Upload 4 files
Browse files- models/__init__.py +10 -0
- models/convnext.py +220 -0
- models/flexible_unet.py +312 -0
- models/flexible_unet_convnext.py +447 -0
models/__init__.py
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
"""
|
4 |
+
Created on Sun Mar 20 14:23:55 2022
|
5 |
+
|
6 |
+
@author: jma
|
7 |
+
"""
|
8 |
+
|
9 |
+
#from .unetr2d import UNETR2D
|
10 |
+
#from .swin_unetr import SwinUNETR
|
models/convnext.py
ADDED
@@ -0,0 +1,220 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
|
3 |
+
# All rights reserved.
|
4 |
+
|
5 |
+
# This source code is licensed under the license found in the
|
6 |
+
# LICENSE file in the root directory of this source tree.
|
7 |
+
|
8 |
+
from functools import partial
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import torch.nn.functional as F
|
12 |
+
from timm.models.layers import trunc_normal_, DropPath
|
13 |
+
from timm.models.registry import register_model
|
14 |
+
from monai.networks.layers.factories import Act, Conv, Pad, Pool
|
15 |
+
from monai.networks.layers.utils import get_norm_layer
|
16 |
+
from monai.utils.module import look_up_option
|
17 |
+
from typing import List, NamedTuple, Optional, Tuple, Type, Union
|
18 |
+
class Block(nn.Module):
|
19 |
+
r""" ConvNeXt Block. There are two equivalent implementations:
|
20 |
+
(1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
|
21 |
+
(2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
|
22 |
+
We use (2) as we find it slightly faster in PyTorch
|
23 |
+
|
24 |
+
Args:
|
25 |
+
dim (int): Number of input channels.
|
26 |
+
drop_path (float): Stochastic depth rate. Default: 0.0
|
27 |
+
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
28 |
+
"""
|
29 |
+
def __init__(self, dim, drop_path=0., layer_scale_init_value=1e-6):
|
30 |
+
super().__init__()
|
31 |
+
self.dwconv = nn.Conv2d(dim, dim, kernel_size=7, padding=3, groups=dim) # depthwise conv
|
32 |
+
self.norm = LayerNorm(dim, eps=1e-6)
|
33 |
+
self.pwconv1 = nn.Linear(dim, 4 * dim) # pointwise/1x1 convs, implemented with linear layers
|
34 |
+
self.act = nn.GELU()
|
35 |
+
self.pwconv2 = nn.Linear(4 * dim, dim)
|
36 |
+
self.gamma = nn.Parameter(layer_scale_init_value * torch.ones((dim)),
|
37 |
+
requires_grad=True) if layer_scale_init_value > 0 else None
|
38 |
+
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
39 |
+
|
40 |
+
def forward(self, x):
|
41 |
+
input = x
|
42 |
+
x = self.dwconv(x)
|
43 |
+
x = x.permute(0, 2, 3, 1) # (N, C, H, W) -> (N, H, W, C)
|
44 |
+
x = self.norm(x)
|
45 |
+
x = self.pwconv1(x)
|
46 |
+
x = self.act(x)
|
47 |
+
x = self.pwconv2(x)
|
48 |
+
if self.gamma is not None:
|
49 |
+
x = self.gamma * x
|
50 |
+
x = x.permute(0, 3, 1, 2) # (N, H, W, C) -> (N, C, H, W)
|
51 |
+
|
52 |
+
x = input + self.drop_path(x)
|
53 |
+
return x
|
54 |
+
|
55 |
+
class ConvNeXt(nn.Module):
|
56 |
+
r""" ConvNeXt
|
57 |
+
A PyTorch impl of : `A ConvNet for the 2020s` -
|
58 |
+
https://arxiv.org/pdf/2201.03545.pdf
|
59 |
+
|
60 |
+
Args:
|
61 |
+
in_chans (int): Number of input image channels. Default: 3
|
62 |
+
num_classes (int): Number of classes for classification head. Default: 1000
|
63 |
+
depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
|
64 |
+
dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
|
65 |
+
drop_path_rate (float): Stochastic depth rate. Default: 0.
|
66 |
+
layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
|
67 |
+
head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
|
68 |
+
"""
|
69 |
+
def __init__(self, in_chans=3, num_classes=21841,
|
70 |
+
depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], drop_path_rate=0.,
|
71 |
+
layer_scale_init_value=1e-6, head_init_scale=1., out_indices=[0, 1, 2, 3],
|
72 |
+
):
|
73 |
+
super().__init__()
|
74 |
+
# conv_type: Type[Union[nn.Conv1d, nn.Conv2d, nn.Conv3d]] = Conv["conv", 2]
|
75 |
+
# self._conv_stem = conv_type(self.in_channels, self.in_channels, kernel_size=3, stride=stride, bias=False)
|
76 |
+
# self._conv_stem_padding = _make_same_padder(self._conv_stem, current_image_size)
|
77 |
+
|
78 |
+
self.downsample_layers = nn.ModuleList() # stem and 3 intermediate downsampling conv layers
|
79 |
+
stem = nn.Sequential(
|
80 |
+
nn.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),
|
81 |
+
LayerNorm(dims[0], eps=1e-6, data_format="channels_first")
|
82 |
+
)
|
83 |
+
self.downsample_layers.append(stem)
|
84 |
+
for i in range(3):
|
85 |
+
downsample_layer = nn.Sequential(
|
86 |
+
LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),
|
87 |
+
nn.Conv2d(dims[i], dims[i+1], kernel_size=2, stride=2),
|
88 |
+
)
|
89 |
+
self.downsample_layers.append(downsample_layer)
|
90 |
+
|
91 |
+
self.stages = nn.ModuleList() # 4 feature resolution stages, each consisting of multiple residual blocks
|
92 |
+
dp_rates=[x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
|
93 |
+
cur = 0
|
94 |
+
for i in range(4):
|
95 |
+
stage = nn.Sequential(
|
96 |
+
*[Block(dim=dims[i], drop_path=dp_rates[cur + j],
|
97 |
+
layer_scale_init_value=layer_scale_init_value) for j in range(depths[i])]
|
98 |
+
)
|
99 |
+
self.stages.append(stage)
|
100 |
+
cur += depths[i]
|
101 |
+
|
102 |
+
|
103 |
+
self.out_indices = out_indices
|
104 |
+
|
105 |
+
norm_layer = partial(LayerNorm, eps=1e-6, data_format="channels_first")
|
106 |
+
for i_layer in range(4):
|
107 |
+
layer = norm_layer(dims[i_layer])
|
108 |
+
layer_name = f'norm{i_layer}'
|
109 |
+
self.add_module(layer_name, layer)
|
110 |
+
self.apply(self._init_weights)
|
111 |
+
|
112 |
+
|
113 |
+
def _init_weights(self, m):
|
114 |
+
if isinstance(m, (nn.Conv2d, nn.Linear)):
|
115 |
+
trunc_normal_(m.weight, std=.02)
|
116 |
+
nn.init.constant_(m.bias, 0)
|
117 |
+
|
118 |
+
def forward_features(self, x):
|
119 |
+
outs = []
|
120 |
+
|
121 |
+
for i in range(4):
|
122 |
+
x = self.downsample_layers[i](x)
|
123 |
+
x = self.stages[i](x)
|
124 |
+
if i in self.out_indices:
|
125 |
+
norm_layer = getattr(self, f'norm{i}')
|
126 |
+
x_out = norm_layer(x)
|
127 |
+
|
128 |
+
outs.append(x_out)
|
129 |
+
|
130 |
+
return tuple(outs)
|
131 |
+
|
132 |
+
def forward(self, x):
|
133 |
+
x = self.forward_features(x)
|
134 |
+
|
135 |
+
return x
|
136 |
+
|
137 |
+
class LayerNorm(nn.Module):
|
138 |
+
r""" LayerNorm that supports two data formats: channels_last (default) or channels_first.
