EfficientNetV2 b0

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

EfficientNetV2 b0Classification, 224 × 224 input, 1,000 classes. Tensor sizes exclude batch.LibreYOLOEfficientNetV2 b0Classification, 224 × 224 input, 1,000 classes. Tensor sizes exclude batch.NetworkInput3 × 224 × 224Pad SAMEL/R=0/1, T/B=0/1Conv2d 3×3 / 232 × 112 × 112; p=0, g=1BatchNorm2d32 × 112 × 112SiLU32 × 112 × 112Layer 1: ConvBnAct16 × 112 × 112, repeat 1Layer 2: EdgeResidual32 × 56 × 56, repeat 1Layer 2: EdgeResidual32 × 56 × 56, repeat 1Layer 3: EdgeResidual48 × 28 × 28, repeat 1Layer 3: EdgeResidual48 × 28 × 28, repeat 1Layer 4: InvertedResidual96 × 14 × 14, repeat 1Layer 4: InvertedResidual96 × 14 × 14, repeat 2Layer 5: InvertedResidual112 × 14 × 14, repeat 1Layer 5: InvertedResidual112 × 14 × 14, repeat 4Layer 6: InvertedResidual192 × 7 × 7, repeat 1Layer 6: InvertedResidual192 × 7 × 7, repeat 7Conv2d 1×1 / 11280 × 7 × 7; p=0, g=1BatchNorm2d1280 × 7 × 7SiLU1280 × 7 × 7AdaptiveAvgPool2d1280 × 1 × 1Flatten1280Linear1280 to 1000Layer 1, unit 1ConvBnActInput 32 × 112 × 112Conv2d 3×3 / 116 × 112 × 112; p=1, g=1BatchNorm2d16 × 112 × 112SiLU16 × 112 × 112Output 16 × 112 × 112Repeat 1 timesLayer 2, unit 1EdgeResidualInput 16 × 112 × 112Pad SAMEL/R=0/1, T/B=0/1Conv2d 3×3 / 264 × 56 × 56; p=0, g=1BatchNorm2d64 × 56 × 56SiLU64 × 56 × 56Conv2d 1×1 / 132 × 56 × 56; p=0, g=1BatchNorm2d32 × 56 × 56Output 32 × 56 × 56Repeat 1 timesLayer 2, unit 2EdgeResidualInput 32 × 56 × 56Conv2d 3×3 / 1128 × 56 × 56; p=1, g=1BatchNorm2d128 × 56 × 56SiLU128 × 56 × 56Conv2d 1×1 / 132 × 56 × 56; p=0, g=1BatchNorm2d32 × 56 × 56+Output 32 × 56 × 56Repeat 1 timesLayer 3, unit 1EdgeResidualInput 32 × 56 × 56Pad SAMEL/R=0/1, T/B=0/1Conv2d 3×3 / 2128 × 28 × 28; p=0, g=1BatchNorm2d128 × 28 × 28SiLU128 × 28 × 28Conv2d 1×1 / 148 × 28 × 28; p=0, g=1BatchNorm2d48 × 28 × 28Output 48 × 28 × 28Repeat 1 timesLayer 3, unit 2EdgeResidualInput 48 × 28 × 28Conv2d 3×3 / 1192 × 28 × 28; p=1, g=1BatchNorm2d192 × 28 × 28SiLU192 × 28 × 28Conv2d 1×1 / 148 × 28 × 28; p=0, g=1BatchNorm2d48 × 28 × 28+Output 48 × 28 × 28Repeat 1 timesLayer 4, unit 1InvertedResidualInput 48 × 28 × 28Conv2d 1×1 / 1192 × 28 × 28; p=0, g=1BatchNorm2d192 × 28 × 28SiLU192 × 28 × 28Pad SAMEL/R=0/1, T/B=0/1Conv2d 3×3 / 2192 × 14 × 14; p=0, g=192BatchNorm2d192 × 14 × 14SiLU192 × 14 × 14Spatial mean192 × 1 × 1Conv2d 1×1 / 112 × 1 × 1; p=0, g=1SiLU12 × 1 × 1Conv2d 1×1 / 1192 × 1 × 1; p=0, g=1Sigmoid192 × 1 × 1Multiply gate with feature192 × 14 × 14Conv2d 1×1 / 196 × 14 × 14; p=0, g=1BatchNorm2d96 × 14 × 