共计 11646 个字符,预计需要花费 30 分钟才能阅读完成。
ConvNext 论文提出了一种新的基于卷积的架构,不仅超过了基于 Transformer 的模型(如 Swin),而且能够随着数据量的减少而扩大!明天咱们应用 Pytorch 来对其进行复现。下图显示了针对不同数据集 / 模型大小的 ConvNext 准确度。
作者首先采纳家喻户晓的 ResNet 架构,并依据过来十年中的新最佳实际和发现对其进行迭代改良。作者专一于 Swin-Transformer,并亲密关注其设计。这篇论文咱们在以前也举荐过,如果你们有浏览过,咱们强烈推荐浏览它:)
下图显示了所有各种改良以及每一项改良之后的各自性能。
论文将设计的路线图分为两局部:宏观设计和宏观设计。宏观设计是从高层次的角度所做的所有扭转,例如架构的设计,而微设计更多的是对于细节的,例如激活函数,归一化等。
上面咱们将从一个经典的 BottleNeck 块开始,并应用 pytorch 一一实现论文中说到的每个更改。
从 ResNet 开始
ResNet 由一个一个的残差(BottleNeck)块,咱们就从这里开始。
fromtorchimportnn
fromtorchimportTensor
fromtypingimportList
classConvNormAct(nn.Sequential):
"""A little util layer composed by (conv) -> (norm) -> (act) layers."""
def__init__(
self,
in_features: int,
out_features: int,
kernel_size: int,
norm = nn.BatchNorm2d,
act = nn.ReLU,
**kwargs
):
super().__init__(
nn.Conv2d(
in_features,
out_features,
kernel_size=kernel_size,
padding=kernel_size//2,
**kwargs
),
norm(out_features),
act(),)
classBottleNeckBlock(nn.Module):
def__init__(
self,
in_features: int,
out_features: int,
reduction: int = 4,
stride: int = 1,
):
super().__init__()
reduced_features = out_features//reduction
self.block = nn.Sequential(
# wide -> narrow
ConvNormAct(in_features, reduced_features, kernel_size=1, stride=stride, bias=False),
# narrow -> narrow
ConvNormAct(reduced_features, reduced_features, kernel_size=3, bias=False),
# narrow -> wide
ConvNormAct(reduced_features, out_features, kernel_size=1, bias=False, act=nn.Identity),
)
self.shortcut = (
nn.Sequential(
ConvNormAct(in_features, out_features, kernel_size=1, stride=stride, bias=False)
)
ifin_features!= out_features
elsenn.Identity())
self.act = nn.ReLU()
defforward(self, x: Tensor) ->Tensor:
res = x
x = self.block(x)
res = self.shortcut(res)
x += res
x = self.act(x)
returnx
看看下面代码是否无效
importtorch
x = torch.rand(1, 32, 7, 7)
block = BottleNeckBlock(32, 64)
block(x).shape
#torch.Size([1, 64, 7, 7])
上面开始定义 Stage,Stage 也叫阶段是残差块的汇合。每个阶段通常将输出下采样 2 倍
classConvNexStage(nn.Sequential):
def__init__(self, in_features: int, out_features: int, depth: int, stride: int = 2, **kwargs):
super().__init__(
# downsample is done here
BottleNeckBlock(in_features, out_features, stride=stride, **kwargs),
*[BottleNeckBlock(out_features, out_features, **kwargs)
for_inrange(depth-1)
],
)
测试
stage = ConvNexStage(32, 64, depth=2)
stage(x).shape
#torch.Size([1, 64, 4, 4])
咱们曾经将输出是从 7×7 缩小到 4×4。
ResNet 也有所谓的 stem,这是模型中对输出图像进行大量下采样的第一层。
classConvNextStem(nn.Sequential):
def__init__(self, in_features: int, out_features: int):
super().__init__(
ConvNormAct(in_features, out_features, kernel_size=7, stride=2),
nn.MaxPool2d(kernel_size=3, stride=2, padding=1),
)
当初咱们能够定义 ConvNextEncoder 来拼接各个阶段,并将图像作为输出生成最终嵌入。
classConvNextEncoder(nn.Module):
def__init__(
self,
in_channels: int,
stem_features: int,
depths: List[int],
widths: List[int],
):
super().__init__()