|
139 |
+
The ordering of the dimensions in the inputs. channels_last corresponds to inputs with
|
140 |
+
shape (batch_size, height, width, channels) while channels_first corresponds to inputs
|
141 |
+
with shape (batch_size, channels, height, width).
|
142 |
+
"""
|
143 |
+
def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
|
144 |
+
super().__init__()
|
145 |
+
self.weight = nn.Parameter(torch.ones(normalized_shape))
|
146 |
+
self.bias = nn.Parameter(torch.zeros(normalized_shape))
|
147 |
+
self.eps = eps
|
148 |
+
self.data_format = data_format
|
149 |
+
if self.data_format not in ["channels_last", "channels_first"]:
|
150 |
+
raise NotImplementedError
|
151 |
+
self.normalized_shape = (normalized_shape, )
|
152 |
+
|
153 |
+
def forward(self, x):
|
154 |
+
if self.data_format == "channels_last":
|
155 |
+
return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
|
156 |
+
elif self.data_format == "channels_first":
|
157 |
+
u = x.mean(1, keepdim=True)
|
158 |
+
s = (x - u).pow(2).mean(1, keepdim=True)
|
159 |
+
x = (x - u) / torch.sqrt(s + self.eps)
|
160 |
+
x = self.weight[:, None, None] * x + self.bias[:, None, None]
|
161 |
+
return x
|
162 |
+
|
163 |
+
|
164 |
+
model_urls = {
|
165 |
+
"convnext_tiny_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_tiny_1k_224_ema.pth",
|
166 |
+
"convnext_small_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_small_1k_224_ema.pth",
|
167 |
+
"convnext_base_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_base_1k_224_ema.pth",
|
168 |
+
"convnext_large_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_large_1k_224_ema.pth",
|
169 |
+
"convnext_tiny_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_tiny_22k_224.pth",
|
170 |
+
"convnext_small_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_small_22k_224.pth",
|
171 |
+
"convnext_base_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_base_22k_224.pth",
|
172 |
+
"convnext_large_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_large_22k_224.pth",
|
173 |
+
"convnext_xlarge_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_xlarge_22k_224.pth",
|
174 |
+
}
|
175 |
+
|
176 |
+
@register_model
|
177 |
+
def convnext_tiny(pretrained=False,in_22k=False, **kwargs):
|
178 |
+
model = ConvNeXt(depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], **kwargs)
|
179 |
+
if pretrained:
|
180 |
+
url = model_urls['convnext_tiny_22k'] if in_22k else model_urls['convnext_tiny_1k']
|
181 |
+
checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu", check_hash=True)
|
182 |
+
model.load_state_dict(checkpoint["model"])
|
183 |
+
return model
|
184 |
+
|
185 |
+
@register_model
|
186 |
+
def convnext_small(pretrained=False,in_22k=False, **kwargs):
|
187 |
+
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[96, 192, 384, 768], **kwargs)
|
188 |
+
if pretrained:
|
189 |
+
url = model_urls['convnext_small_22k'] if in_22k else model_urls['convnext_small_1k']
|
190 |
+
checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
|
191 |
+
model.load_state_dict(checkpoint["model"], strict=False)
|
192 |
+
return model
|
193 |
+
|
194 |
+
@register_model
|
195 |
+
def convnext_base(pretrained=False, in_22k=False, **kwargs):
|
196 |
+
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[128, 256, 512, 1024], **kwargs)
|
197 |
+
if pretrained:
|
198 |
+
url = model_urls['convnext_base_22k'] if in_22k else model_urls['convnext_base_1k']
|
199 |
+
checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
|
200 |
+
model.load_state_dict(checkpoint["model"], strict=False)
|
201 |
+
return model
|
202 |
+
|
203 |
+
@register_model
|
204 |
+
def convnext_large(pretrained=False, in_22k=False, **kwargs):
|
205 |
+
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[192, 384, 768, 1536], **kwargs)
|
206 |
+
if pretrained:
|
207 |
+
url = model_urls['convnext_large_22k'] if in_22k else model_urls['convnext_large_1k']
|
208 |
+
checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
|
209 |
+
model.load_state_dict(checkpoint["model"])
|
210 |
+
return model
|
211 |
+
|
212 |
+
@register_model
|
213 |
+
def convnext_xlarge(pretrained=False, in_22k=False, **kwargs):
|
214 |
+
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[256, 512, 1024, 2048], **kwargs)
|
215 |
+
if pretrained:
|
216 |
+
assert in_22k, "only ImageNet-22K pre-trained ConvNeXt-XL is available; please set in_22k=True"
|
217 |
+
url = model_urls['convnext_xlarge_22k']
|
218 |
+
checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
|
219 |
+
model.load_state_dict(checkpoint["model"])
|
220 |
+
return model
|
models/flexible_unet.py
ADDED
@@ -0,0 +1,312 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) MONAI Consortium
