modules.resnet.EncoderNet3d¶
net = mdnc.modules.resnet.EncoderNet3d(
channel, layers, block='bottleneck',
kernel_size=3, in_planes=1, out_length=2
)
This moule is a built-in model for 3D residual encoder network. This network could be used as a part of the auto-encoder, or just a network for down-sampling data.
The network would down-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] blocks"]
b2["Block 2<br>Stack of layers[1] blocks"]
bi["Block ...<br>Stack of ... layers"]
bn["Block n<br>Stack of layers[n-1] blocks"]
optional:::blockoptional
subgraph optional [Optional]
fc["FC layer"]
end
b1 -->|down<br>sampling| b2 -->|down<br>sampling| bi -->|down<br>sampling| bn
bn -.->|flatten| fc
linkStyle 0,1,2 stroke-width:4px, stroke:#800 ;
classDef blockoptional fill:none, stroke-dasharray:10,10, stroke:#9370DB, width:100;The argument layers is a sequence of int. For each block \(i\), it contains layers[i-1] repeated residual blocks (see mdnc.modules.resnet.BlockPlain3d and mdnc.modules.resnet.BlockBottleneck3d). Each down-sampling or up-sampling is configured by stride=2. The channel number would be doubled in the down-sampling route. An optional flattener and fully-connected layer could be appended to the last layer when the argument out_length != None.
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. |
layers | (int,) | A sequence of layer numbers for each block. Each number represents the number of residual blocks of a stage (block). The stage numer, i.e. the depth of the network is the length of this list. |
block | str | The residual block type, could be:
|
kernel_size | int or(int, int, int) | The kernel size of each residual block. |
in_planes | int | The channel number of the input data. |
out_length | int | The length of the output vector, if not set, the output would not be flattened. |
Operators¶
__call__¶
y = net(x)
The forward operator implemented by the forward() method. The input is a 3D tensor, and the output is the final output of this network.
Requries
| Argument | Type | Description |
|---|---|---|
x | torch.Tensor | A 3D tensor, the size should be (B, C, L1, L2, L3), where B is the batch size, C is the input channel number, and (L1, L2, L3) is the input data size. |
Returns
| Argument | Description |
|---|---|
y | A tensor,
|
Properties¶
nlayers¶
net.nlayers
The total number of convolutional layers along the depth of the network. This value would not take the fully-connected layer into consideration.
Examples¶
Example 1
1 2 3 4 5 | |
The number of convolutional layers along the depth is 32.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv3d-1 [-1, 64, 31, 32, 30] 24,000
