docs

vision-rs / Core Concepts

The YOLO26 Model

vision_rs::models::yolo::yolo26 implements ultralytics/cfg/models/26/yolo26.yaml: a CSP-style backbone, an FPN neck, and one or two detection heads, with reg_max = 1 (YOLO26 drops DFL compared to earlier YOLO versions).

Variants

pub enum Yolo26Variant { N, S, M, L, XL }

Each variant has a depth/width/mc (max channels) scaling triple, returned by Yolo26Variant::config():

Variant depth width max channels
N 0.5 0.25 1024
S 0.5 0.50 1024
M 0.5 1.00 512
L 1.0 1.00 512
XL 1.0 1.50 512

Two helper functions scale the yaml’s base values by these multipliers: ch(base, width, mc) scales a channel count (capped at mc), and rep(base, depth) scales a block repeat count (minimum 1).

Backbone + FPN neck

build_neck constructs the shared backbone and neck, returning a closure Fn(SymTensor) -> (SymTensor, SymTensor, SymTensor) producing the three FPN feature maps (p3d, p4d, p5d) at strides 8/16/32, plus the three corresponding channel widths.

graph TD
    In[Input image] --> L0["conv (stride 2)"] --> L1["conv (stride 2)"]
    L1 --> L2[c3k2] --> L3["conv (stride 2)"]
    L3 --> L4[c3k2] --> P3["p3 (stride 8)"]
    P3 --> L5["conv (stride 2)"] --> L6[c3k2] --> P4["p4 (stride 16)"]
    P4 --> L7["conv (stride 2)"] --> L8[c3k2] --> L9[sppf] --> L10[c2psa] --> P5["p5 (stride 32)"]

    P5 --> Up1[upsample] --> Cat1[concat with p4]
    Cat1 --> L13[c3k2] --> Nk4[nk4]
    Nk4 --> Up2[upsample] --> Cat2[concat with p3]
    Cat2 --> L16[c3k2] --> P3D["p3d (to head)"]
    P3D --> L17["conv (stride 2)"] --> Cat3[concat with nk4]
    Cat3 --> L19[c3k2] --> P4D["p4d (to head)"]
    P4D --> L20["conv (stride 2)"] --> Cat4[concat with p5]
    Cat4 --> L22[c3k2_psa] --> P5D["p5d (to head)"]

Every layer is wrapped in a name_scope matching its yaml layer index (model.0 through model.22) — this is what lets weight loading map a pretrained checkpoint’s parameter names onto the traced graph.

Blocks (see models::yolo::yolo26::blocks):

  • conv — Conv2d + BatchNorm + activation.
  • c3k2 — the CSP bottleneck variant used throughout backbone/neck.
  • c2psa / c3k2_psa — cross-stage-partial blocks with position-sensitive attention (see Custom Kernels).
  • sppf — Spatial Pyramid Pooling - Fast.
  • upsample / concat — nearest-neighbor upsampling and channel-wise concatenation, used to build the FPN top-down path.

Detection heads

pub enum DetectHead { OneToMany, OneToOne }

OneToMany binds to the cv2/cv3 weight namespace (the dense training head); OneToOne binds to one2one_cv2/one2one_cv3 (the head used for inference, matching ultralytics eval-mode mAP). yolo26(nc, variant, head) builds a single-head forward closure producing raw DetectOutput { boxes, scores } (training-mode layout — apply detect-decode with the anchor grid/strides for inference-ready boxes; see Custom Kernels).

yolo26_dual(nc, variant) traces both heads in one graph, sharing the backbone/neck, for dual-assignment training — see Training.