|
2 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3 |
+
# you may not use this file except in compliance with the License.
|
4 |
+
# You may obtain a copy of the License at
|
5 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
6 |
+
# Unless required by applicable law or agreed to in writing, software
|
7 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
8 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
9 |
+
# See the License for the specific language governing permissions and
|
10 |
+
# limitations under the License.
|
11 |
+
|
12 |
+
from typing import List, Optional, Sequence, Tuple, Union
|
13 |
+
|
14 |
+
import torch
|
15 |
+
from torch import nn
|
16 |
+
|
17 |
+
from monai.networks.blocks import UpSample
|
18 |
+
from monai.networks.layers.factories import Conv
|
19 |
+
from monai.networks.layers.utils import get_act_layer
|
20 |
+
from monai.networks.nets import EfficientNetBNFeatures
|
21 |
+
from monai.networks.nets.basic_unet import UpCat
|
22 |
+
from monai.utils import InterpolateMode
|
23 |
+
|
24 |
+
__all__ = ["FlexibleUNet"]
|
25 |
+
|
26 |
+
encoder_feature_channel = {
|
27 |
+
"efficientnet-b0": (16, 24, 40, 112, 320),
|
28 |
+
"efficientnet-b1": (16, 24, 40, 112, 320),
|
29 |
+
"efficientnet-b2": (16, 24, 48, 120, 352),
|
30 |
+
"efficientnet-b3": (24, 32, 48, 136, 384),
|
31 |
+
"efficientnet-b4": (24, 32, 56, 160, 448),
|
32 |
+
"efficientnet-b5": (24, 40, 64, 176, 512),
|
33 |
+
"efficientnet-b6": (32, 40, 72, 200, 576),
|
34 |
+
"efficientnet-b7": (32, 48, 80, 224, 640),
|
35 |
+
"efficientnet-b8": (32, 56, 88, 248, 704),
|
36 |
+
"efficientnet-l2": (72, 104, 176, 480, 1376),
|
37 |
+
}
|
38 |
+
|
39 |
+
|
40 |
+
def _get_encoder_channels_by_backbone(backbone: str, in_channels: int = 3) -> tuple:
|
41 |
+
"""
|
42 |
+
Get the encoder output channels by given backbone name.
|
43 |
+
|
44 |
+
Args:
|
45 |
+
backbone: name of backbone to generate features, can be from [efficientnet-b0, ..., efficientnet-b7].
|
46 |
+
in_channels: channel of input tensor, default to 3.
|
47 |
+
|
48 |
+
Returns:
|
49 |
+
A tuple of output feature map channels' length .
|
50 |
+
"""
|
51 |
+
encoder_channel_tuple = encoder_feature_channel[backbone]
|
52 |
+
encoder_channel_list = [in_channels] + list(encoder_channel_tuple)
|
53 |
+
encoder_channel = tuple(encoder_channel_list)
|
54 |
+
return encoder_channel
|
55 |
+
|
56 |
+
|
57 |
+
class UNetDecoder(nn.Module):
|
58 |
+
"""
|
59 |
+
UNet Decoder.
|
60 |
+
This class refers to `segmentation_models.pytorch
|
61 |
+
<https://github.com/qubvel/segmentation_models.pytorch>`_.
|
62 |
+
|
63 |
+
Args:
|
64 |
+
spatial_dims: number of spatial dimensions.
|
65 |
+
encoder_channels: number of output channels for all feature maps in encoder.
|
66 |
+
`len(encoder_channels)` should be no less than 2.
|
67 |
+
decoder_channels: number of output channels for all feature maps in decoder.
|
68 |
+
`len(decoder_channels)` should equal to `len(encoder_channels) - 1`.
|
69 |
+
act: activation type and arguments.
|
70 |
+
norm: feature normalization type and arguments.
|
71 |
+
dropout: dropout ratio.
|
72 |
+
bias: whether to have a bias term in convolution blocks in this decoder.
|
73 |
+
upsample: upsampling mode, available options are
|
74 |
+
``"deconv"``, ``"pixelshuffle"``, ``"nontrainable"``.
|
75 |
+
pre_conv: a conv block applied before upsampling.
|
76 |
+
Only used in the "nontrainable" or "pixelshuffle" mode.
|
77 |
+
interp_mode: {``"nearest"``, ``"linear"``, ``"bilinear"``, ``"bicubic"``, ``"trilinear"``}
|
78 |
+
Only used in the "nontrainable" mode.
|
79 |
+
align_corners: set the align_corners parameter for upsample. Defaults to True.
|
80 |
+
Only used in the "nontrainable" mode.
|
81 |
+
is_pad: whether to pad upsampling features to fit the encoder spatial dims.
|
82 |
+
|
83 |
+
"""
|
84 |
+
|
85 |
+
def __init__(
|
86 |
+
self,
|
87 |
+
spatial_dims: int,
|
88 |
+
encoder_channels: Sequence[int],
|
89 |
+
decoder_channels: Sequence[int],
|
90 |
+
act: Union[str, tuple],
|
91 |
+
norm: Union[str, tuple],
|
92 |
+
dropout: Union[float, tuple],
|
93 |
+
bias: bool,
|
94 |
+
upsample: str,
|
95 |
+
pre_conv: Optional[str],
|
96 |
+
interp_mode: str,
|
97 |
+
align_corners: Optional[bool],
|
98 |
+
is_pad: bool,
|
99 |
+
):
|
100 |
+
|
101 |
+
super().__init__()
|
102 |
+
if len(encoder_channels) < 2:
|
103 |
+
raise ValueError("the length of `encoder_channels` should be no less than 2.")
|
104 |
+
if len(decoder_channels) != len(encoder_channels) - 1:
|
105 |
+
raise ValueError("`len(decoder_channels)` should equal to `len(encoder_channels) - 1`.")
|
106 |
+
|
107 |
+
in_channels = [encoder_channels[-1]] + list(decoder_channels[:-1])
|
108 |
+
skip_channels = list(encoder_channels[1:-1][::-1]) + [0]
|
109 |
+
halves = [True] * (len(skip_channels) - 1)
|
110 |
+
halves.append(False)
|
111 |
+
blocks = []
|
112 |
+
for in_chn, skip_chn, out_chn, halve in zip(in_channels, skip_channels, decoder_channels, halves):
|
113 |
+
blocks.append(
|
114 |
+
UpCat(
|
115 |
+
spatial_dims=spatial_dims,
|
116 |
+
in_chns=in_chn,
|
117 |
+
cat_chns=skip_chn,
|
118 |
+
out_chns=out_chn,
|
119 |
+
act=act,
|
120 |
+
norm=norm,
|
121 |
+
dropout=dropout,
|
122 |
+
bias=bias,
|
123 |
+
upsample=upsample,
|
124 |
+
pre_conv=pre_conv,
|
125 |
+
interp_mode=interp_mode,
|
126 |
+
align_corners=align_corners,
|
127 |
+
halves=halve,
|
128 |
+
is_pad=is_pad,
|
129 |
+
)
|
130 |
+
)
|
131 |
+
self.blocks = nn.ModuleList(blocks)
|
132 |
+
|
133 |
+
def forward(self, features: List[torch.Tensor], skip_connect: int = 4):
|
134 |
+
skips = features[:-1][::-1]
|
135 |
+
features = features[1:][::-1]
|
136 |
+
|
137 |
+
x = features[0]
|
138 |
+
for i, block in enumerate(self.blocks):
|
139 |
+
if i < skip_connect:
|
140 |
+
skip = skips[i]
|
141 |
+
else:
|
142 |
+
skip = None
|
143 |
+
x = block(x, skip)
|
144 |
+
|
145 |
+
return x
|
146 |
+
|
147 |
+
|
148 |
+
class SegmentationHead(nn.Sequential):
|
149 |
+
"""
|
150 |
+
Segmentation head.