InstanceNorm3d-2 [-1, 64, 31, 32, 30] 128
PReLU-3 [-1, 64, 31, 32, 30] 64
Conv3d-4 [-1, 64, 31, 32, 30] 4,096
InstanceNorm3d-5 [-1, 64, 31, 32, 30] 128
PReLU-6 [-1, 64, 31, 32, 30] 64
Conv3d-7 [-1, 64, 31, 32, 30] 110,592
InstanceNorm3d-8 [-1, 64, 31, 32, 30] 128
PReLU-9 [-1, 64, 31, 32, 30] 64
Conv3d-10 [-1, 64, 31, 32, 30] 4,096
_BlockBo...neckNd-11 [-1, 64, 31, 32, 30] 0
InstanceNorm3d-12 [-1, 64, 31, 32, 30] 128
PReLU-13 [-1, 64, 31, 32, 30] 64
Conv3d-14 [-1, 64, 31, 32, 30] 4,096
InstanceNorm3d-15 [-1, 64, 31, 32, 30] 128
PReLU-16 [-1, 64, 31, 32, 30] 64
Conv3d-17 [-1, 64, 16, 16, 15] 110,592
InstanceNorm3d-18 [-1, 64, 16, 16, 15] 128
PReLU-19 [-1, 64, 16, 16, 15] 64
Conv3d-20 [-1, 64, 16, 16, 15] 4,096
Conv3d-21 [-1, 64, 16, 16, 15] 4,096
InstanceNorm3d-22 [-1, 64, 16, 16, 15] 128
_BlockBo...neckNd-23 [-1, 64, 16, 16, 15] 0
_BlockResStkNd-24 [-1, 64, 16, 16, 15] 0
InstanceNorm3d-25 [-1, 64, 16, 16, 15] 128
PReLU-26 [-1, 64, 16, 16, 15] 64
Conv3d-27 [-1, 64, 16, 16, 15] 4,096
InstanceNorm3d-28 [-1, 64, 16, 16, 15] 128
PReLU-29 [-1, 64, 16, 16, 15] 64
Conv3d-30 [-1, 64, 16, 16, 15] 110,592
InstanceNorm3d-31 [-1, 64, 16, 16, 15] 128
PReLU-32 [-1, 64, 16, 16, 15] 64
Conv3d-33 [-1, 128, 16, 16, 15] 8,192
Conv3d-34 [-1, 128, 16, 16, 15] 8,192
InstanceNorm3d-35 [-1, 128, 16, 16, 15] 256
_BlockBo...neckNd-36 [-1, 128, 16, 16, 15] 0
InstanceNorm3d-37 [-1, 128, 16, 16, 15] 256
PReLU-38 [-1, 128, 16, 16, 15] 128
Conv3d-39 [-1, 128, 16, 16, 15] 16,384
InstanceNorm3d-40 [-1, 128, 16, 16, 15] 256
PReLU-41 [-1, 128, 16, 16, 15] 128
Conv3d-42 [-1, 128, 8, 8, 8] 442,368
InstanceNorm3d-43 [-1, 128, 8, 8, 8] 256
PReLU-44 [-1, 128, 8, 8, 8] 128
Conv3d-45 [-1, 128, 8, 8, 8] 16,384
Conv3d-46 [-1, 128, 8, 8, 8] 16,384
InstanceNorm3d-47 [-1, 128, 8, 8, 8] 256
_BlockBo...neckNd-48 [-1, 128, 8, 8, 8] 0
_BlockResStkNd-49 [-1, 128, 8, 8, 8] 0
InstanceNorm3d-50 [-1, 128, 8, 8, 8] 256
PReLU-51 [-1, 128, 8, 8, 8] 128
Conv3d-52 [-1, 128, 8, 8, 8] 16,384
InstanceNorm3d-53 [-1, 128, 8, 8, 8] 256
PReLU-54 [-1, 128, 8, 8, 8] 128
Conv3d-55 [-1, 128, 8, 8, 8] 442,368
InstanceNorm3d-56 [-1, 128, 8, 8, 8] 256
PReLU-57 [-1, 128, 8, 8, 8] 128
Conv3d-58 [-1, 256, 8, 8, 8] 32,768
Conv3d-59 [-1, 256, 8, 8, 8] 32,768
InstanceNorm3d-60 [-1, 256, 8, 8, 8] 512
_BlockBo...neckNd-61 [-1, 256, 8, 8, 8] 0
InstanceNorm3d-62 [-1, 256, 8, 8, 8] 512
PReLU-63 [-1, 256, 8, 8, 8] 256
Conv3d-64 [-1, 256, 8, 8, 8] 65,536
InstanceNorm3d-65 [-1, 256, 8, 8, 8] 512
PReLU-66 [-1, 256, 8, 8, 8] 256
Conv3d-67 [-1, 256, 4, 4, 4] 1,769,472
InstanceNorm3d-68 [-1, 256, 4, 4, 4] 512
PReLU-69 [-1, 256, 4, 4, 4] 256
Conv3d-70 [-1, 256, 4, 4, 4] 65,536
Conv3d-71 [-1, 256, 4, 4, 4] 65,536
