point transformer v3结构

最新推荐文章于 2024-11-07 20:20:01 发布
最新推荐文章于 2024-11-07 20:20:01 发布
阅读量267 收藏 1
点赞数 3
文章标签： transformer 深度学习 人工智能
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/weixin_45625419/article/details/139679448
版权

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): PointSequential(
            (0): DropPath(drop_prob=0.069)
          )
        )
      )
      (enc2): PointSequential(
        (down): SerializedPooling(
          (proj): Linear(in_features=64, out_features=128, bias=True)
          (norm): PointSequential(
            (0): BatchNorm1d(128, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
          )
          (act): PointSequential(
            (0): GELU(approximate='none')
          )
        )
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(128, 128, 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=128, out_features=128, bias=True)
            (2): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=128, out_features=384, bias=True)
            (proj): Linear(in_features=128, out_features=128, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=128, out_features=512, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=512, out_features=128, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.092)
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(128, 128, 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=128, out_features=128, bias=True)
            (2): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=128, out_features=384, bias=True)
            (proj): Linear(in_features=128, out_features=128, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=128, out_features=512, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=512, out_features=128, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.115)
          )
        )
      )
      (enc3): PointSequential(
        (down): SerializedPooling(
          (proj): Linear(in_features=128, out_features=256, bias=True)
          (norm): PointSequential(
            (0): BatchNorm1d(256, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
          )
          (act): PointSequential(
            (0): GELU(approximate='none')
          )
        )
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.138)
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.162)
          )
        )
        (block2): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.185)
          )
        )
        (block3): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.208)
          )
        )
        (block4): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.231)
          )
        )
        (block5): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.254)
          )
        )
      )
      (enc4): PointSequential(
        (down): SerializedPooling(
          (proj): Linear(in_features=256, out_features=512, bias=True)
          (norm): PointSequential(
            (0): BatchNorm1d(512, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
          )
          (act): PointSequential(
            (0): GELU(approximate='none')
          )
        )
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(512, 512, 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=512, out_features=512, bias=True)
            (2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=512, out_features=1536, bias=True)
            (proj): Linear(in_features=512, out_features=512, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=512, out_features=2048, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=2048, out_features=512, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.277)
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(512, 512, 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=512, out_features=512, bias=True)
            (2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=512, out_features=1536, bias=True)
            (proj): Linear(in_features=512, out_features=512, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((512,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=512, out_features=2048, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=2048, out_features=512, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.300)
          )
        )
      )
    )
    (dec): PointSequential(
      (dec3): PointSequential(
        (up): SerializedUnpooling(
          (proj): PointSequential(
            (0): Linear(in_features=512, out_features=256, bias=True)
            (1): BatchNorm1d(256, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): GELU(approximate='none')
          )
          (proj_skip): PointSequential(
            (0): Linear(in_features=256, out_features=256, bias=True)
            (1): BatchNorm1d(256, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): GELU(approximate='none')
          )
        )
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.300)
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(256, 256, 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=256, out_features=256, bias=True)
            (2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=256, out_features=768, bias=True)
            (proj): Linear(in_features=256, out_features=256, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((256,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=256, out_features=1024, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=1024, out_features=256, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.257)