|
151 |
+
This class refers to `segmentation_models.pytorch
|
152 |
+
<https://github.com/qubvel/segmentation_models.pytorch>`_.
|
153 |
+
|
154 |
+
Args:
|
155 |
+
spatial_dims: number of spatial dimensions.
|
156 |
+
in_channels: number of input channels for the block.
|
157 |
+
out_channels: number of output channels for the block.
|
158 |
+
kernel_size: kernel size for the conv layer.
|
159 |
+
act: activation type and arguments.
|
160 |
+
scale_factor: multiplier for spatial size. Has to match input size if it is a tuple.
|
161 |
+
|
162 |
+
"""
|
163 |
+
|
164 |
+
def __init__(
|
165 |
+
self,
|
166 |
+
spatial_dims: int,
|
167 |
+
in_channels: int,
|
168 |
+
out_channels: int,
|
169 |
+
kernel_size: int = 3,
|
170 |
+
act: Optional[Union[Tuple, str]] = None,
|
171 |
+
scale_factor: float = 1.0,
|
172 |
+
):
|
173 |
+
|
174 |
+
conv_layer = Conv[Conv.CONV, spatial_dims](
|
175 |
+
in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, padding=kernel_size // 2
|
176 |
+
)
|
177 |
+
up_layer: nn.Module = nn.Identity()
|
178 |
+
if scale_factor > 1.0:
|
179 |
+
up_layer = UpSample(
|
180 |
+
spatial_dims=spatial_dims,
|
181 |
+
scale_factor=scale_factor,
|
182 |
+
mode="nontrainable",
|
183 |
+
pre_conv=None,
|
184 |
+
interp_mode=InterpolateMode.LINEAR,
|
185 |
+
)
|
186 |
+
if act is not None:
|
187 |
+
act_layer = get_act_layer(act)
|
188 |
+
else:
|
189 |
+
act_layer = nn.Identity()
|
190 |
+
super().__init__(conv_layer, up_layer, act_layer)
|
191 |
+
|
192 |
+
|
193 |
+
class FlexibleUNet(nn.Module):
|
194 |
+
"""
|
195 |
+
A flexible implementation of UNet-like encoder-decoder architecture.
|
196 |
+
"""
|
197 |
+
|
198 |
+
def __init__(
|
199 |
+
self,
|
200 |
+
in_channels: int,
|
201 |
+
out_channels: int,
|
202 |
+
backbone: str,
|
203 |
+
pretrained: bool = False,
|
204 |
+
decoder_channels: Tuple = (256, 128, 64, 32, 16),
|
205 |
+
spatial_dims: int = 2,
|
206 |
+
norm: Union[str, tuple] = ("batch", {"eps": 1e-3, "momentum": 0.1}),
|
207 |
+
act: Union[str, tuple] = ("relu", {"inplace": True}),
|
208 |
+
dropout: Union[float, tuple] = 0.0,
|
209 |
+
decoder_bias: bool = False,
|
210 |
+
upsample: str = "nontrainable",
|
211 |
+
interp_mode: str = "nearest",
|
212 |
+
is_pad: bool = True,
|
213 |
+
) -> None:
|
214 |
+
"""
|
215 |
+
A flexible implement of UNet, in which the backbone/encoder can be replaced with
|
216 |
+
any efficient network. Currently the input must have a 2 or 3 spatial dimension
|
217 |
+
and the spatial size of each dimension must be a multiple of 32 if is pad parameter
|
218 |
+
is False
|
219 |
+
|
220 |
+
Args:
|
221 |
+
in_channels: number of input channels.
|
222 |
+
out_channels: number of output channels.
|
223 |
+
backbone: name of backbones to initialize, only support efficientnet right now,
|
224 |
+
can be from [efficientnet-b0,..., efficientnet-b8, efficientnet-l2].
|
225 |
+
pretrained: whether to initialize pretrained ImageNet weights, only available
|
226 |
+
for spatial_dims=2 and batch norm is used, default to False.
|
227 |
+
decoder_channels: number of output channels for all feature maps in decoder.
|
228 |
+
`len(decoder_channels)` should equal to `len(encoder_channels) - 1`,default
|
229 |
+
to (256, 128, 64, 32, 16).
|
230 |
+
spatial_dims: number of spatial dimensions, default to 2.
|
231 |
+
norm: normalization type and arguments, default to ("batch", {"eps": 1e-3,
|
232 |
+
"momentum": 0.1}).
|
233 |
+
act: activation type and arguments, default to ("relu", {"inplace": True}).
|
234 |
+
dropout: dropout ratio, default to 0.0.
|
235 |
+
decoder_bias: whether to have a bias term in decoder's convolution blocks.
|
236 |
+
upsample: upsampling mode, available options are``"deconv"``, ``"pixelshuffle"``,
|
237 |
+
``"nontrainable"``.
|
238 |
+
interp_mode: {``"nearest"``, ``"linear"``, ``"bilinear"``, ``"bicubic"``, ``"trilinear"``}
|
239 |
+
Only used in the "nontrainable" mode.
|
240 |
+
is_pad: whether to pad upsampling features to fit features from encoder. Default to True.
|
241 |
+
If this parameter is set to "True", the spatial dim of network input can be arbitary
|
242 |
+
size, which is not supported by TensorRT. Otherwise, it must be a multiple of 32.
|
243 |
+
"""
|
244 |
+
super().__init__()
|
245 |
+
|
246 |
+
if backbone not in encoder_feature_channel:
|
247 |
+
raise ValueError(f"invalid model_name {backbone} found, must be one of {encoder_feature_channel.keys()}.")
|
248 |
+
|
249 |
+
if spatial_dims not in (2, 3):
|
250 |
+
raise ValueError("spatial_dims can only be 2 or 3.")
