point transformer v3结构

weixin_45625419

最新推荐文章于 2026-05-10 16:36:39 发布

关注

VIPpoint transformer v3结构

分割结构：embedding->enc->decembedding->enc0->enc1->enc2->enc3->enc4->dec3->dec2->dec1->dec0embedding->block0->block1->dow->block0->block1->dow->block0->block1->dow->block0->block1-&gt

开通可读全文 ·575 人阅读

最低0.47元/天解锁文章

在这里插入图片描述

分割结构：

embedding->enc->dec

embedding->enc0->enc1->enc2->enc3->enc4->dec3->dec2->dec1->dec0

embedding
->block0->block1
->dow->block0->block1
->dow->block0->block1
->dow->block0->block1->block2->block3->block4->block5
->dow->block0->block1
->up->block0->block1
->up->block0->block1
->up->block0->block1
->up->block0->block1

embedding:SubMConv3d->BatchNorm1d->GELU

SubMConv3d：输入为上一步sparsify()生成的sparse_conv_feat，它由feat[108413, 6]和grid_coord[108413, 3]变化而来。
            经过这一步，feat变成[108413, 32]。
            feat中存着法向量和颜色，grid是体素化后的坐标。
BatchNorm1d：输入feat。更新了feat和sparse_conv_feat。
GELU：输入feat。更新了feat和sparse_conv_feat。

block:cpe->norm1->attn->norm2->mlp

cpe:SubMConv3d->Liner->LayerNorm

SubMConv3d：输入为sparse_conv_feat[108413, 32]，更新feat和sparse_conv_feat[108413, 32]。
Liner：输入为sparse_conv_feat[108413, 32]，更新feat和sparse_conv_feat[108413, 32]。
LayerNorm：输入为sparse_conv_feat[108413, 32]，更新feat和sparse_conv_feat[108413, 32]。

DefaultSegmentorV2(
  (seg_head): Linear(in_features=64, out_features=13, bias=True)
  (backbone): PointTransformerV3(
    (embedding): Embedding(
      (stem): PointSequential(
        (conv): SubMConv3d(6, 32, kernel_size=[5, 5, 5], stride=[1, 1, 1], padding=[1, 1, 1], dilation=[1, 1, 1], output_padding=[0, 0, 0], bias=False, algo=ConvAlgo.Native)
        (norm): BatchNorm1d(32, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
        (act): GELU(approximate='none')
      )
    )
    (enc): PointSequential(
      (enc0): PointSequential(
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(32, 32, kernel_size=[3, 3, 3], stride=[1, 1, 1], padding=[0, 0, 0], dilation=[1, 1, 1], output_padding=[0, 0, 0], algo=ConvAlgo.MaskImplicitGemm)
            (1): Linear(in_features=32, out_features=32, bias=True)
            (2): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=32, out_features=96, bias=True)
            (proj): Linear(in_features=32, out_features=32, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=32, out_features=128, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=128, out_features=32, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): Identity()
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(32, 32, kernel_size=[3, 3, 3], stride=[1, 1, 1], padding=[0, 0, 0], dilation=[1, 1, 1], output_padding=[0, 0, 0], algo=ConvAlgo.MaskImplicitGemm)
            (1): Linear(in_features=32, out_features=32, bias=True)
            (2): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=32, out_features=96, bias=True)
            (proj): Linear(in_features=32, out_features=32, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=32, out_features=128, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=128, out_features=32, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.023)
          )
        )
      )
      (enc1): PointSequential(
        (down): SerializedPooling(
          (proj): Linear(in_features=32, out_features=64, bias=True)
          (norm): PointSequential(
            (0): BatchNorm1d(64, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
          )
          (act): PointSequential(
            (0): GELU(approximate='none')
          )
        )
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(64, 64, kernel_size=[3, 3, 3], stride=[1, 1, 1], padding=[0, 0, 0], dilation=[1, 1, 1], output_padding=[0, 0, 0], algo=ConvAlgo.MaskImplicitGemm)
            (1): Linear(in_features=64, out_features=64, bias=True)
            (2): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=64, out_features=192, bias=True)
            (proj): Linear(in_features=64, out_features=64, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=64, out_features=256, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=256, out_features=64, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.046)
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(64, 64, kernel_size=[3, 3, 3], stride=[1, 1, 1], padding=[0, 0, 0], dilation=[1, 1, 1], output_padding=[0, 0, 0], algo=ConvAlgo.MaskImplicitGemm)
            (1): Linear(in_features=64, out_features=64, bias=True)
            (2): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=64, out_features=192, bias=True)
            (proj): Linear(in_features=64, out_features=64, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=64, out_features=256, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=256, out_features=64, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path)