LingBot-Vision S

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

LingBot-Vision SSemantic segmentation, 150 classes, native eval, input 3 × 512 × 512. Shapes exclude batch.LibreYOLOLingBot-Vision SSemantic segmentation, 150 classes, native eval, input 3 × 512 × 512. Shapes exclude batch.Backbone and semantic probeRGB + ImageNet normalization3 × 512 × 512Conv2d patch embedding 16×163 to 384; stride16; 32 × 32 patch gridFlatten 1024 patches1024 × 384Prepend CLS + 4 storage tokens1029 × 384; learnable prefixSelf-attention block, n=126 heads; head dimension64Final LayerNorm384 channels; epsilon1e-5Select patch tokensDiscard CLS and 4 storage tokens; 1024 × 384Restore patch grid384 × 32 × 32Conv2d 1×1 linear probe384 to 150; bias=Truesemantic_logits150 × 32 × 32Network forward returns patch-grid logits in a dictionary.Caller-level resizing is outside this network graph.Mask token is stored for checkpoint compatibility.Masked-token pretraining is not executed by this port.g is implemented but unpublished in the public family.Self-attention blockInput tokens1029 × 384LayerNorm384 channels; epsilon1e-5Rotary self-attention6 heads, width64LayerScale384 learned scalars; init1e-5+LayerNorm384 channels; epsilon1e-5Linear384 to 1536GELULinear1536 to 384LayerScale384 learned scalars; init1e-5+Output tokens1029 × 384No stochastic depth or dropout in this inference port.s/b/l use GELU MLP; g uses the separately drawn SwiGLU.Rotary self-attentionFused QKV linear384 to 1152; K bias masked to zeroSplit Q, K, V and heads6 × 1029 × 64 for each tensorApply RoPE to Q and KOnly 1024 patch tokens; keep prefixV unchanged1029 tokensQ × transpose(K) / 8Each head: 1029 × 1029Softmax over keysAll prefix and patch tokens attendAttention weights × V64 channels per headConcat heads1029 × 384Output linear384 to 384; bias=TrueQ/K rotation uses fp32 sine/cosine tables.Q/K convert back to input dtype before attention.Head dimension is64 for every registered size.Q,V biases exist for s/b/l; K bias is masked to zero.The giant g configuration disables fused QKV bias.Axial RoPE on patch tokens32×32 patch-center coordinatesNormalize each axis independently to [-1,1]Periods and angles16 frequencies per axis; base100; 2π × coordinate / periodSine and cosine tables1024 × 64; duplicate axial angle vectorRotate half of Q or KSplit [x1,x2], concatenate [-x2,x1]Multiply original x by cosineElementwiseMultiply rotated x by sineElementwise+x_rotated = x*cos + rotate_half(x)*sin. Prefix tokens bypass this rotation.Coordinates and periods construct positions; there is no learned absolute position table.MLP family configurationSize D (width) N (blocks) A (heads)s 384 12 6b 768 12 12l 1024 24 16Every MLP expands width by4 with GELU.s:384/1536; b:768/3072; l:1024/4096.g:1536 width,40 blocks,24 heads, SwiGLU4096.g has a separate concrete view because its FFN differs.Source: libreyolo/models/lingbotvision/nn.py and model.py. Revision a4d0ecc9e17f.libreyolo.com