],
ds_ratios=[8, 4, 2, 1], depths=[3, 4, 8, 3], drop_rate=0.,
drop_path_rate=0., act_layer=nn.SiLU, kernel_sizes=[3, 3, 3, 3]):
super().__init__()
self.depths = depths
self.patch_embed = nn.Sequential(
nn.Conv2d(3, 16, 3, padding=1, stride=2),
nn.GroupNorm(1, 16, eps=1e-6),
act_layer(inplace=True),
nn.Conv2d(16, dims[0], 3, padding=1, stride=2),
)
self.blocks = []
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
for stage in range(len(dims)):
self.blocks.append(nn.ModuleList([Block(
dims[stage], head_dim, grid_size=grid_sizes[stage], ds_ratio=ds_ratios[stage],
expansion=expansions[stage], drop=drop_rate, drop_path=dpr[sum(depths[:stage]) + i],
kernel_size=kernel_sizes[stage], act_layer=act_layer)
for i in range(depths[stage])]))
self.blocks = nn.ModuleList(self.blocks)
self.ds2 = Downsample(dims[0], dims[1])
self.ds3 = Downsample(dims[1], dims[2])
self.ds4 = Downsample(dims[2], dims[3])
self.classifier = nn.Sequential(
nn.Dropout(0.2, inplace=True),
nn.Linear(dims[-1], num_classes),
)
# init weights
self.apply(self._init_weights)
def reset_drop_path(self, drop_path_rate):
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
cur = 0
for stage in range(len(self.blocks)):
for idx in range(self.depths[stage]):
self.blocks[stage][idx].drop_path.drop_prob = dpr[cur + idx]
cur += self.depths[stage]
def _init_weights(self, m):
if isinstance(m, (nn.Linear, nn.Conv2d)):
trunc_normal_(m.weight, std=.02)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, (nn.LayerNorm, nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward(self, x):
x = self.patch_embed(x)
for block in self.blocks[0]:
x = block(x)
x = self.ds2(x)
for block in self.blocks[1]:
x = block(x)
x = self.ds3(x)
for block in self.blocks[2]:
x = block(x)
x = self.ds4(x)
for block in self.blocks[3]:
x = block(x)
x = F.adaptive_avg_pool2d(x, (1, 1)).flatten(1)
x = self.classifier(x)
return x
@register_model
def HAT_Net_tiny(pretrained=False, **kwargs):
model = HAT_Net(dims=[48, 96, 240, 384], head_dim=48, expansions=[8, 8, 4, 4],
grid_sizes=[8, 7, 7, 1], ds_ratios=[8, 4, 2, 1], depths=[2, 2, 6, 3], **kwargs)
model.default_cfg = _cfg()
return model
@register_model
def HAT_Net_small(pretrained=False, **kwargs):
model = HAT_Net(dims=[64, 128, 320, 512], head_dim=64, expansions=[8, 8, 4, 4],
grid_sizes=[8, 7, 7, 1], ds_ratios=[8, 4, 2, 1], depths=[2, 3, 8, 3], **kwargs)
model.default_cfg = _cfg()
return model
@register_model
def HAT_Net_medium(pretrained=False, **kwargs):
model = HAT_Net(dims=[64, 128, 320, 512], head_dim=64, expansions=[8, 8, 4, 4],
grid_sizes=[8, 7, 7, 1], ds_ratios=[8, 4, 2, 1], depths=[3, 6, 18, 3], kernel_sizes=[5, 3, 5, 3], **kwargs)
model.default_cfg = _cfg()
return model
@register_model
def HAT_Net_large(pretrained=False, **kwargs):
model = HAT_Net(dims=[64, 128, 320, 640], head_dim=64, expansions=[8, 8, 4, 4],
grid_sizes=[8, 7, 7, 1], ds_ratios=[8, 4, 2, 1], depths=[3, 8, 27, 3], kernel_sizes=[3, 3, 3, 3], **kwargs)
model.default_cfg = _cfg()
return model