|
251 |
+
|
252 |
+
adv_prop = "ap" in backbone
|
253 |
+
|
254 |
+
self.backbone = backbone
|
255 |
+
self.spatial_dims = spatial_dims
|
256 |
+
model_name = backbone
|
257 |
+
encoder_channels = _get_encoder_channels_by_backbone(backbone, in_channels)
|
258 |
+
self.encoder = EfficientNetBNFeatures(
|
259 |
+
model_name=model_name,
|
260 |
+
pretrained=pretrained,
|
261 |
+
in_channels=in_channels,
|
262 |
+
spatial_dims=spatial_dims,
|
263 |
+
norm=norm,
|
264 |
+
adv_prop=adv_prop,
|
265 |
+
)
|
266 |
+
self.decoder = UNetDecoder(
|
267 |
+
spatial_dims=spatial_dims,
|
268 |
+
encoder_channels=encoder_channels,
|
269 |
+
decoder_channels=decoder_channels,
|
270 |
+
act=act,
|
271 |
+
norm=norm,
|
272 |
+
dropout=dropout,
|
273 |
+
bias=decoder_bias,
|
274 |
+
upsample=upsample,
|
275 |
+
interp_mode=interp_mode,
|
276 |
+
pre_conv=None,
|
277 |
+
align_corners=None,
|
278 |
+
is_pad=is_pad,
|
279 |
+
)
|
280 |
+
self.dist_head = SegmentationHead(
|
281 |
+
spatial_dims=spatial_dims,
|
282 |
+
in_channels=decoder_channels[-1],
|
283 |
+
out_channels=32,
|
284 |
+
kernel_size=1,
|
285 |
+
act='relu',
|
286 |
+
)
|
287 |
+
self.prob_head = SegmentationHead(
|
288 |
+
spatial_dims=spatial_dims,
|
289 |
+
in_channels=decoder_channels[-1],
|
290 |
+
out_channels=1,
|
291 |
+
kernel_size=1,
|
292 |
+
act='sigmoid',
|
293 |
+
)
|
294 |
+
|
295 |
+
def forward(self, inputs: torch.Tensor):
|
296 |
+
"""
|
297 |
+
Do a typical encoder-decoder-header inference.
|
298 |
+
|
299 |
+
Args:
|
300 |
+
inputs: input should have spatially N dimensions ``(Batch, in_channels, dim_0[, dim_1, ..., dim_N])``,
|
301 |
+
N is defined by `dimensions`.
|
302 |
+
|
303 |
+
Returns:
|
304 |
+
A torch Tensor of "raw" predictions in shape ``(Batch, out_channels, dim_0[, dim_1, ..., dim_N])``.
|
305 |
+
|
306 |
+
"""
|
307 |
+
x = inputs
|
308 |
+
enc_out = self.encoder(x)
|
309 |
+
decoder_out = self.decoder(enc_out)
|
310 |
+
dist = self.dist_head(decoder_out)
|
311 |
+
prob = self.prob_head(decoder_out)
|
312 |
+
return dist,prob
|
models/flexible_unet_convnext.py
ADDED
@@ -0,0 +1,447 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) MONAI Consortium
|
2 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
3 |
+
# you may not use this file except in compliance with the License.
|
4 |
+
# You may obtain a copy of the License at
|
5 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
6 |
+
# Unless required by applicable law or agreed to in writing, software
|
7 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
8 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
9 |
+
# See the License for the specific language governing permissions and
|
10 |
+
# limitations under the License.
|
11 |
+
|
12 |
+
from typing import List, Optional, Sequence, Tuple, Union
|
13 |
+
|
14 |
+
import torch
|
15 |
+
from torch import nn
|
16 |
+
from . import convnext
|
17 |
+
from monai.networks.blocks import UpSample
|
18 |
+
from monai.networks.layers.factories import Conv
|
19 |
+
from monai.networks.layers.utils import get_act_layer
|
20 |
+
from monai.networks.nets import EfficientNetBNFeatures
|
21 |
+
from monai.networks.nets.basic_unet import UpCat
|
22 |
+
from monai.utils import InterpolateMode
|
23 |
+
|
24 |
+
__all__ = ["FlexibleUNet"]
|
25 |
+
|
26 |
+
encoder_feature_channel = {
|
27 |
+
"efficientnet-b0": (16, 24, 40, 112, 320),
|
28 |
+
"efficientnet-b1": (16, 24, 40, 112, 320),
|
29 |
+
"efficientnet-b2": (16, 24, 48, 120, 352),
|
30 |
+
"efficientnet-b3": (24, 32, 48, 136, 384),
|
31 |
+
"efficientnet-b4": (24, 32, 56, 160, 448),
|
32 |
+
"efficientnet-b5": (24, 40, 64, 176, 512),
|
33 |
+
"efficientnet-b6": (32, 40, 72, 200, 576),
|
34 |
+
"efficientnet-b7": (32, 48, 80, 224, 640),
|
35 |
+
"efficientnet-b8": (32, 56, 88, 248, 704),
|
36 |
+
"efficientnet-l2": (72, 104, 176, 480, 1376),
|
37 |
+
"convnext_small": (96, 192, 384, 768),
|
38 |
+
"convnext_base": (128, 256, 512, 1024),
|
39 |
+
"van_b2": (64, 128, 320, 512),
|
40 |
+
"van_b1": (64, 128, 320, 512),
|
41 |
+
}
|
42 |
+
|
43 |
+
|
44 |
+
def _get_encoder_channels_by_backbone(backbone: str, in_channels: int = 3) -> tuple:
|
45 |
+
"""
|
46 |
+
Get the encoder output channels by given backbone name.
|
47 |
+
|
48 |
+
Args:
|
49 |
+
backbone: name of backbone to generate features, can be from [efficientnet-b0, ..., efficientnet-b7].
|
50 |
+
in_channels: channel of input tensor, default to 3.
|
51 |
+
|
52 |
+
Returns:
|
53 |
+
A tuple of output feature map channels' length .
|
54 |
+
"""
|
55 |
+
encoder_channel_tuple = encoder_feature_channel[backbone]
|
56 |
+
encoder_channel_list = [in_channels] + list(encoder_channel_tuple)
|
57 |
+
encoder_channel = tuple(encoder_channel_list)
|
58 |
+
return encoder_channel
|
59 |
+
|
60 |
+
|
61 |
+
class UNetDecoder(nn.Module):
|
62 |
+
"""
|
63 |
+
UNet Decoder.
|
64 |
+
This class refers to `segmentation_models.pytorch
|
65 |
+
<https://github.com/qubvel/segmentation_models.pytorch>`_.
|
66 |
+
|
67 |
+
Args:
|
68 |
+
spatial_dims: number of spatial dimensions.
|
69 |
+
encoder_channels: number of output channels for all feature maps in encoder.
|
70 |
+
`len(encoder_channels)` should be no less than 2.
|
71 |
+
decoder_channels: number of output channels for all feature maps in decoder.
|
72 |
+
`len(decoder_channels)` should equal to `len(encoder_channels) - 1`.
|
73 |
+
act: activation type and arguments.
|
74 |
+
norm: feature normalization type and arguments.
