modules.conv.UNet2d¶
net = mdnc.modules.conv.UNet2d(
channel, layers,
kernel_size=3, in_planes=1, out_planes=1
)
This moule is a built-in model for 2D convolutional U-Net. The network is inspired by:
The network would down-sample and up-sample the input data according to the network depth. The depth is given by the length of the argument layers. The network structure is shown in the following chart:
flowchart TB
b1["Block 1<br>Stack of layers[0] layers"]
b2["Block 2<br>Stack of layers[1] layers"]
bi["Block ...<br>Stack of ... layers"]
bn["Block n<br>Stack of layers[n-1] layers"]
u1["Block 2n-1<br>Stack of layers[0] layers"]
u2["Block 2n-2<br>Stack of layers[1] layers"]
ui["Block ...<br>Stack of ... layers"]
b1 -->|down<br>sampling| b2 -->|down<br>sampling| bi -->|down<br>sampling| bn
bn -->|up<br>sampling| ui -->|up<br>sampling| u2 -->|up<br>sampling| u1
b1 -->|skip<br>connection| u1
b2 -->|skip<br>connection| u2
bi -->|skip<br>connection| ui
linkStyle 0,1,2 stroke-width:4px, stroke:#800 ;
linkStyle 3,4,5 stroke-width:4px, stroke:#080 ;
linkStyle 6,7,8 stroke-width:4px, stroke:#888 ;The argument layers is a sequence of int. For each block \(i\), it contains layers[i-1] repeated modern convolutional layers (see mdnc.modules.conv.ConvModern2d). Each down-sampling or up-sampling is configured by stride=2. The channel number would be doubled in the down-sampling route and reduced to ½ in the up-sampling route. The skip connection is perfromed by concatenation.
Arguments¶
Requries
| Argument | Type | Description |
|---|---|---|
channel | int | The channel number of the first hidden block (layer). After each down-sampling, the channel number would be doubled. After each up-sampling, the channel number would be reduced to ½. |
layers | (int,) | A sequence of layer numbers for each block. Each number represents the number of convolutional layers of a stage (block). The stage numer, i.e. the depth of the network is the length of this list. |
kernel_size | int or(int, int) | The kernel size of each convolutional layer. |
in_planes | int | The channel number of the input data. |
out_planes | int | The channel number of the output data. |
Operators¶
__call__¶
y = net(x)
The forward operator implemented by the forward() method. The input is a 2D tensor, and the output is the final output of this network.
Requries
| Argument | Type | Description |
|---|---|---|
x | torch.Tensor | A 2D tensor, the size should be (B, C, L1, L2), where B is the batch size, C is the input channel number, and (L1, L2) is the input data size. |
Returns
| Argument | Description |
|---|---|
y | A 2D tensor, the size should be (B, C, L1, L2), where B is the batch size, C is the output channel number, and (L1, L2) is the input data size. |
Properties¶
nlayers¶
net.nlayers
The total number of convolutional layers along the depth of the network.
Examples¶
Example
1 2 3 4 5 | |
The number of convolutional layers along the depth is 25.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 64, 63] 4,800
InstanceNorm2d-2 [-1, 64, 64, 63] 128
PReLU-3 [-1, 64, 64, 63] 64
Conv2d-4 [-1, 64, 64, 63] 36,864
_ConvModernNd-5 [-1, 64, 64, 63] 0
InstanceNorm2d-6 [-1, 64, 64, 63] 128
PReLU-7 [-1, 64, 64, 63] 64
Conv2d-8 [-1, 64, 32, 32] 36,864
_ConvModernNd-9 [-1, 64, 32, 32] 0
_BlockConvStkNd-10 [-1, 64, 32, 32] 0
[-1, 64, 64, 63]
InstanceNorm2d-11 [-1, 64, 32, 32] 128
PReLU-12 [-1, 64, 32, 32] 64
Conv2d-13 [-1, 128, 32, 32] 73,728
_ConvModernNd-14 [-1, 128, 32, 32] 0
InstanceNorm2d-15 [-1, 128, 32, 32] 256
PReLU-16 [-1, 128, 32, 32] 128
Conv2d-17 [-1, 128, 16, 16] 147,456
_ConvModernNd-18 [-1, 128, 16, 16] 0
