YOLOv1-B

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

YOLOv1-BDetection; 448 × 448 RGB; 20 classes; batch 1; unfused eval. Every cfg layer is shown.LibreYOLOYOLOv1-BDetection; 448 × 448 RGB; 20 classes; batch 1; unfused eval. Every cfg layer is shown.L = cfg layer index. Matching F labels continue the same tensor across columns. Full routes remain visible.Layers 0 to 15L0 Conv 7×7, s=2 + norm + leaky64 × 224 × 224L1 MaxPool2d 2×2, s=264 × 112 × 112L2 Conv 3×3, s=1 + norm + leaky192 × 112 × 112L3 MaxPool2d 2×2, s=2192 × 56 × 56L4 Conv 1×1, s=1 + norm + leaky128 × 56 × 56L5 Conv 3×3, s=1 + norm + leaky256 × 56 × 56L6 Conv 1×1, s=1 + norm + leaky256 × 56 × 56L7 Conv 3×3, s=1 + norm + leaky512 × 56 × 56L8 MaxPool2d 2×2, s=2512 × 28 × 28L9 Conv 1×1, s=1 + norm + leaky256 × 28 × 28L10 Conv 3×3, s=1 + norm + leaky512 × 28 × 28L11 Conv 1×1, s=1 + norm + leaky256 × 28 × 28L12 Conv 3×3, s=1 + norm + leaky512 × 28 × 28L13 Conv 1×1, s=1 + norm + leaky256 × 28 × 28L14 Conv 3×3, s=1 + norm + leaky512 × 28 × 28L15 Conv 1×1, s=1 + norm + leaky256 × 28 × 28Input: 3 × 448 × 448F15Layers 16 to 31L16 Conv 3×3, s=1 + norm + leaky512 × 28 × 28L17 Conv 1×1, s=1 + norm + leaky512 × 28 × 28L18 Conv 3×3, s=1 + norm + leaky1024 × 28 × 28L19 MaxPool2d 2×2, s=21024 × 14 × 14L20 Conv 1×1, s=1 + norm + leaky512 × 14 × 14L21 Conv 3×3, s=1 + norm + leaky1024 × 14 × 14L22 Conv 1×1, s=1 + norm + leaky512 × 14 × 14L23 Conv 3×3, s=1 + norm + leaky1024 × 14 × 14L24 Conv 3×3, s=1 + norm + leaky1024 × 14 × 14L25 Conv 3×3, s=2 + norm + leaky1024 × 7 × 7L26 Conv 3×3, s=1 + norm + leaky1024 × 7 × 7L27 Conv 3×3, s=1 + norm + leaky1024 × 7 × 7L28 Local 3×3, leaky256 × 7 × 7L29 Dropout (identity in eval)256 × 7 × 7L30 Flatten + Linear1470L31 Raw detection head1470F15Convolution blockConv2dk, stride and channels from layer labelDarknet normalizationOnly layers labeled + normActivationleaky: LeakyReLU(0.1); linear: identityNorm layers: bias=False. Without norm: bias=True.Mish uses its separate definition when present.Normalization and poolingSubtract running meanx - meanDivide by standard deviationsqrt(running variance) + 0.000001Multiply scale, add biasLearned channel-wise affine transformMaxPool2d: p = floor((k - 1) / 2).k=2, s=1: pad right/bottom with negative infinity.That pool preserves its input spatial size.Fully connected headFlatten C, H, W12,544 features; channel-majorLinear12,544 inputs; 1,470 outputs; biasActivationIdentity for the final prediction vectorLocally connected layerUnfold image patcheskernel 3; padding 1; stride 1Per-location matrix multiply49 separate banks; 256 × 9216 eachAdd per-location bias256 × 49 biasesReshape + LeakyReLU(0.1)256 × 7 × 7Detection decoding (postprocessing)Raw head outputs: 1470Raw prediction tensorField selection is explicit in the three branchesSelect xy; add grid; divide by 7Centers scaled to 448 pixelsSelect wh; square dimensionsScale dimensions to 448 pixelsSelect confidence and class valuesMultiply; 20 scores for each of 98 boxesStack cx, cy, width, heightConvert to corner coordinatesConfidence filter + class NMSThen invert resize to the original imageVOC 20 classes. No anchors. Shared class values per 7 × 7 cell; 2 boxes per cell. Stretch preprocessing.Source: darknet/cfgs/yolov1.cfg; darknet/net.py; darknet/blocks.py. Revision a4d0ecc9e17f.libreyolo.com