|
75 |
+
dropout: dropout ratio.
|
76 |
+
bias: whether to have a bias term in convolution blocks in this decoder.
|
77 |
+
upsample: upsampling mode, available options are
|
78 |
+
``"deconv"``, ``"pixelshuffle"``, ``"nontrainable"``.
|
79 |
+
pre_conv: a conv block applied before upsampling.
|
80 |
+
Only used in the "nontrainable" or "pixelshuffle" mode.
|
81 |
+
interp_mode: {``"nearest"``, ``"linear"``, ``"bilinear"``, ``"bicubic"``, ``"trilinear"``}
|
82 |
+
Only used in the "nontrainable" mode.
|
83 |
+
align_corners: set the align_corners parameter for upsample. Defaults to True.
|
84 |
+
Only used in the "nontrainable" mode.
|
85 |
+
is_pad: whether to pad upsampling features to fit the encoder spatial dims.
|
86 |
+
|
87 |
+
"""
|
88 |
+
|
89 |
+
def __init__(
|
90 |
+
self,
|
91 |
+
spatial_dims: int,
|
92 |
+
encoder_channels: Sequence[int],
|
93 |
+
decoder_channels: Sequence[int],
|
94 |
+
act: Union[str, tuple],
|
95 |
+
norm: Union[str, tuple],
|
96 |
+
dropout: Union[float, tuple],
|
97 |
+
bias: bool,
|
98 |
+
upsample: str,
|
99 |
+
pre_conv: Optional[str],
|
100 |
+
interp_mode: str,
|
101 |
+
align_corners: Optional[bool],
|
102 |
+
is_pad: bool,
|
103 |
+
):
|
104 |
+
|
105 |
+
super().__init__()
|
106 |
+
if len(encoder_channels) < 2:
|
107 |
+
raise ValueError("the length of `encoder_channels` should be no less than 2.")
|
108 |
+
if len(decoder_channels) != len(encoder_channels) - 1:
|
109 |
+
raise ValueError("`len(decoder_channels)` should equal to `len(encoder_channels) - 1`.")
|
110 |
+
|
111 |
+
in_channels = [encoder_channels[-1]] + list(decoder_channels[:-1])
|
112 |
+
skip_channels = list(encoder_channels[1:-1][::-1]) + [0]
|
113 |
+
halves = [True] * (len(skip_channels) - 1)
|
114 |
+
halves.append(False)
|
115 |
+
blocks = []
|
116 |
+
for in_chn, skip_chn, out_chn, halve in zip(in_channels, skip_channels, decoder_channels, halves):
|
117 |
+
blocks.append(
|
118 |
+
UpCat(
|
119 |
+
spatial_dims=spatial_dims,
|
120 |
+
in_chns=in_chn,
|
121 |
+
cat_chns=skip_chn,
|
122 |
+
out_chns=out_chn,
|
123 |
+
act=act,
|
124 |
+
norm=norm,
|
125 |
+
dropout=dropout,
|
126 |
+
bias=bias,
|
127 |
+
upsample=upsample,
|
128 |
+
pre_conv=pre_conv,
|
129 |
+
interp_mode=interp_mode,
|
130 |
+
align_corners=align_corners,
|
131 |
+
halves=halve,
|
132 |
+
is_pad=is_pad,
|
133 |
+
)
|
134 |
+
)
|
135 |
+
self.blocks = nn.ModuleList(blocks)
|
136 |
+
|
137 |
+
def forward(self, features: List[torch.Tensor], skip_connect: int = 3):
|
138 |
+
skips = features[:-1][::-1]
|
139 |
+
features = features[1:][::-1]
|
140 |
+
|
141 |
+
x = features[0]
|
142 |
+
for i, block in enumerate(self.blocks):
|
143 |
+
if i < skip_connect:
|
144 |
+
skip = skips[i]
|
145 |
+
else:
|
146 |
+
skip = None
|
147 |
+
x = block(x, skip)
|
148 |
+
|
149 |
+
return x
|
150 |
+
|
151 |
+
|
152 |
+
class SegmentationHead(nn.Sequential):
|
153 |
+
"""
|
154 |
+
Segmentation head.
|
155 |
+
This class refers to `segmentation_models.pytorch
|
156 |
+
<https://github.com/qubvel/segmentation_models.pytorch>`_.
|
157 |
+
|
158 |
+
Args:
|
159 |
+
spatial_dims: number of spatial dimensions.
|
160 |
+
in_channels: number of input channels for the block.
|
161 |
+
out_channels: number of output channels for the block.
|
162 |
+
kernel_size: kernel size for the conv layer.
|
163 |
+
act: activation type and arguments.
|
164 |
+
scale_factor: multiplier for spatial size. Has to match input size if it is a tuple.
|
165 |
+
|
166 |
+
"""
|
167 |
+
|
168 |
+
def __init__(
|
169 |
+
self,
|
170 |
+
spatial_dims: int,
|
171 |
+
in_channels: int,
|
172 |
+
out_channels: int,
|
173 |
+
kernel_size: int = 3,
|
174 |
+
act: Optional[Union[Tuple, str]] = None,
|
175 |
+
scale_factor: float = 1.0,
|
176 |
+
):
|
177 |
+
|
178 |
+
conv_layer = Conv[Conv.CONV, spatial_dims](
|
179 |
+
in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, padding=kernel_size // 2
|
180 |
+
)
|
181 |
+
up_layer: nn.Module = nn.Identity()
|
182 |
+
# if scale_factor > 1.0:
|
183 |
+
# up_layer = UpSample(
|
184 |
+
# in_channels=out_channels,
|
185 |
+
# spatial_dims=spatial_dims,
|
186 |
+
# scale_factor=scale_factor,
|
187 |
+
# mode="deconv",
|
188 |
+
# pre_conv=None,
|
189 |
+
# interp_mode=InterpolateMode.LINEAR,
|
190 |
+
# )
|
191 |
+
if scale_factor > 1.0:
|
192 |
+
up_layer = UpSample(
|
193 |
+
spatial_dims=spatial_dims,
|
194 |
+
scale_factor=scale_factor,
|
195 |
+
mode="nontrainable",
|
196 |
+
pre_conv=None,
|
197 |
+
interp_mode=InterpolateMode.LINEAR,
|
198 |
+
)
|
199 |
+
if act is not None:
|
200 |
+
act_layer = get_act_layer(act)
|
201 |
+
else:
|
202 |
+
act_layer = nn.Identity()
|
203 |
+
super().__init__(conv_layer, up_layer, act_layer)
|
204 |
+
|
205 |
+
|
206 |
+
class FlexibleUNet_star(nn.Module):
|
207 |
+
"""
|
208 |
+
A flexible implementation of UNet-like encoder-decoder architecture.