_BlockConvStkNd-19 [-1, 128, 16, 16] 0
[-1, 128, 32, 32]
InstanceNorm2d-20 [-1, 128, 16, 16] 256
PReLU-21 [-1, 128, 16, 16] 128
Conv2d-22 [-1, 256, 16, 16] 294,912
_ConvModernNd-23 [-1, 256, 16, 16] 0
InstanceNorm2d-24 [-1, 256, 16, 16] 512
PReLU-25 [-1, 256, 16, 16] 256
Conv2d-26 [-1, 256, 16, 16] 589,824
_ConvModernNd-27 [-1, 256, 16, 16] 0
InstanceNorm2d-28 [-1, 256, 16, 16] 512
PReLU-29 [-1, 256, 16, 16] 256
Conv2d-30 [-1, 256, 8, 8] 589,824
_ConvModernNd-31 [-1, 256, 8, 8] 0
_BlockConvStkNd-32 [-1, 256, 8, 8] 0
[-1, 256, 16, 16]
InstanceNorm2d-33 [-1, 256, 8, 8] 512
PReLU-34 [-1, 256, 8, 8] 256
Conv2d-35 [-1, 512, 8, 8] 1,179,648
_ConvModernNd-36 [-1, 512, 8, 8] 0
InstanceNorm2d-37 [-1, 512, 8, 8] 1,024
PReLU-38 [-1, 512, 8, 8] 512
Conv2d-39 [-1, 512, 8, 8] 2,359,296
_ConvModernNd-40 [-1, 512, 8, 8] 0
InstanceNorm2d-41 [-1, 512, 8, 8] 1,024
PReLU-42 [-1, 512, 8, 8] 512
Conv2d-43 [-1, 512, 4, 4] 2,359,296
_ConvModernNd-44 [-1, 512, 4, 4] 0
_BlockConvStkNd-45 [-1, 512, 4, 4] 0
[-1, 512, 8, 8]
InstanceNorm2d-46 [-1, 512, 4, 4] 1,024
PReLU-47 [-1, 512, 4, 4] 512
Conv2d-48 [-1, 1024, 4, 4] 4,718,592
_ConvModernNd-49 [-1, 1024, 4, 4] 0
InstanceNorm2d-50 [-1, 1024, 4, 4] 2,048
PReLU-51 [-1, 1024, 4, 4] 1,024
Conv2d-52 [-1, 1024, 4, 4] 9,437,184
_ConvModernNd-53 [-1, 1024, 4, 4] 0
InstanceNorm2d-54 [-1, 1024, 4, 4] 2,048
PReLU-55 [-1, 1024, 4, 4] 1,024
Upsample-56 [-1, 1024, 8, 8] 0
Conv2d-57 [-1, 512, 8, 8] 4,718,592
_ConvModernNd-58 [-1, 512, 8, 8] 0
_BlockConvStkNd-59 [-1, 512, 8, 8] 0
InstanceNorm2d-60 [-1, 1024, 8, 8] 2,048
PReLU-61 [-1, 1024, 8, 8] 1,024
Conv2d-62 [-1, 512, 8, 8] 4,718,592
_ConvModernNd-63 [-1, 512, 8, 8] 0
InstanceNorm2d-64 [-1, 512, 8, 8] 1,024
PReLU-65 [-1, 512, 8, 8] 512
Conv2d-66 [-1, 512, 8, 8] 2,359,296
_ConvModernNd-67 [-1, 512, 8, 8] 0
InstanceNorm2d-68 [-1, 512, 8, 8] 1,024
PReLU-69 [-1, 512, 8, 8] 512
Upsample-70 [-1, 512, 16, 16] 0
Conv2d-71 [-1, 256, 16, 16] 1,179,648
_ConvModernNd-72 [-1, 256, 16, 16] 0
_BlockConvStkNd-73 [-1, 256, 16, 16] 0
InstanceNorm2d-74 [-1, 512, 16, 16] 1,024
PReLU-75 [-1, 512, 16, 16] 512
Conv2d-76 [-1, 256, 16, 16] 1,179,648
_ConvModernNd-77 [-1, 256, 16, 16] 0
InstanceNorm2d-78 [-1, 256, 16, 16] 512
PReLU-79 [-1, 256, 16, 16] 256
Conv2d-80 [-1, 256, 16, 16] 589,824
_ConvModernNd-81 [-1, 256, 16, 16] 0
InstanceNorm2d-82 [-1, 256, 16, 16] 512
PReLU-83 [-1, 256, 16, 16] 256
Upsample-84 [-1, 256, 32, 32] 0
Conv2d-85 [-1, 128, 32, 32] 294,912
_ConvModernNd-86 [-1, 128, 32, 32] 0
_BlockConvStkNd-87 [-1, 128, 32, 32] 0
InstanceNorm2d-88 [-1, 256, 32, 32] 512
PReLU-89 [-1, 256, 32, 32] 256
Conv2d-90 [-1, 128, 32, 32] 294,912
_ConvModernNd-91 [-1, 128, 32, 32] 0
InstanceNorm2d-92 [-1, 128, 32, 32] 256
PReLU-93 [-1, 128, 32, 32] 128
Upsample-94 [-1, 128, 64, 64] 0
Conv2d-95 [-1, 64, 64, 64] 73,728
_ConvModernNd-96 [-1, 64, 64, 64] 0
_BlockConvStkNd-97 [-1, 64, 64, 64] 0
InstanceNorm2d-98 [-1, 128, 64, 63] 256
PReLU-99 [-1, 128, 64, 63] 128
Conv2d-100 [-1, 64, 64, 63] 73,728
_ConvModernNd-101 [-1, 64, 64, 63] 0
InstanceNorm2d-102 [-1, 64, 64, 63] 128
PReLU-103 [-1, 64, 64, 63] 64
Conv2d-104 [-1, 64, 64, 63] 36,864
_ConvModernNd-105 [-1, 64, 64, 63] 0
_BlockConvStkNd-106 [-1, 64, 64, 63] 0
Conv2d-107 [-1, 1, 64, 63] 1,601
UNet2d-108 [-1, 1, 64, 63] 0
================================================================
Total params: 37,374,977
Trainable params: 37,374,977
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.05
Forward/backward pass size (MB): 90.66
Params size (MB): 142.57
Estimated Total Size (MB): 233.28
----------------------------------------------------------------
Last update: March 14, 2021