-
Notifications
You must be signed in to change notification settings - Fork 111
/
convnextv2_sparse.py
139 lines (121 loc) · 4.73 KB
/
convnextv2_sparse.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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn as nn
from timm.models.layers import trunc_normal_
from .utils import (
LayerNorm,
MinkowskiLayerNorm,
MinkowskiGRN,
MinkowskiDropPath
)
from MinkowskiEngine import (
MinkowskiConvolution,
MinkowskiDepthwiseConvolution,
MinkowskiLinear,
MinkowskiGELU
)
from MinkowskiOps import (
to_sparse,
)
class Block(nn.Module):
""" Sparse ConvNeXtV2 Block.
Args:
dim (int): Number of input channels.
drop_path (float): Stochastic depth rate. Default: 0.0
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
"""
def __init__(self, dim, drop_path=0., D=3):
super().__init__()
self.dwconv = MinkowskiDepthwiseConvolution(dim, kernel_size=7, bias=True, dimension=D)
self.norm = MinkowskiLayerNorm(dim, 1e-6)
self.pwconv1 = MinkowskiLinear(dim, 4 * dim)
self.act = MinkowskiGELU()
self.pwconv2 = MinkowskiLinear(4 * dim, dim)
self.grn = MinkowskiGRN(4 * dim)
self.drop_path = MinkowskiDropPath(drop_path)
def forward(self, x):
input = x
x = self.dwconv(x)
x = self.norm(x)
x = self.pwconv1(x)
x = self.act(x)
x = self.grn(x)
x = self.pwconv2(x)
x = input + self.drop_path(x)
return x
class SparseConvNeXtV2(nn.Module):
""" Sparse ConvNeXtV2.
Args:
in_chans (int): Number of input image channels. Default: 3
num_classes (int): Number of classes for classification head. Default: 1000
depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
drop_path_rate (float): Stochastic depth rate. Default: 0.
head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
"""
def __init__(self,
in_chans=3,
num_classes=1000,
depths=[3, 3, 9, 3],
dims=[96, 192, 384, 768],
drop_path_rate=0.,
D=3):
super().__init__()
self.depths = depths
self.num_classes = num_classes
self.downsample_layers = nn.ModuleList() # stem and 3 intermediate downsampling conv layers
stem = nn.Sequential(
nn.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),
LayerNorm(dims[0], eps=1e-6, data_format="channels_first")
)
self.downsample_layers.append(stem)
for i in range(3):
downsample_layer = nn.Sequential(
MinkowskiLayerNorm(dims[i], eps=1e-6),
MinkowskiConvolution(dims[i], dims[i+1], kernel_size=2, stride=2, bias=True, dimension=D)
)
self.downsample_layers.append(downsample_layer)
self.stages = nn.ModuleList() # 4 feature resolution stages, each consisting of multiple residual blocks
dp_rates=[x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
cur = 0
for i in range(4):
stage = nn.Sequential(
*[Block(dim=dims[i], drop_path=dp_rates[cur + j], D=D) for j in range(depths[i])]
)
self.stages.append(stage)
cur += depths[i]
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, MinkowskiConvolution):
trunc_normal_(m.kernel, std=.02)
nn.init.constant_(m.bias, 0)
if isinstance(m, MinkowskiDepthwiseConvolution):
trunc_normal_(m.kernel, std=.02)
nn.init.constant_(m.bias, 0)
if isinstance(m, MinkowskiLinear):
trunc_normal_(m.linear.weight, std=.02)
nn.init.constant_(m.linear.bias, 0)
def upsample_mask(self, mask, scale):
assert len(mask.shape) == 2
p = int(mask.shape[1] ** .5)
return mask.reshape(-1, p, p).\
repeat_interleave(scale, axis=1).\
repeat_interleave(scale, axis=2)
def forward(self, x, mask):
num_stages = len(self.stages)
mask = self.upsample_mask(mask, 2**(num_stages-1))
mask = mask.unsqueeze(1).type_as(x)
# patch embedding
x = self.downsample_layers[0](x)
x *= (1.-mask)
# sparse encoding
x = to_sparse(x)
for i in range(4):
x = self.downsample_layers[i](x) if i > 0 else x
x = self.stages[i](x)
# densify
x = x.dense()[0]
return x