-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
224 lines (196 loc) · 10 KB
/
model.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
import torch
import torch.nn as nn
from spikingjelly.clock_driven.neuron import MultiStepParametricLIFNode as MultiStepLIFNode
from timm.models.layers import to_2tuple, trunc_normal_, DropPath
from timm.models.registry import register_model
from timm.models.vision_transformer import _cfg
import torch.nn.functional as F
from functools import partial
__all__ = ['AutoST']
class MLP(nn.Module):
def __init__(self, in_features, hidden_features=None, out_features=None, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1_linear = nn.Linear(in_features, hidden_features)
self.fc1_bn = nn.BatchNorm1d(hidden_features)
self.fc1_lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.fc2_linear = nn.Linear(hidden_features, out_features)
self.fc2_bn = nn.BatchNorm1d(out_features)
self.fc2_lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.c_hidden = hidden_features
self.c_output = out_features
def forward(self, x):
T, B, N, C = x.shape
x = self.fc1_lif(x)
x = x.flatten(0, 1)
x = self.fc1_linear(x)
x = self.fc1_bn(x.transpose(-1, -2)).transpose(-1, -2).reshape(T, B, N, self.c_hidden).contiguous()
x = self.fc2_lif(x)
x = self.fc2_linear(x.flatten(0, 1))
x = self.fc2_bn(x.transpose(-1, -2)).transpose(-1, -2).reshape(T, B, N, C).contiguous()
return x
class SSA(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
super().__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
self.scale = 0.125
self.q_linear = nn.Linear(dim, dim)
self.q_bn = nn.BatchNorm1d(dim)
self.q_lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.k_linear = nn.Linear(dim, dim)
self.k_bn = nn.BatchNorm1d(dim)
self.k_lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.v_linear = nn.Linear(dim, dim)
self.v_bn = nn.BatchNorm1d(dim)
self.v_lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.attn_lif = MultiStepLIFNode(v_threshold=0.5, detach_reset=True, backend='cupy')
self.proj_linear = nn.Linear(dim, dim)
self.proj_bn = nn.BatchNorm1d(dim)
self.lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
def forward(self, x):
T, B, N, C = x.shape
x = self.lif(x)
x_for_qkv = x.flatten(0, 1) # TB, N, C
q_linear_out = self.q_linear(x_for_qkv) # [TB, N, C]
q_linear_out = self.q_bn(q_linear_out.transpose(-1, -2)).transpose(-1, -2).reshape(T, B, N, C).contiguous()
q_linear_out = self.q_lif(q_linear_out)
q = q_linear_out.reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4).contiguous()
k_linear_out = self.k_linear(x_for_qkv)
k_linear_out = self.k_bn(k_linear_out.transpose(-1, -2)).transpose(-1, -2).reshape(T, B, N, C).contiguous()
k_linear_out = self.k_lif(k_linear_out)
k = k_linear_out.reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4).contiguous()
v_linear_out = self.v_linear(x_for_qkv)
v_linear_out = self.v_bn(v_linear_out.transpose(-1, -2)).transpose(-1, -2).reshape(T, B, N, C).contiguous()
v_linear_out = self.v_lif(v_linear_out)
v = v_linear_out.reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4).contiguous()
attn = (q @ k.transpose(-2, -1)) * self.scale
x = attn @ v
x = x.transpose(2, 3).reshape(T, B, N, C).contiguous()
x = self.attn_lif(x)
x = x.flatten(0, 1)
x = self.proj_bn(self.proj_linear(x).transpose(-1, -2)).transpose(-1, -2).reshape(T, B, N, C).contiguous()
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., norm_layer=nn.LayerNorm, sr_ratio=1):
super().__init__()
# self.norm1 = norm_layer(dim)
self.attn = SSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
# self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = MLP(in_features=dim, hidden_features=mlp_hidden_dim, drop=drop)
def forward(self, x):
x = x + self.attn(x)
x = x + self.mlp(x)
return x
class SPS(nn.Module):
def __init__(self, img_size_h=128, img_size_w=128, patch_size=4, in_channels=2, embed_dims=256):
super().__init__()
self.image_size = [img_size_h, img_size_w]
patch_size = to_2tuple(patch_size)