          )
        )
      )
      (dec2): PointSequential(
        (up): SerializedUnpooling(
          (proj): PointSequential(
            (0): Linear(in_features=256, out_features=128, bias=True)
            (1): BatchNorm1d(128, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): GELU(approximate='none')
          )
          (proj_skip): PointSequential(
            (0): Linear(in_features=128, out_features=128, bias=True)
            (1): BatchNorm1d(128, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): GELU(approximate='none')
          )
        )
        (block0): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(128, 128, 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=128, out_features=128, bias=True)
            (2): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=128, out_features=384, bias=True)
            (proj): Linear(in_features=128, out_features=128, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=128, out_features=512, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=512, out_features=128, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.214)
          )
        )
        (block1): Block(
          (cpe): PointSequential(
            (0): SubMConv3d(128, 128, 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=128, out_features=128, bias=True)
            (2): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (norm1): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (attn): SerializedAttention(
            (attn_drop): Dropout(p=0.0, inplace=False)
            (qkv): Linear(in_features=128, out_features=384, bias=True)
            (proj): Linear(in_features=128, out_features=128, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
            (rpe): RPE()
          )
          (norm2): PointSequential(
            (0): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          )
          (mlp): PointSequential(
            (0): MLP(
              (fc1): Linear(in_features=128, out_features=512, bias=True)
              (act): GELU(approximate='none')
              (fc2): Linear(in_features=512, out_features=128, bias=True)
              (drop): Dropout(p=0.0, inplace=False)
            )
          )
          (drop_path): PointSequential(
            (0): DropPath(drop_prob=0.171)
          )
        )
      )
      (dec1): PointSequential(
        (up): SerializedUnpooling(
          (proj): PointSequential(
            (0): Linear(in_features=128, out_features=64, bias=True)
            (1): BatchNorm1d(64, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): GELU(approximate='none')
          )
          (proj_skip): PointSequential(
            (0): Linear(in_features=64, out_features=64, bias=True)
            (1): BatchNorm1d(64, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): 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=0.129)
          )
        )
        (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): PointSequential(
            (0): DropPath(drop_prob=0.086)
          )
        )
      )
      (dec0): PointSequential(
        (up): SerializedUnpooling(
          (proj): PointSequential(
            (0): Linear(in_features=64, out_features=64, bias=True)
            (1): BatchNorm1d(64, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): GELU(approximate='none')
          )
          (proj_skip): PointSequential(
            (0): Linear(in_features=32, out_features=64, bias=True)
            (1): BatchNorm1d(64, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
            (2): 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=0.043)
          )
        )
        (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): PointSequential(
            (0): Identity()
          )
        )
      )
    )
  )
)
weixin_45625419
关注
Point Transformer V3阅读
qq_49130177的博客
04-19 1153
不创新注意力机制，而是权衡现有的准确性和效率(利用scale的力量)。模型性能更多受到规模的影响，而不是复杂的设计。将simple和efficient置于某些机制的准确性之上，从而实现可扩展性，而这些机制在scale后对整体性能来说是次要的。实现了显著的缩放，将感受野从16扩展到1024点，同时很高效(相比于其前身PTv2 ,处理速度提高了3倍,内存效率提高了10倍)。
Point Transformer V3 论文复现
qq_61247019的博客
06-28 2543
在Github上下载好代码文件进行环境配置这边由于使用Windows进行安装会出现很多BUG。
Point Transformer V3适合小白，看完就懂
博客简介
03-23 3986
PTv3设计原理介绍
Point Transformer V3: 更简单，更快，更强 详细解读
晨的博客
03-22 3735
主题: "更简单，更快，更强大"，v3版本在保持高性能的同时，着重优化了模型的简洁性和计算效率。 核心改进: 简化设计: v3版本对Point Transformer层进行了简化，减少了模型复杂度，提高了训练和推理速度。 效率提升: 引入了新的优化算法和技术，如高效的点采样和特征提取方法，显著提升了处理速度。 性能增强: 即便在简化模型结构的同时，v3版本通过更加精细的注意力机制和改进的网络架构，依然实现了在多个3D点云处理任务上的性能提升。
Point Transformer V3
w_t_f的博客
09-14 663
本篇论文不是寻求注意力机制的新的创新，而是平衡点云处理中准确性和效率的trade-offs。借鉴3D large-scale表示学习，作者认识到模型性能更受规模scale的影响，而不是复杂设计。因此作者提出Point Transformer V3 (PTv3)，优先考虑简单性和效率。例如，将KNN替换成序列化的邻域映射*。这一准则支持大规模scaling，将感受野receptive field从16个点扩展到1024个点，同时保持高效（与PTv2相比，处理速度提高了3倍，内存效率提高了10倍）。
point transformer v3复现及核心代码详解
weixin_50557558的博客
09-11 2118
point transformer v3复现及核心代码详解
探索点云变换的未来：Point Transformer V3 强势来袭
gitblog_00099的博客
06-08 658
探索点云变换的未来：Point Transformer V3 强势来袭 PointTransformerV3[CVPR 2024] Official project repository of Point Transformer V3 (PTv3)项目地址:https://gitcode.com/gh_mirrors/po/PointTransformerV3 在深度学习与三维感知领域，一个新星...
Point Transformer V3：更简单、更快、更强!
CV_Autobot的博客
12-24 983
点击下方卡片，关注“自动驾驶之心”公众号ADAS巨卷干货，即可获取>>点击进入→自动驾驶之心【点云处理】技术交流群“原标题：Point Transformer V3: Simpler, Faster, Stronger论文链接：https://arxiv.org/pdf/2312.10035.pdf代码链接：https://github.com/Pointcept/PointTrans...