InstanceNorm3d-72 [-1, 256, 4, 4, 4] 512
_BlockBo...neckNd-73 [-1, 256, 4, 4, 4] 0
_BlockResStkNd-74 [-1, 256, 4, 4, 4] 0
InstanceNorm3d-75 [-1, 256, 4, 4, 4] 512
PReLU-76 [-1, 256, 4, 4, 4] 256
Conv3d-77 [-1, 256, 4, 4, 4] 65,536
InstanceNorm3d-78 [-1, 256, 4, 4, 4] 512
PReLU-79 [-1, 256, 4, 4, 4] 256
Conv3d-80 [-1, 256, 4, 4, 4] 1,769,472
InstanceNorm3d-81 [-1, 256, 4, 4, 4] 512
PReLU-82 [-1, 256, 4, 4, 4] 256
Conv3d-83 [-1, 512, 4, 4, 4] 131,072
Conv3d-84 [-1, 512, 4, 4, 4] 131,072
InstanceNorm3d-85 [-1, 512, 4, 4, 4] 1,024
_BlockBo...neckNd-86 [-1, 512, 4, 4, 4] 0
InstanceNorm3d-87 [-1, 512, 4, 4, 4] 1,024
PReLU-88 [-1, 512, 4, 4, 4] 512
Conv3d-89 [-1, 512, 4, 4, 4] 262,144
InstanceNorm3d-90 [-1, 512, 4, 4, 4] 1,024
PReLU-91 [-1, 512, 4, 4, 4] 512
Conv3d-92 [-1, 512, 2, 2, 2] 7,077,888
InstanceNorm3d-93 [-1, 512, 2, 2, 2] 1,024
PReLU-94 [-1, 512, 2, 2, 2] 512
Conv3d-95 [-1, 512, 2, 2, 2] 262,144
Conv3d-96 [-1, 512, 2, 2, 2] 262,144
InstanceNorm3d-97 [-1, 512, 2, 2, 2] 1,024
_BlockBo...neckNd-98 [-1, 512, 2, 2, 2] 0
_BlockResStkNd-99 [-1, 512, 2, 2, 2] 0
InstanceNorm3d-100 [-1, 512, 2, 2, 2] 1,024
PReLU-101 [-1, 512, 2, 2, 2] 512
Conv3d-102 [-1, 512, 2, 2, 2] 262,144
InstanceNorm3d-103 [-1, 512, 2, 2, 2] 1,024
PReLU-104 [-1, 512, 2, 2, 2] 512
Conv3d-105 [-1, 512, 2, 2, 2] 7,077,888
InstanceNorm3d-106 [-1, 512, 2, 2, 2] 1,024
PReLU-107 [-1, 512, 2, 2, 2] 512
Conv3d-108 [-1, 1024, 2, 2, 2] 524,288
Conv3d-109 [-1, 1024, 2, 2, 2] 524,288
InstanceNorm3d-110 [-1, 1024, 2, 2, 2] 2,048
_BlockBo...eckNd-111 [-1, 1024, 2, 2, 2] 0
InstanceNorm3d-112 [-1, 1024, 2, 2, 2] 2,048
PReLU-113 [-1, 1024, 2, 2, 2] 1,024
Conv3d-114 [-1, 1024, 2, 2, 2] 1,048,576
InstanceNorm3d-115 [-1, 1024, 2, 2, 2] 2,048
PReLU-116 [-1, 1024, 2, 2, 2] 1,024
Conv3d-117 [-1, 1024, 1, 1, 1] 28,311,552
InstanceNorm3d-118 [-1, 1024, 1, 1, 1] 2,048
PReLU-119 [-1, 1024, 1, 1, 1] 1,024
Conv3d-120 [-1, 1024, 1, 1, 1] 1,048,576
Conv3d-121 [-1, 1024, 1, 1, 1] 1,048,576
InstanceNorm3d-122 [-1, 1024, 1, 1, 1] 2,048
_BlockBo...eckNd-123 [-1, 1024, 1, 1, 1] 0
_BlockResStkNd-124 [-1, 1024, 1, 1, 1] 0
Conv3d-125 [-1, 1024, 1, 1, 1] 28,312,576
Adaptive...ool3d-126 [-1, 1024, 1, 1, 1] 0
Linear-127 [-1, 32] 32,800
EncoderNet3d-128 [-1, 32] 0
================================================================
Total params: 81,566,240
Trainable params: 81,566,240
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.34
Forward/backward pass size (MB): 318.64
Params size (MB): 311.15
Estimated Total Size (MB): 630.13
----------------------------------------------------------------
Example 2
1 2 3 4 5 | |
The number of convolutional layers along the depth is 32.
