FOMO-M

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

FOMO-MPoint detection; 1 foreground class + background; unfused PyTorch eval. Sizes exclude batch (batch = 1).LibreYOLOFOMO-MPoint detection; 1 foreground class + background; unfused PyTorch eval. Sizes exclude batch (batch = 1).Truncated MobileNetV2Input3 × 192 × 192ConvBNReLU6 3×3, s=216 × 96 × 96InvertedResidual (expansion 1)8 × 96 × 96InvertedResidual (block 1, s=2)16 × 48 × 48InvertedResidual (block 2)16 × 48 × 48InvertedResidual (block 3, s=2)16 × 24 × 24InvertedResidual (block 4)16 × 24 × 24InvertedResidual (block 5)16 × 24 × 24ConvBNReLU6 1×196 × 24 × 24No box-regression tower or anchor decoding.Feature FPoint head and postprocessingFeature map96 × 24 × 24Conv2d 1×1, bias=True2 × 24 × 24Softmax over 2 classes2 × 24 × 24Foreground probability1 × 24 × 24Threshold, sort scoresDescending foreground confidenceGrid-radius suppressionDefault radius = 1 cellGrid-cell centersx + 0.5; y + 0.5Rescale to original imageRows: x, y, class, confidenceThe neural network ends at logits. Softmax and point selection run in utils.postprocess().Random-weight CPU output: 1 × 2 × 24 × 24. Input resolution is fixed when building the model.Feature F from backboneInitial blockInput16 channelsStatic zero paddingL, R, T, B: (1, 1, 1, 1)Depthwise Conv2d 3×316 ch; groups=16; s=1BatchNorm2d + ReLU6eps=0.001Conv2d 1×18 ch; bias=FalseBatchNorm2deps=0.001; no activationNo residual connection.Block 1Input8 channelsConv2d 1×148 ch; bias=FalseBatchNorm2d + ReLU6eps=0.001Static zero paddingL, R, T, B: (0, 1, 0, 1)Depthwise Conv2d 3×348 ch; groups=48; s=2BatchNorm2d + ReLU6eps=0.001Conv2d 1×116 ch; bias=FalseBatchNorm2deps=0.001; no activationNo residual connection.Block 2Input16 channelsConv2d 1×196 ch; bias=FalseBatchNorm2d + ReLU6eps=0.001Static zero paddingL, R, T, B: (1, 1, 1, 1)Depthwise Conv2d 3×396 ch; groups=96; s=1BatchNorm2d + ReLU6eps=0.001Conv2d 1×116 ch; bias=FalseBatchNorm2deps=0.001; no activation+identityResidual enabled.Block 3Input16 channelsConv2d 1×196 ch; bias=FalseBatchNorm2d + ReLU6eps=0.001Static zero paddingL, R, T, B: (0, 1, 0, 1)Depthwise Conv2d 3×396 ch; groups=96; s=2BatchNorm2d + ReLU6eps=0.001Conv2d 1×116 ch; bias=FalseBatchNorm2deps=0.001; no activationNo residual connection.Blocks 4 and 5Input16 channelsConv2d 1×196 ch; bias=FalseBatchNorm2d + ReLU6eps=0.001Static zero paddingL, R, T, B: (1, 1, 1, 1)Depthwise Conv2d 3×396 ch; groups=96; s=1BatchNorm2d + ReLU6eps=0.001Conv2d 1×116 ch; bias=FalseBatchNorm2deps=0.001; no activation+identityEach block has its own weights.ConvBNReLU6Static zero paddingSAME for the model input geometryConv2dk and stride given by stageBatchNorm2deps=0.001ReLU6Clamp activations to [0, 6]Depthwise uses one group per input channel.Projection Conv2d has no ReLU6.Source-verified architecture. CPU shape evidence covers s, m, l with no downloaded weights. Point postprocessing is shown separately.Source: models/fomo/nn.py; models/fomo/utils.py. Revision a4d0ecc9e17f.libreyolo.com