Talk｜CVPR‘24 Oral：超越3D - Point Transformer V3中的多模态特征提取新构想
TechBeat人工智能社区博客
06-13 987
在本次Talk中，我们将超越3D感知与表征的范畴，从多模态数据特征提取的角度介绍我们被接收为CVPR 2024 Oral的工作Point Transformer V3 (PTv3) 的思想与设计。点云作为3D表征与感知的基础模态，其本身也是高维度稀疏非结构化数据的代表。将图像的每一个像素视为点，图像本身也可被视为点云，这佐证了这类数据结构的普适性。本次Talk将通过PTv3的两个核心思想——骨干网络设计的规模准则与非结构化数据的序列化技术，探究3D点云骨干网络作为一种多模态泛用特征提取器的构想与挑战。
基于Pytorch复现Point-Transformer，用于ShapeNet数据集点云分割
04-25
基于Point Transformers复现点云分割任务 ## 准备数据： 使用连接下载 **ShapeNet** 数据集 ## 训练： ```bash python train.py ...| Models | Accuracy | cat.mIOU...| Point Transformer (paper) | None | 0.837 |
（参考写法）Transformer-Based Visual Segmentation:A Survey
sinat_41942180的博客
11-04 384
视觉分割旨在将图像、视频帧或点云分割为多个片段或组。这种技术具有许多现实世界的应用，如自动驾驶、图像编辑、机器人感知和医学分析。（视觉分割定义+应用举例）在过去的十年里，基于深度学习的方法在这个领域取得了显著的进展。（之前是基于什么发展的）最近，Transformer成为一种基于自注意力机制的神经网络，最初设计用于自然语言处理，在各种视觉处理任务中明显超越了以往的卷积或循环方法。（提出本文要用的方法Transformer）
如何解决RNN在处理深层序列数据时遇到的如梯度消失、长期以来等问题
m0_73640344的博客
11-06 317
这些技术各自以不同的方式解决了RNN在处理深层序列数据时遇到的关键问题，如梯度消失和长期依赖。选择哪种技术取决于特定任务的需求，包括模型的复杂性、训练时间和性能要求。通过适当选择和配置这些改进的RNN架构，可以显著提高序列数据处理任务的效果和效率。
LLMs之LoLCATs：《LoLCATs: On Low-Rank Linearizing of Large Language Models》翻译与解读
近期请国内外头部出版社可尽快私信博主！——心比天高，仗剑走天涯，保持热爱，奔赴向梦想！低调，谦虚，自律，反思，成长，还算是比较正能量的博主，公益免费传播……内心特别想在AI界做出一些可以推进历史进程影响力的东西(兴趣使然，有点小情怀，也有点使命感呀)…
11-07 862
​ LLMs之LoLCATs：《LoLCATs: On Low-Rank Linearizing of Large Language Models》翻译与解读 目录 《LoLCATs: On Low-Rank Linearizing of Large Language Models》翻译与解读 Abstract 1 Introduction Conclusion Limitations and Future Work局限性和未来工作 《Lo
【语义分割|代码解析】CMTFNet-4: CNN and Multiscale Transformer Fusion Network 用于遥感图像分割！
985小水博的摸鱼日常
11-03 844
【语义分割|代码解析】CMTFNet-4: CNN and Multiscale Transformer Fusion Network 用于遥感图像分割！
深度学习基础练习：从pytorch API出发复现LSTM与LSTMP
SashiMoore的博客
11-07 339
对pytorch api的lstm与lstmp做了简单的代码复现
《深度学习》——深度学习基础知识（全连接神经网络）
最新发布
qq_59702185的博客
11-07 620
在介绍神经网络之前，我们先看下面这幅图：人工智能>机器学习>深度学习深度学习是机器学习的⼀个⼦集，也就是说深度学习是实现机器学习的⼀种⽅法。与机器学习算法的主要区别如下图所示：传统机器学习依赖于人工设计特征，并进行特征提取，而深度学习算法不需要人工，而是依赖算法自动提取特征，这也是深度学习被看做黑盒子、可解释性差的原因。卷积神经网络（CNN）循环神经网络（RNN）生成对抗网络（GAN）深度强化网络（DRN）
深度学习（十）：伦理与社会影响的深度剖析（10/10）
小相探索IT世界
11-03 1602
深度学习作为一项强大的技术，在带来诸多益处的同时，也引发了一系列伦理和社会问题。在伦理方面，数据隐私、偏见以及责任归属等挑战亟待解决。我们需要借鉴成功的法规案例，如欧盟的 GDPR，加强数据管理和隐私保护，努力消除算法偏见，并建立明确的责任归属机制。在社会影响方面，自动化虽然对传统职业造成冲击，但也创造了新的职业机会。个人、社会和教育机构都应积极应对，提升技能和加强人才培养，以适应就业结构的变化。
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值