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv3d-1 [-1, 64, 31, 32, 30] 24,000
InstanceNorm3d-2 [-1, 64, 31, 32, 30] 128
PReLU-3 [-1, 64, 31, 32, 30] 64
Conv3d-4 [-1, 64, 31, 32, 30] 4,096
InstanceNorm3d-5 [-1, 64, 31, 32, 30] 128
PReLU-6 [-1, 64, 31, 32, 30] 64
Conv3d-7 [-1, 64, 31, 32, 30] 110,592
InstanceNorm3d-8 [-1, 64, 31, 32, 30] 128
PReLU-9 [-1, 64, 31, 32, 30] 64
Conv3d-10 [-1, 64, 31, 32, 30] 4,096
_BlockBo...neckNd-11 [-1, 64, 31, 32, 30] 0
InstanceNorm3d-12 [-1, 64, 31, 32, 30] 128
PReLU-13 [-1, 64, 31, 32, 30] 64
Conv3d-14 [-1, 64, 31, 32, 30] 4,096
InstanceNorm3d-15 [-1, 64, 31, 32, 30] 128
PReLU-16 [-1, 64, 31, 32, 30] 64
Conv3d-17 [-1, 64, 16, 16, 15] 110,592
InstanceNorm3d-18 [-1, 64, 16, 16, 15] 128
PReLU-19 [-1, 64, 16, 16, 15] 64
Conv3d-20 [-1, 64, 16, 16, 15] 4,096
Conv3d-21 [-1, 64, 16, 16, 15] 4,096
InstanceNorm3d-22 [-1, 64, 16, 16, 15] 128
_BlockBo...neckNd-23 [-1, 64, 16, 16, 15] 0
_BlockResStkNd-24 [-1, 64, 16, 16, 15] 0
InstanceNorm3d-25 [-1, 64, 16, 16, 15] 128
PReLU-26 [-1, 64, 16, 16, 15] 64
Conv3d-27 [-1, 64, 16, 16, 15] 4,096
InstanceNorm3d-28 [-1, 64, 16, 16, 15] 128
PReLU-29 [-1, 64, 16, 16, 15] 64
Conv3d-30 [-1, 64, 16, 16, 15] 110,592
InstanceNorm3d-31 [-1, 64, 16, 16, 15] 128
PReLU-32 [-1, 64, 16, 16, 15] 64
Conv3d-33 [-1, 128, 16, 16, 15] 8,192
Conv3d-34 [-1, 128, 16, 16, 15] 8,192
InstanceNorm3d-35 [-1, 128, 16, 16, 15] 256
_BlockBo...neckNd-36 [-1, 128, 16, 16, 15] 0
InstanceNorm3d-37 [-1, 128, 16, 16, 15] 256
PReLU-38 [-1, 128, 16, 16, 15] 128
Conv3d-39 [-1, 128, 16, 16, 15] 16,384
InstanceNorm3d-40 [-1, 128, 16, 16, 15] 256
PReLU-41 [-1, 128, 16, 16, 15] 128
Conv3d-42 [-1, 128, 8, 8, 8] 442,368
InstanceNorm3d-43 [-1, 128, 8, 8, 8] 256
PReLU-44 [-1, 128, 8, 8, 8] 128
Conv3d-45 [-1, 128, 8, 8, 8] 16,384
Conv3d-46 [-1, 128, 8, 8, 8] 16,384
InstanceNorm3d-47 [-1, 128, 8, 8, 8] 256
_BlockBo...neckNd-48 [-1, 128, 8, 8, 8] 0
_BlockResStkNd-49 [-1, 128, 8, 8, 8] 0
InstanceNorm3d-50 [-1, 128, 8, 8, 8] 256
PReLU-51 [-1, 128, 8, 8, 8] 128
Conv3d-52 [-1, 128, 8, 8, 8] 16,384
InstanceNorm3d-53 [-1, 128, 8, 8, 8] 256
PReLU-54 [-1, 128, 8, 8, 8] 128
Conv3d-55 [-1, 128, 8, 8, 8] 442,368
InstanceNorm3d-56 [-1, 128, 8, 8, 8] 256
PReLU-57 [-1, 128, 8, 8, 8] 128
Conv3d-58 [-1, 256, 8, 8, 8] 32,768
Conv3d-59 [-1, 256, 8, 8, 8] 32,768
InstanceNorm3d-60 [-1, 256, 8, 8, 8] 512
_BlockBo...neckNd-61 [-1, 256, 8, 8, 8] 0
InstanceNorm3d-62 [-1, 256, 8, 8, 8] 512
PReLU-63 [-1, 256, 8, 8, 8] 256
Conv3d-64 [-1, 256, 8, 8, 8] 65,536
InstanceNorm3d-65 [-1, 256, 8, 8, 8] 512
PReLU-66 [-1, 256, 8, 8, 8] 256
Conv3d-67 [-1, 256, 4, 4, 4] 1,769,472
InstanceNorm3d-68 [-1, 256, 4, 4, 4] 512