self.stem = ConvNextStem(in_channels, stem_features)
in_out_widths = list(zip(widths, widths[1:]))
self.stages = nn.ModuleList(
[ConvNexStage(stem_features, widths[0], depths[0], stride=1),
*[ConvNexStage(in_features, out_features, depth)
for (in_features, out_features), depthinzip(in_out_widths, depths[1:]
)
],
]
)
defforward(self, x):
x = self.stem(x)
forstageinself.stages:
x = stage(x)
returnx
测试后果如下:
image = torch.rand(1, 3, 224, 224)
encoder = ConvNextEncoder(in_channels=3, stem_features=64, depths=[3,4,6,4], widths=[256, 512, 1024, 2048])
encoder(image).shape
#torch.Size([1, 2048, 7, 7])
当初咱们实现了 resnet50 编码器,如果你附加一个分类头,那么他就能够在图像分类工作上工作。上面开始进入本文的正题实现 ConvNext。
Macro Design
1、扭转阶段计算比率
传统的 ResNet 中蕴含了 4 个阶段,而 Swin Transformer 这 4 个阶段应用的比例为 1:1:3:1(第一个阶段有一个区块,第二个阶段有一个区块,第三个阶段有三个区块……)将 ResNet50 调整为这个比率 ((3,4,6,3)->(3,3,9,3)) 能够使性能从 78.8% 进步到 79.4%。
encoder = ConvNextEncoder(in_channels=3, stem_features=64, depths=[3,3,9,3], widths=[256, 512, 1024, 2048])
2、将 stem 改为“Patchify”
ResNet stem 应用的是十分激进的 7 ×7 和 maxpool 来大量采样输出图像。然而,Transfomers 应用了 被称为“Patchify”的骨干,这意味着他们将输出图像嵌入到补丁中。Vision transforms 应用十分激进的补丁(16×16),而 ConvNext 的作者应用应用 conv 层实现的 4 ×4 补丁,这使得性能从 79.4% 晋升到 79.5%。
classConvNextStem(nn.Sequential):
def__init__(self, in_features: int, out_features: int):
super().__init__(nn.Conv2d(in_features, out_features, kernel_size=4, stride=4),
nn.BatchNorm2d(out_features)
)
3、ResNeXtify
ResNetXt 对 BottleNeck 中的 3×3 卷积层采纳分组卷积来缩小 FLOPS。在 ConvNext 中应用 depth-wise convolution(如 MobileNet 和起初的 EfficientNet)。depth-wise convolution 也是是分组卷积的一种模式,其中组数等于输出通道数。
作者留神到这与 self-attention 中的加权求和操作十分类似,后者仅在空间维度上混合信息。应用 depth-wise convs 会升高精度(因为没有像 ResNetXt 那样减少宽度),这是意料之中的毕竟晋升了速度。
所以咱们将 BottleNeck 块内的 3×3 conv 更改为上面代码
ConvNormAct(reduced_features, reduced_features, kernel_size=3, bias=False, groups=reduced_features)
4、Inverted Bottleneck(倒置瓶颈)
个别的 BottleNeck 首先通过 1×1 conv 缩小特色,而后用 3×3 conv,最初将特色扩大为原始大小,而倒置瓶颈块则相同。
所以上面咱们从宽 -> 窄 -> 宽 批改到到 窄 -> 宽 -> 窄。
这与 Transformer 相似,因为 MLP 层遵循窄 -> 宽 -> 窄设计,MLP 中的第二个浓密层将输出的特色扩大了四倍。
classBottleNeckBlock(nn.Module):
def__init__(
self,
in_features: int,
out_features: int,
expansion: int = 4,
stride: int = 1,
):
super().__init__()
expanded_features = out_features*expansion
self.block = nn.Sequential(
# narrow -> wide
ConvNormAct(in_features, expanded_features, kernel_size=1, stride=stride, bias=False),
# wide -> wide (with depth-wise)
ConvNormAct(expanded_features, expanded_features, kernel_size=3, bias=False, groups=in_features),
# wide -> narrow
ConvNormAct(expanded_features, out_features, kernel_size=1, bias=False, act=nn.Identity),
)
self.shortcut = (
nn.Sequential(
ConvNormAct(in_features, out_features, kernel_size=1, stride=stride, bias=False)
)
ifin_features!= out_features
elsenn.Identity())
self.act = nn.ReLU()
defforward(self, x: Tensor) ->Tensor:
res = x
x = self.block(x)
res = self.shortcut(res)
