DexiNed Base

Click a block to read its description, or select it with Tab and Enter.

DexiNed BaseEdge probability, input3 × 352 × 352 RGB, native eval. Shapes exclude batch.LibreYOLODexiNed BaseEdge probability, input3 × 352 × 352 RGB, native eval. Shapes exclude batch.Feature path and additionsRGB toBGR, multiply255Subtract103.939,116.779,123.68 BGR meansB1: DoubleConvBlock3/32/64 channels; first stride2;64×176×176B2: DoubleConvBlock64/128/128; stride1;no finalReLU;128×176×176D2: MaxPool3×3,stride2128 × 88 × 88; padding1+S1A2: 128×88×88B3: DenseBlock n=2Primary A2:128ch; residual PD3:256ch;88×88D3: MaxPool3×3,stride2256 × 44 × 44; padding1+S2A3: 256×44×44B4: DenseBlock n=3Primary A3:256ch; residual PD4:512ch;44×44D4: MaxPool3×3,stride2512 × 22 × 22; padding1+S3A4: 512×22×22B5: DenseBlock n=3Primary A4:512ch; residual PD5:512ch;22×22+S4A5: 512×22×22B6: DenseBlock n=3Primary A5:512ch; residual PD6:256ch;22×22Named S/PD inputs are computed in the adjacent projection panel.Cross-stage projectionsB164 input channelsConv1×1 + BatchNorm64 to128; stride2S1128 × 88 × 88A2128 input channelsConv1×1 + BatchNorm128 to256; stride2S2256 × 44 × 44A3256 input channelsConv1×1 + BatchNorm256 to512; stride2S3512 × 22 × 22A4512 input channelsConv1×1 + BatchNorm512 to512; stride1S4512 × 22 × 22D2128 input channelsConv1×1 + BatchNorm128 to256; stride1PD3256 × 88 × 88D2128 input channelsConv1×1 + BatchNorm128 to256; stride2R2256 × 44 × 44+D3R2Conv1×1 + BatchNorm256 to512; stride1PD4512 × 44 × 44D4512 input channelsConv1×1 + BatchNorm512 to512; stride1PD5512 × 22 × 22B5512 input channelsConv1×1 + BatchNorm512 to256; stride1PD6256 × 22 × 22D3+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 probabilityUpConvBlock from B164 channels; 1 upsample stepsL1: side logits1 × 352 × 352UpConvBlock from B2128 channels; 1 upsample stepsL2: side logits1 × 352 × 352UpConvBlock from B3256 channels; 2 upsample stepsL3: side logits1 × 352 × 352UpConvBlock from B4512 channels; 3 upsample stepsL4: side logits1 × 352 × 352UpConvBlock from B5512 channels; 4 upsample stepsL5: side logits1 × 352 × 352UpConvBlock from B6256 channels; 4 upsample stepsL6: side logits1 × 352 × 352L1L2L3L4L5L6Concat six side logits6 × 352 × 352Conv2d1×1, stride16 to1; bias=True; noBatchNormSigmoid1 × 352 × 352 edge probabilityCore returns six side logits + fusion; wrapper selects fusion.Dense layer: first 128 to256, repeat 256 to256Primary feature PFirst layer 128 channels; later 256ReLUBefore first convolutionConv2d 3×3128 to256 first; 256 to256 later; padding2BatchNorm2d256 channelsReLUConv2d 3×3256 to256; padding0BatchNorm2d256 channels+R256 chMultiply0.5Return (new primary, unchanged residual R); n=2Padding2 expands by2; padding0 restores the original grid.Dense layer: first 256 to512, repeat 512 to512Primary feature PFirst layer 256 channels; later 512ReLUBefore first convolutionConv2d 3×3256 to512 first; 512 to512 later; padding2BatchNorm2d512 channelsReLUConv2d 3×3512 to512; padding0BatchNorm2d512 channels+R512 chMultiply0.5Return (new primary, unchanged residual R); n=3Padding2 expands by2; padding0 restores the original grid.Dense layer: first 512 to512, repeat 512 to512Primary feature PFirst layer 512 channels; later 512ReLUBefore first convolutionConv2d 3×3512 to512 first; 512 to512 later; padding2BatchNorm2d512 channelsReLUConv2d 3×3512 to512; padding0BatchNorm2d512 channels+R512 chMultiply0.5Return (new primary, unchanged residual R); n=3Padding2 expands by2; padding0 restores the original grid.Dense layer: first 512 to256, repeat 256 to256Primary feature PFirst layer 512 channels; later 256ReLUBefore first convolutionConv2d 3×3512 to256 first; 256 to256 later; padding2BatchNorm2d256 channelsReLUConv2d 3×3256 to256; padding0BatchNorm2d256 channels+R256 chMultiply0.5Return (new primary, unchanged residual R); n=3Padding2 expands by2; padding0 restores the original grid.Stem and projection primitivesB1 DoubleConvBlockConv2d3×33 to32;s2,p1BatchNorm32ReLUConv2d3×332 to64;s1,p1BatchNorm64ReLUB2 DoubleConvBlockConv2d3×364 to128;s1,p1BatchNorm128ReLUConv2d3×3128 to128;s1,p1BatchNorm128SingleConvBlock: Conv2d1×1Numeric Cin/Cout/stride in projection panelBatchNorm2dNo activationUpConvBlock side1Conv1×164 to1ReLUTranspose2×21ch;s2,p0Every step doubles the grid; output1 × 352 × 352.UpConvBlock side2Conv1×1128 to1ReLUTranspose2×21ch;s2,p0Every step doubles the grid; output1 × 352 × 352.UpConvBlock side3Conv1×1256 to16ReLUTranspose4×416ch;s2,p1Conv1×116 to1ReLUTranspose4×41ch;s2,p1Every step doubles the grid; output1 × 352 × 352.UpConvBlock side4Conv1×1512 to16ReLUTranspose8×816ch;s2,p3Conv1×116 to16ReLUTranspose8×816ch;s2,p3Conv1×116 to1ReLUTranspose8×81ch;s2,p3Every step doubles the grid; output1 × 352 × 352.UpConvBlock side5Conv1×1512 to16ReLUTranspose16×1616ch;s2,p7Conv1×116 to16ReLUTranspose16×1616ch;s2,p7Conv1×116 to16ReLUTranspose16×1616ch;s2,p7Conv1×116 to1ReLUTranspose16×161ch;s2,p7Every step doubles the grid; output1 × 352 × 352.UpConvBlock side6Conv1×1256 to16ReLUTranspose16×1616ch;s2,p7Conv1×116 to16ReLUTranspose16×1616ch;s2,p7Conv1×116 to16ReLUTranspose16×1616ch;s2,p7Conv1×116 to1ReLUTranspose16×161ch;s2,p7Every step doubles the grid; output1 × 352 × 352.Source: libreyolo/models/dexined/nn.py and model.py. Revision a4d0ecc9e17f.libreyolo.com