14Output 96 × 14 × 14Repeat 1 timesLayer 4, unit 2InvertedResidualInput 96 × 14 × 14Conv2d 1×1 / 1384 × 14 × 14; p=0, g=1BatchNorm2d384 × 14 × 14SiLU384 × 14 × 14Conv2d 3×3 / 1384 × 14 × 14; p=1, g=384BatchNorm2d384 × 14 × 14SiLU384 × 14 × 14Spatial mean384 × 1 × 1Conv2d 1×1 / 124 × 1 × 1; p=0, g=1SiLU24 × 1 × 1Conv2d 1×1 / 1384 × 1 × 1; p=0, g=1Sigmoid384 × 1 × 1Multiply gate with feature384 × 14 × 14Conv2d 1×1 / 196 × 14 × 14; p=0, g=1BatchNorm2d96 × 14 × 14+Output 96 × 14 × 14Repeat 2 timesLayer 5, unit 1InvertedResidualInput 96 × 14 × 14Conv2d 1×1 / 1576 × 14 × 14; p=0, g=1BatchNorm2d576 × 14 × 14SiLU576 × 14 × 14Conv2d 3×3 / 1576 × 14 × 14; p=1, g=576BatchNorm2d576 × 14 × 14SiLU576 × 14 × 14Spatial mean576 × 1 × 1Conv2d 1×1 / 124 × 1 × 1; p=0, g=1SiLU24 × 1 × 1Conv2d 1×1 / 1576 × 1 × 1; p=0, g=1Sigmoid576 × 1 × 1Multiply gate with feature576 × 14 × 14Conv2d 1×1 / 1112 × 14 × 14; p=0, g=1BatchNorm2d112 × 14 × 14Output 112 × 14 × 14Repeat 1 timesLayer 5, unit 2InvertedResidualInput 112 × 14 × 14Conv2d 1×1 / 1672 × 14 × 14; p=0, g=1BatchNorm2d672 × 14 × 14SiLU672 × 14 × 14Conv2d 3×3 / 1672 × 14 × 14; p=1, g=672BatchNorm2d672 × 14 × 14SiLU672 × 14 × 14Spatial mean672 × 1 × 1Conv2d 1×1 / 128 × 1 × 1; p=0, g=1SiLU28 × 1 × 1Conv2d 1×1 / 1672 × 1 × 1; p=0, g=1Sigmoid672 × 1 × 1Multiply gate with feature672 × 14 × 14Conv2d 1×1 / 1112 × 14 × 14; p=0, g=1BatchNorm2d112 × 14 × 14+Output 112 × 14 × 14Repeat 4 timesLayer 6, unit 1InvertedResidualInput 112 × 14 × 14Conv2d 1×1 / 1672 × 14 × 14; p=0, g=1BatchNorm2d672 × 14 × 14SiLU672 × 14 × 14Pad SAMEL/R=0/1, T/B=0/1Conv2d 3×3 / 2672 × 7 × 7; p=0, g=672BatchNorm2d672 × 7 × 7SiLU672 × 7 × 7Spatial mean672 × 1 × 1Conv2d 1×1 / 128 × 1 × 1; p=0, g=1SiLU28 × 1 × 1Conv2d 1×1 / 1672 × 1 × 1; p=0, g=1Sigmoid672 × 1 × 1Multiply gate with feature672 × 7 × 7Conv2d 1×1 / 1192 × 7 × 7; p=0, g=1BatchNorm2d192 × 7 × 7Output 192 × 7 × 7Repeat 1 timesLayer 6, unit 2InvertedResidualInput 192 × 7 × 7Conv2d 1×1 / 11152 × 7 × 7; p=0, g=1BatchNorm2d1152 × 7 × 7SiLU1152 × 7 × 7Conv2d 3×3 / 11152 × 7 × 7; p=1, g=1152BatchNorm2d1152 × 7 × 7SiLU1152 × 7 × 7Spatial mean1152 × 1 × 1Conv2d 1×1 / 148 × 1 × 1; p=0, g=1SiLU48 × 1 × 1Conv2d 1×1 / 11152 × 1 × 1; p=0, g=1Sigmoid1152 × 1 × 1Multiply gate with feature1152 × 7 × 7Conv2d 1×1 / 1192 × 7 × 7; p=0, g=1BatchNorm2d192 × 7 × 7+Output 192 × 7 × 7Repeat 7 timesRepeated units are grouped only when operation parameters, tensor shapes and shortcut behavior match.Source: libreyolo/models/efficientnetv2/nn.py. Revision a4d0ecc9e17f.libreyolo.com