|
209 |
+
"""
|
210 |
+
|
211 |
+
def __init__(
|
212 |
+
self,
|
213 |
+
in_channels: int,
|
214 |
+
out_channels: int,
|
215 |
+
backbone: str,
|
216 |
+
pretrained: bool = False,
|
217 |
+
decoder_channels: Tuple = (256, 128, 64, 32),
|
218 |
+
#decoder_channels: Tuple = (1024, 512, 256, 128),
|
219 |
+
spatial_dims: int = 2,
|
220 |
+
norm: Union[str, tuple] = ("batch", {"eps": 1e-3, "momentum": 0.1}),
|
221 |
+
act: Union[str, tuple] = ("relu", {"inplace": True}),
|
222 |
+
dropout: Union[float, tuple] = 0.0,
|
223 |
+
decoder_bias: bool = False,
|
224 |
+
upsample: str = "nontrainable",
|
225 |
+
interp_mode: str = "nearest",
|
226 |
+
is_pad: bool = True,
|
227 |
+
n_rays: int = 32,
|
228 |
+
prob_out_channels: int = 1,
|
229 |
+
) -> None:
|
230 |
+
"""
|
231 |
+
A flexible implement of UNet, in which the backbone/encoder can be replaced with
|
232 |
+
any efficient network. Currently the input must have a 2 or 3 spatial dimension
|
233 |
+
and the spatial size of each dimension must be a multiple of 32 if is pad parameter
|
234 |
+
is False
|
235 |
+
|
236 |
+
Args:
|
237 |
+
in_channels: number of input channels.
|
238 |
+
out_channels: number of output channels.
|
239 |
+
backbone: name of backbones to initialize, only support efficientnet right now,
|
240 |
+
can be from [efficientnet-b0,..., efficientnet-b8, efficientnet-l2].
|
241 |
+
pretrained: whether to initialize pretrained ImageNet weights, only available
|
242 |
+
for spatial_dims=2 and batch norm is used, default to False.
|
243 |
+
decoder_channels: number of output channels for all feature maps in decoder.
|
244 |
+
`len(decoder_channels)` should equal to `len(encoder_channels) - 1`,default
|
245 |
+
to (256, 128, 64, 32, 16).
|
246 |
+
spatial_dims: number of spatial dimensions, default to 2.
|
247 |
+
norm: normalization type and arguments, default to ("batch", {"eps": 1e-3,
|
248 |
+
"momentum": 0.1}).
|
249 |
+
act: activation type and arguments, default to ("relu", {"inplace": True}).
|
250 |
+
dropout: dropout ratio, default to 0.0.
|
251 |
+
decoder_bias: whether to have a bias term in decoder's convolution blocks.
|
252 |
+
upsample: upsampling mode, available options are``"deconv"``, ``"pixelshuffle"``,
|
253 |
+
``"nontrainable"``.
|
254 |
+
interp_mode: {``"nearest"``, ``"linear"``, ``"bilinear"``, ``"bicubic"``, ``"trilinear"``}
|
255 |
+
Only used in the "nontrainable" mode.
|
256 |
+
is_pad: whether to pad upsampling features to fit features from encoder. Default to True.
|
257 |
+
If this parameter is set to "True", the spatial dim of network input can be arbitary
|
258 |
+
size, which is not supported by TensorRT. Otherwise, it must be a multiple of 32.
|
259 |
+
"""
|
260 |
+
super().__init__()
|
261 |
+
|
262 |
+
if backbone not in encoder_feature_channel:
|
263 |
+
raise ValueError(f"invalid model_name {backbone} found, must be one of {encoder_feature_channel.keys()}.")
|
264 |
+
|
265 |
+
if spatial_dims not in (2, 3):
|
266 |
+
raise ValueError("spatial_dims can only be 2 or 3.")
|
267 |
+
|
268 |
+
adv_prop = "ap" in backbone
|
269 |
+
|
270 |
+
self.backbone = backbone
|
271 |
+
self.spatial_dims = spatial_dims
|
272 |
+
model_name = backbone
|
273 |
+
encoder_channels = _get_encoder_channels_by_backbone(backbone, in_channels)
|
274 |
+
|
275 |
+
self.encoder = convnext.convnext_small(pretrained=False,in_22k=True)
|
276 |
+
|
277 |
+
self.decoder = UNetDecoder(
|
278 |
+
spatial_dims=spatial_dims,
|
279 |
+
encoder_channels=encoder_channels,
|
280 |
+
decoder_channels=decoder_channels,
|
281 |
+
act=act,
|
282 |
+
norm=norm,
|
283 |
+
dropout=dropout,
|
284 |
+
bias=decoder_bias,
|
285 |
+
upsample=upsample,
|
286 |
+
interp_mode=interp_mode,
|
287 |
+
pre_conv=None,
|
288 |
+
align_corners=None,
|
289 |
+
is_pad=is_pad,
|
290 |
+
)
|
291 |
+
self.dist_head = SegmentationHead(
|
292 |
+
spatial_dims=spatial_dims,
|
293 |
+
in_channels=decoder_channels[-1],
|
294 |
+
out_channels=n_rays,
|
295 |
+
kernel_size=1,
|
296 |
+
act='relu',
|
297 |
+
scale_factor = 2,
|
298 |
+
)
|
299 |
+
self.prob_head = SegmentationHead(
|
300 |
+
spatial_dims=spatial_dims,
|
301 |
+
in_channels=decoder_channels[-1],
|
302 |
+
out_channels=prob_out_channels,
|
303 |
+
kernel_size=1,
|
304 |
+
act='sigmoid',
|
305 |
+
scale_factor = 2,
|
306 |
+
)
|
307 |
+
|
308 |
+
def forward(self, inputs: torch.Tensor):
|
309 |
+
"""
|
310 |
+
Do a typical encoder-decoder-header inference.
|
311 |
+
|
312 |
+
Args:
|
313 |
+
inputs: input should have spatially N dimensions ``(Batch, in_channels, dim_0[, dim_1, ..., dim_N])``,
|
314 |
+
N is defined by `dimensions`.
|
315 |
+
|
316 |
+
Returns:
|
317 |
+
A torch Tensor of "raw" predictions in shape ``(Batch, out_channels, dim_0[, dim_1, ..., dim_N])``.
|
318 |
+
|
319 |
+
"""
|
320 |
+
x = inputs
|
321 |
+
enc_out = self.encoder(x)
|
322 |
+
decoder_out = self.decoder(enc_out)
|
323 |
+
|
324 |
+
dist = self.dist_head(decoder_out)
|
325 |
+
prob = self.prob_head(decoder_out)
|
326 |
+
|
327 |
+
return dist,prob
|
328 |
+
|
329 |
+
|
330 |
+
|
331 |
+
class FlexibleUNet_hv(nn.Module):
|
332 |
+
"""
|
333 |
+
A flexible implementation of UNet-like encoder-decoder architecture.
