Skip to content

YOLOPAFPN

YOLOPAFPN based on YOLOX: Exceeding YOLO Series in 2021.

YOLOPAFPN is a modified PAFPN for YOLOX model. Therefore, although YOLOPAFP is compatible with various backbones, we recommend to use it when constructing YOLOX models. The size is determined by dep_mul value, which defines the repetition of CSPLayers.

Compatibility matrix

Supporting backbones Supporting heads torch.fx NetsPresso
ResNet
MobileNetV3
MixNet
CSPDarkNet
MobileViT
MixTransformer
EfficientFormer 		ALLMLPDecoder
AnchorDecoupledHead
AnchorFreeDecoupledHead 		Supported Supported

Field list

Field Description
name (str) Name must be "yolopafpn" to use YOLOPAFPN neck.
params.depthwise (bool) Whether to enable depthwise convolution for the YOLOPAFPN neck.
params.dep_mul (int) Multiplying factor determining the repetition count of CSPLayer in the backbone.
params.act_type (int) Type of activation function for the model. Supporting activation functions are described in [here].

Model configuration examples

PAFPN for YOLOX-nano 
model:
  architecture:
    neck:
      name: yolopafpn
      params:
        depthwise: True
        dep_mul: 0.33
        act_type: "silu"
PAFPN for YOLOX-s 
model:
  architecture:
    neck:
      name: yolopafpn
      params:
        depthwise: False
        dep_mul: 0.33
        act_type: "silu"