DexiNed Base
Fit diagram
Read at 100%
Clear selection
Download SVG
Download PNG
Click a block to read its description, or select it with Tab and Enter.
DexiNed Base
Edge probability, input3 × 352 × 352 RGB, native eval. Shapes exclude batch.
LibreYOLO
DexiNed Base
Edge probability, input3 × 352 × 352 RGB, native eval. Shapes exclude batch.
Feature path and additions
RGB toBGR, multiply255
Subtract103.939,116.779,123.68 BGR means
B1: DoubleConvBlock
3/32/64 channels; first stride2;64×176×176
B2: DoubleConvBlock
64/128/128; stride1;no finalReLU;128×176×176
D2: MaxPool3×3,stride2
128 × 88 × 88; padding1
+
S1
A2: 128×88×88
B3: DenseBlock n=2
Primary A2:128ch; residual PD3:256ch;88×88
D3: MaxPool3×3,stride2
256 × 44 × 44; padding1
+
S2
A3: 256×44×44
B4: DenseBlock n=3
Primary A3:256ch; residual PD4:512ch;44×44
D4: MaxPool3×3,stride2
512 × 22 × 22; padding1
+
S3
A4: 512×22×22
B5: DenseBlock n=3
Primary A4:512ch; residual PD5:512ch;22×22
+
S4
A5: 512×22×22
B6: DenseBlock n=3
Primary A5:512ch; residual PD6:256ch;22×22
Named S/PD inputs are computed in the adjacent projection panel.
Cross-stage projections
B1
64 input channels
Conv1×1 + BatchNorm
64 to128; stride2
S1
128 × 88 × 88
A2
128 input channels
Conv1×1 + BatchNorm
128 to256; stride2
S2
256 × 44 × 44
A3
256 input channels
Conv1×1 + BatchNorm
256 to512; stride2
S3
512 × 22 × 22
A4
512 input channels
Conv1×1 + BatchNorm
512 to512; stride1
S4
512 × 22 × 22
D2
128 input channels
Conv1×1 + BatchNorm
128 to256; stride1
PD3
256 × 88 × 88
D2
128 input channels
Conv1×1 + BatchNorm
128 to256; stride2
R2
256 × 44 × 44
+
D3
R2
Conv1×1 + BatchNorm
256 to512; stride1
PD4
512 × 44 × 44
D4
512 input channels
Conv1×1 + BatchNorm
512 to512; stride1
PD5
512 × 22 × 22
B5
512 input channels
Conv1×1 + BatchNorm
512 to256; stride1
PD6
256 × 22 × 22
D3+R2 is an elementwise sum before the PD4 projection.
All projection convolutions have bias; BatchNorm follows.
No activation inside SingleConvBlock.
Six side logits and fused probability
UpConvBlock from B1
64 channels; 1 upsample steps
L1: side logits
1 × 352 × 352
UpConvBlock from B2
128 channels; 1 upsample steps
L2: side logits
1 × 352 × 352
UpConvBlock from B3
256 channels; 2 upsample steps
L3: side logits
1 × 352 × 352
UpConvBlock from B4
512 channels; 3 upsample steps
L4: side logits
1 × 352 × 352
UpConvBlock from B5
512 channels; 4 upsample steps
L5: side logits
1 × 352 × 352
UpConvBlock from B6
256 channels; 4 upsample steps
L6: side logits
1 × 352 × 352
L1
L2
L3
L4
L5
L6
Concat six side logits
6 × 352 × 352
Conv2d1×1, stride1
6 to1; bias=True; noBatchNorm
Sigmoid
1 × 352 × 352 edge probability
Core returns six side logits + fusion; wrapper selects fusion.
Dense layer: first 128 to256, repeat 256 to256
Primary feature P
First layer 128 channels; later 256
ReLU
Before first convolution
Conv2d 3×3
128 to256 first; 256 to256 later; padding2
BatchNorm2d
256 channels
ReLU
Conv2d 3×3
256 to256; padding0
BatchNorm2d
256 channels
+
R
256 ch
Multiply0.5
Return (new primary, unchanged residual R); n=2
Padding2 expands by2; padding0 restores the original grid.