|
334 |
+
"""
|
335 |
+
|
336 |
+
def __init__(
|
337 |
+
self,
|
338 |
+
in_channels: int,
|
339 |
+
out_channels: int,
|
340 |
+
backbone: str,
|
341 |
+
pretrained: bool = False,
|
342 |
+
decoder_channels: Tuple = (1024, 512, 256, 128),
|
343 |
+
spatial_dims: int = 2,
|
344 |
+
norm: Union[str, tuple] = ("batch", {"eps": 1e-3, "momentum": 0.1}),
|
345 |
+
act: Union[str, tuple] = ("relu", {"inplace": True}),
|
346 |
+
dropout: Union[float, tuple] = 0.0,
|
347 |
+
decoder_bias: bool = False,
|
348 |
+
upsample: str = "nontrainable",
|
349 |
+
interp_mode: str = "nearest",
|
350 |
+
is_pad: bool = True,
|
351 |
+
n_rays: int = 32,
|
352 |
+
prob_out_channels: int = 1,
|
353 |
+
) -> None:
|
354 |
+
"""
|
355 |
+
A flexible implement of UNet, in which the backbone/encoder can be replaced with
|
356 |
+
any efficient network. Currently the input must have a 2 or 3 spatial dimension
|
357 |
+
and the spatial size of each dimension must be a multiple of 32 if is pad parameter
|
358 |
+
is False
|
359 |
+
|
360 |
+
Args:
|
361 |
+
in_channels: number of input channels.
|
362 |
+
out_channels: number of output channels.
|
363 |
+
backbone: name of backbones to initialize, only support efficientnet right now,
|
364 |
+
can be from [efficientnet-b0,..., efficientnet-b8, efficientnet-l2].
|
365 |
+
pretrained: whether to initialize pretrained ImageNet weights, only available
|
366 |
+
for spatial_dims=2 and batch norm is used, default to False.
|
367 |
+
decoder_channels: number of output channels for all feature maps in decoder.
|
368 |
+
`len(decoder_channels)` should equal to `len(encoder_channels) - 1`,default
|
369 |
+
to (256, 128, 64, 32, 16).
|
370 |
+
spatial_dims: number of spatial dimensions, default to 2.
|
371 |
+
norm: normalization type and arguments, default to ("batch", {"eps": 1e-3,
|
372 |
+
"momentum": 0.1}).
|
373 |
+
act: activation type and arguments, default to ("relu", {"inplace": True}).
|
374 |
+
dropout: dropout ratio, default to 0.0.
|
375 |
+
decoder_bias: whether to have a bias term in decoder's convolution blocks.
|
376 |
+
upsample: upsampling mode, available options are``"deconv"``, ``"pixelshuffle"``,
|
377 |
+
``"nontrainable"``.
|
378 |
+
interp_mode: {``"nearest"``, ``"linear"``, ``"bilinear"``, ``"bicubic"``, ``"trilinear"``}
|
379 |
+
Only used in the "nontrainable" mode.
|
380 |
+
is_pad: whether to pad upsampling features to fit features from encoder. Default to True.
|
381 |
+
If this parameter is set to "True", the spatial dim of network input can be arbitary
|
382 |
+
size, which is not supported by TensorRT. Otherwise, it must be a multiple of 32.
|
383 |
+
"""
|
384 |
+
super().__init__()
|
385 |
+
|
386 |
+
if backbone not in encoder_feature_channel:
|
387 |
+
raise ValueError(f"invalid model_name {backbone} found, must be one of {encoder_feature_channel.keys()}.")
|
388 |
+
|
389 |
+
if spatial_dims not in (2, 3):
|
390 |
+
raise ValueError("spatial_dims can only be 2 or 3.")
|
391 |
+
|
392 |
+
adv_prop = "ap" in backbone
|
393 |
+
|
394 |
+
self.backbone = backbone
|
395 |
+
self.spatial_dims = spatial_dims
|
396 |
+
model_name = backbone
|
397 |
+
encoder_channels = _get_encoder_channels_by_backbone(backbone, in_channels)
|
398 |
+
self.encoder = convnext.convnext_small(pretrained=False,in_22k=True)
|
399 |
+
self.decoder = UNetDecoder(
|
400 |
+
spatial_dims=spatial_dims,
|
401 |
+
encoder_channels=encoder_channels,
|
402 |
+
decoder_channels=decoder_channels,
|
403 |
+
act=act,
|
404 |
+
norm=norm,
|
405 |
+
dropout=dropout,
|
406 |
+
bias=decoder_bias,
|
407 |
+
upsample=upsample,
|
408 |
+
interp_mode=interp_mode,
|
409 |
+
pre_conv=None,
|
410 |
+
align_corners=None,
|
411 |
+
is_pad=is_pad,
|
412 |
+
)
|
413 |
+
self.dist_head = SegmentationHead(
|
414 |
+
spatial_dims=spatial_dims,
|
415 |
+
in_channels=decoder_channels[-1],
|
416 |
+
out_channels=n_rays,
|
417 |
+
kernel_size=1,
|
418 |
+
act=None,
|
419 |
+
scale_factor = 2,
|
420 |
+
)
|
421 |
+
self.prob_head = SegmentationHead(
|
422 |
+
spatial_dims=spatial_dims,
|
423 |
+
in_channels=decoder_channels[-1],
|
424 |
+
out_channels=prob_out_channels,
|
425 |
+
kernel_size=1,
|
426 |
+
act='sigmoid',
|
427 |
+
scale_factor = 2,
|
428 |
+
)
|
429 |
+
|
430 |
+
def forward(self, inputs: torch.Tensor):
|
431 |
+
"""
|
432 |
+
Do a typical encoder-decoder-header inference.
|
433 |
+
|
434 |
+
Args:
|
435 |
+
inputs: input should have spatially N dimensions ``(Batch, in_channels, dim_0[, dim_1, ..., dim_N])``,
|
436 |
+
N is defined by `dimensions`.
|
437 |
+
|
438 |
+
Returns:
|
439 |
+
A torch Tensor of "raw" predictions in shape ``(Batch, out_channels, dim_0[, dim_1, ..., dim_N])``.
|
440 |
+
|
441 |
+
"""
|
442 |
+
x = inputs
|
443 |
+
enc_out = self.encoder(x)
|
444 |
+
decoder_out = self.decoder(enc_out)
|
445 |
+
dist = self.dist_head(decoder_out)
|
446 |
+
prob = self.prob_head(decoder_out)
|
447 |
+
return dist,prob
|