PReLU-69 [-1, 256, 4, 4, 4] 256
Conv3d-70 [-1, 256, 4, 4, 4] 65,536
Conv3d-71 [-1, 256, 4, 4, 4] 65,536
InstanceNorm3d-72 [-1, 256, 4, 4, 4] 512
_BlockBo...neckNd-73 [-1, 256, 4, 4, 4] 0
_BlockResStkNd-74 [-1, 256, 4, 4, 4] 0
InstanceNorm3d-75 [-1, 256, 4, 4, 4] 512
PReLU-76 [-1, 256, 4, 4, 4] 256
Conv3d-77 [-1, 256, 4, 4, 4] 65,536
InstanceNorm3d-78 [-1, 256, 4, 4, 4] 512
PReLU-79 [-1, 256, 4, 4, 4] 256
Conv3d-80 [-1, 256, 4, 4, 4] 1,769,472
InstanceNorm3d-81 [-1, 256, 4, 4, 4] 512
PReLU-82 [-1, 256, 4, 4, 4] 256
Conv3d-83 [-1, 512, 4, 4, 4] 131,072
Conv3d-84 [-1, 512, 4, 4, 4] 131,072
InstanceNorm3d-85 [-1, 512, 4, 4, 4] 1,024
_BlockBo...neckNd-86 [-1, 512, 4, 4, 4] 0
InstanceNorm3d-87 [-1, 512, 4, 4, 4] 1,024
PReLU-88 [-1, 512, 4, 4, 4] 512
Conv3d-89 [-1, 512, 4, 4, 4] 262,144
InstanceNorm3d-90 [-1, 512, 4, 4, 4] 1,024
PReLU-91 [-1, 512, 4, 4, 4] 512
Conv3d-92 [-1, 512, 2, 2, 2] 7,077,888
InstanceNorm3d-93 [-1, 512, 2, 2, 2] 1,024
PReLU-94 [-1, 512, 2, 2, 2] 512
Conv3d-95 [-1, 512, 2, 2, 2] 262,144
Conv3d-96 [-1, 512, 2, 2, 2] 262,144
InstanceNorm3d-97 [-1, 512, 2, 2, 2] 1,024
_BlockBo...neckNd-98 [-1, 512, 2, 2, 2] 0
_BlockResStkNd-99 [-1, 512, 2, 2, 2] 0
InstanceNorm3d-100 [-1, 512, 2, 2, 2] 1,024
PReLU-101 [-1, 512, 2, 2, 2] 512
Conv3d-102 [-1, 512, 2, 2, 2] 262,144
InstanceNorm3d-103 [-1, 512, 2, 2, 2] 1,024
PReLU-104 [-1, 512, 2, 2, 2] 512
Conv3d-105 [-1, 512, 2, 2, 2] 7,077,888
InstanceNorm3d-106 [-1, 512, 2, 2, 2] 1,024
PReLU-107 [-1, 512, 2, 2, 2] 512
Conv3d-108 [-1, 1024, 2, 2, 2] 524,288
Conv3d-109 [-1, 1024, 2, 2, 2] 524,288
InstanceNorm3d-110 [-1, 1024, 2, 2, 2] 2,048
_BlockBo...eckNd-111 [-1, 1024, 2, 2, 2] 0
InstanceNorm3d-112 [-1, 1024, 2, 2, 2] 2,048
PReLU-113 [-1, 1024, 2, 2, 2] 1,024
Conv3d-114 [-1, 1024, 2, 2, 2] 1,048,576
InstanceNorm3d-115 [-1, 1024, 2, 2, 2] 2,048
PReLU-116 [-1, 1024, 2, 2, 2] 1,024
Conv3d-117 [-1, 1024, 1, 1, 1] 28,311,552
InstanceNorm3d-118 [-1, 1024, 1, 1, 1] 2,048
PReLU-119 [-1, 1024, 1, 1, 1] 1,024
Conv3d-120 [-1, 1024, 1, 1, 1] 1,048,576
Conv3d-121 [-1, 1024, 1, 1, 1] 1,048,576
InstanceNorm3d-122 [-1, 1024, 1, 1, 1] 2,048
_BlockBo...eckNd-123 [-1, 1024, 1, 1, 1] 0
_BlockResStkNd-124 [-1, 1024, 1, 1, 1] 0
Conv3d-125 [-1, 1024, 1, 1, 1] 28,312,576
EncoderNet3d-126 [-1, 1024, 1, 1, 1] 0
================================================================
Total params: 81,533,440
Trainable params: 81,533,440
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.34
Forward/backward pass size (MB): 318.64
Params size (MB): 311.03
Estimated Total Size (MB): 630.01
----------------------------------------------------------------
Last update: March 14, 2021