Skip to content

Docs for NetsPresso Trainer

RTMCC

nota-netspresso/netspresso-trainer

RTMCC¶

RTMCC head based on RTMPose: Real-Time Multi-Person Pose Estimation based on MMPose.

Compatibility matrix¶

Supporting backbones	Supporting necks	torch.fx	NetsPresso
To be updated ...	To be updated ...	To be updated ...	To be updated ...

Field list¶

Field	Description
`name`	(str) Name must be "rtmcc" to use `RTMCC` head.
`params.conv_kernel`	(int) Kernel size of convolution layer.
`params.attention_channels`	(int) Dimension of gated attention unit.
`params.attention_act_type`	(str) Activation type of gated attention unit.
`params.attention_pos_enc`	(bool) Whether to use rotary position embedding for gated attention unit.
`params.s`	(int) Self attention feature dimension of gated attention unit.
`params.expansion_factor`	(int) Expansion factor of gated attention unit.
`params.dropout_rate`	(float) Dropout rate of gated attention unit.
`params.drop_path`	(float) Drop path rate of gated attention unit.
`params.use_rel_bias`	(bool) Whether to use relative bias for gated attention unit.
`params.simcc_split_ratio`	(float) Split ratio of pixels.
`params.target_size`	(list) Original input image size.
`params.backbone_stride`	(int) Stride of input feature from original image produced by backbone.

Model configuration example¶

RTMCC head

model:
  architecture:
    head:
    name: rtmcc
    params:
      conv_kernel: 7
      attention_channels: 256
      attention_act_type: 'silu'
      attention_pos_enc: False
      s: 128
      expansion_factor: 2
      dropout_rate: 0.
      drop_path: 0.
      use_rel_bias: False
      simcc_split_ratio: 2.
      target_size: [256, 256]
      backbone_stride: 32

RTMPose