self.patch_size = patch_size
self.C = in_channels
self.H, self.W = self.image_size[0] // patch_size[0], self.image_size[1] // patch_size[1]
self.num_patches = self.H * self.W
self.proj_conv = nn.Conv2d(in_channels, embed_dims // 8, kernel_size=3, stride=1, padding=1, bias=False)
self.proj_bn = nn.BatchNorm2d(embed_dims // 8)
self.proj_lif = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.proj_conv1 = nn.Conv2d(embed_dims // 8, embed_dims // 4, kernel_size=3, stride=1, padding=1, bias=False)
self.proj_bn1 = nn.BatchNorm2d(embed_dims // 4)
self.proj_lif1 = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.proj_conv2 = nn.Conv2d(embed_dims // 4, embed_dims // 2, kernel_size=3, stride=1, padding=1, bias=False)
self.proj_bn2 = nn.BatchNorm2d(embed_dims // 2)
self.proj_lif2 = MultiStepLIFNode(detach_reset=True, backend='cupy')
self.maxpool2 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
self.proj_conv3 = nn.Conv2d(embed_dims // 2, embed_dims, kernel_size=3, stride=1, padding=1, bias=False)
self.proj_bn3 = nn.BatchNorm2d(embed_dims)
self.maxpool3 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
self.rpe_conv = nn.Conv2d(embed_dims, embed_dims, kernel_size=3, stride=1, padding=1, bias=False)
self.rpe_bn = nn.BatchNorm2d(embed_dims)
def forward(self, x):
T, B, C, H, W = x.shape
x = self.proj_conv(x.flatten(0, 1)) # have some fire value
x = self.proj_bn(x).reshape(T, B, -1, H, W).contiguous()
x = self.proj_lif(x).flatten(0, 1).contiguous()
x = self.proj_conv1(x)
x = self.proj_bn1(x).reshape(T, B, -1, H, W).contiguous()
x = self.proj_lif1(x).flatten(0, 1).contiguous()
x = self.proj_conv2(x)
x = self.proj_bn2(x).reshape(T, B, -1, H, W).contiguous()
x = self.proj_lif2(x).flatten(0, 1).contiguous()
x = self.maxpool2(x)
x = self.proj_conv3(x)
x = self.proj_bn3(x).reshape(T, B, -1, H // 2, W // 2).contiguous()
x = self.maxpool3(x.flatten(0, 1))
x_feat = x.reshape(T, B, -1, H // 4, W // 4).contiguous()
x = self.rpe_conv(x)
x = self.rpe_bn(x).reshape(T, B, -1, H // 4, W // 4).contiguous()
x = x + x_feat
x = x.flatten(-2).transpose(-1, -2) # T,B,N,C
return x
class AutoST(nn.Module):
def __init__(self,
img_size_h=128, img_size_w=128, patch_size=16, in_channels=2, num_classes=11,
embed_dims=[64, 128], num_heads=[1, 2], mlp_ratios=[4, 4], qkv_bias=False, qk_scale=None,
drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
depths=[6, 8], sr_ratios=[8, 4], T=4
):
super().__init__()
self.T = T # time step
self.num_classes = num_classes
self.depths = depths
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depths)] # stochastic depth decay rule
patch_embed = SPS(img_size_h=img_size_h,
img_size_w=img_size_w,
patch_size=patch_size,
in_channels=in_channels,
embed_dims=embed_dims)
block = nn.ModuleList([Block(
dim=embed_dims, num_heads=num_heads[j], mlp_ratio=mlp_ratios[j], qkv_bias=qkv_bias,
qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[j],
norm_layer=norm_layer, sr_ratio=sr_ratios)
for j in range(depths)])
setattr(self, f"patch_embed", patch_embed)
setattr(self, f"block", block)
# classification head
self.head = nn.Linear(embed_dims, num_classes) if num_classes > 0 else nn.Identity()
self.apply(self._init_weights)
@torch.jit.ignore
def _get_pos_embed(self, pos_embed, patch_embed, H, W):
if H * W == self.patch_embed1.num_patches:
return pos_embed
else:
return F.interpolate(
pos_embed.reshape(1, patch_embed.H, patch_embed.W, -1).permute(0, 3, 1, 2),
size=(H, W), mode="bilinear").reshape(1, -1, H * W).permute(0, 2, 1)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward_features(self, x):
block = getattr(self, f"block")
patch_embed = getattr(self, f"patch_embed")
x = patch_embed(x)
for blk in block:
x = blk(x)
return x.mean(2)
def forward(self, x):
x = (x.unsqueeze(0)).repeat(self.T, 1, 1, 1, 1)
x = self.forward_features(x)
x = self.head(x.mean(0))
return x
@register_model
def AutoST(pretrained=False, **kwargs):
model = AutoST(
**kwargs
)
model.default_cfg = _cfg()
return model