Dense layer: first 256 to512, repeat 512 to512
Primary feature P
First layer 256 channels; later 512
ReLU
Before first convolution
Conv2d 3×3
256 to512 first; 512 to512 later; padding2
BatchNorm2d
512 channels
ReLU
Conv2d 3×3
512 to512; padding0
BatchNorm2d
512 channels
+
R
512 ch
Multiply0.5
Return (new primary, unchanged residual R); n=3
Padding2 expands by2; padding0 restores the original grid.
Dense layer: first 512 to512, repeat 512 to512
Primary feature P
First layer 512 channels; later 512
ReLU
Before first convolution
Conv2d 3×3
512 to512 first; 512 to512 later; padding2
BatchNorm2d
512 channels
ReLU
Conv2d 3×3
512 to512; padding0
BatchNorm2d
512 channels
+
R
512 ch
Multiply0.5
Return (new primary, unchanged residual R); n=3
Padding2 expands by2; padding0 restores the original grid.
Dense layer: first 512 to256, repeat 256 to256
Primary feature P
First layer 512 channels; later 256
ReLU
Before first convolution
Conv2d 3×3
512 to256 first; 256 to256 later; padding2
BatchNorm2d
256 channels
ReLU
Conv2d 3×3
256 to256; padding0
BatchNorm2d
256 channels
+
R
256 ch
Multiply0.5
Return (new primary, unchanged residual R); n=3
Padding2 expands by2; padding0 restores the original grid.
Stem and projection primitives
B1 DoubleConvBlock
Conv2d3×3
3 to32;s2,p1
BatchNorm
32
ReLU
Conv2d3×3
32 to64;s1,p1
BatchNorm
64
ReLU
B2 DoubleConvBlock
Conv2d3×3
64 to128;s1,p1
BatchNorm
128
ReLU
Conv2d3×3
128 to128;s1,p1
BatchNorm
128
SingleConvBlock: Conv2d1×1
Numeric Cin/Cout/stride in projection panel
BatchNorm2d
No activation
UpConvBlock side1
Conv1×1
64 to1
ReLU
Transpose2×2
1ch;s2,p0
Every step doubles the grid; output1 × 352 × 352.
UpConvBlock side2
Conv1×1
128 to1
ReLU
Transpose2×2
1ch;s2,p0
Every step doubles the grid; output1 × 352 × 352.
UpConvBlock side3
Conv1×1
256 to16
ReLU
Transpose4×4
16ch;s2,p1
Conv1×1
16 to1
ReLU
Transpose4×4
1ch;s2,p1
Every step doubles the grid; output1 × 352 × 352.
UpConvBlock side4
Conv1×1
512 to16
ReLU
Transpose8×8
16ch;s2,p3
Conv1×1
16 to16
ReLU
Transpose8×8
16ch;s2,p3
Conv1×1
16 to1
ReLU
Transpose8×8
1ch;s2,p3
Every step doubles the grid; output1 × 352 × 352.
UpConvBlock side5
Conv1×1
512 to16
ReLU
Transpose16×16
16ch;s2,p7
Conv1×1
16 to16
ReLU
Transpose16×16
16ch;s2,p7
Conv1×1
16 to16
ReLU
Transpose16×16
16ch;s2,p7
Conv1×1
16 to1
ReLU
Transpose16×16
1ch;s2,p7
Every step doubles the grid; output1 × 352 × 352.
UpConvBlock side6
Conv1×1
256 to16
ReLU
Transpose16×16
16ch;s2,p7
Conv1×1
16 to16
ReLU
Transpose16×16
16ch;s2,p7
Conv1×1
16 to16
ReLU
Transpose16×16
16ch;s2,p7
Conv1×1
16 to1
ReLU
Transpose16×16
1ch;s2,p7
Every step doubles the grid; output1 × 352 × 352.
Source: libreyolo/models/dexined/nn.py and model.py. Revision a4d0ecc9e17f.
libreyolo.com