x += res
x = self.act(x)
returnx
5、扩充卷积核大小
像 Swin 一样,ViT 应用更大的内核尺寸(7×7)。减少内核的大小会使计算量更大,所以才应用下面提到的 depth-wise convolution,通过应用更少的通道来缩小计算量。作者指出,这相似于 Transformers 模型,其中多头自我留神 (MSA) 在 MLP 层之前实现。
classBottleNeckBlock(nn.Module):
def__init__(
self,
in_features: int,
out_features: int,
expansion: int = 4,
stride: int = 1,
):
super().__init__()
expanded_features = out_features*expansion
self.block = nn.Sequential(# narrow -> wide (with depth-wise and bigger kernel)
ConvNormAct(in_features, in_features, kernel_size=7, stride=stride, bias=False, groups=in_features),
# wide -> wide
ConvNormAct(in_features, expanded_features, kernel_size=1),
# wide -> narrow
ConvNormAct(expanded_features, out_features, kernel_size=1, bias=False, act=nn.Identity),
)
self.shortcut = (
nn.Sequential(
ConvNormAct(in_features, out_features, kernel_size=1, stride=stride, bias=False)
)
ifin_features!= out_features
elsenn.Identity())
self.act = nn.ReLU()
defforward(self, x: Tensor) ->Tensor:
res = x
x = self.block(x)
res = self.shortcut(res)
x += res
x = self.act(x)
returnx
这将准确度从 79.9% 进步到 80.6%
Micro Design
1、用 GELU 替换 ReLU
transformers 应用的是 GELU,为什么咱们不必呢?作者测试替换后准确率放弃不变。PyTorch 的 GELU 是 在 nn.GELU。
2、更少的激活函数
残差块有三个激活函数。而在 Transformer 块中,只有一个激活函数,即 MLP 块中的激活函数。作者除去了除中间层之后的所有激活。这是与 swing – t 一样的,这使得精度进步到 81.3% !
3、更少的归一化层
与激活相似,Transformers 块具备较少的归一化层。作者决定删除所有 BatchNorm,只保留两头转换之前的那个。
4、用 LN 代替 BN
作者用 LN 代替了 BN 层。他们留神到在原始 ResNet 中提到这样做会侵害性能,但通过作者以上的所有的更改后,性能进步到 81.5%
下面 4 个步骤让咱们整合起来操作:
classBottleNeckBlock(nn.Module):
def__init__(
self,
in_features: int,
out_features: int,
expansion: int = 4,
stride: int = 1,
):
super().__init__()
expanded_features = out_features*expansion
self.block = nn.Sequential(# narrow -> wide (with depth-wise and bigger kernel)
nn.Conv2d(in_features, in_features, kernel_size=7, stride=stride, bias=False, groups=in_features),
# GroupNorm with num_groups=1 is the same as LayerNorm but works for 2D data
nn.GroupNorm(num_groups=1, num_channels=in_features),
# wide -> wide
nn.Conv2d(in_features, expanded_features, kernel_size=1),
nn.GELU(),
# wide -> narrow
nn.Conv2d(expanded_features, out_features, kernel_size=1),
)
self.shortcut = (
nn.Sequential(
ConvNormAct(in_features, out_features, kernel_size=1, stride=stride, bias=False)
)
ifin_features!= out_features
elsenn.Identity())
defforward(self, x: Tensor) ->Tensor:
res = x
x = self.block(x)
res = self.shortcut(res)
x += res
returnx
拆散下采样层
在 ResNet 中,下采样是通过 stride=2 conv 实现的。Transformers(以及其余卷积网络)也有一个独自的下采样模块。作者删除了 stride=2 并在三个 conv 之前增加了一个下采样块,为了放弃训练期间的稳定性在,在下采样操作之前须要进行归一化。将此模块增加到 ConvNexStage。达到了超过 Swin 的 82.0%!
classConvNexStage(nn.Sequential):
def__init__(self, in_features: int, out_features: int, depth: int, **kwargs):
super().__init__(
# add the downsampler
nn.Sequential(nn.GroupNorm(num_groups=1, num_channels=in_features),
nn.Conv2d(in_features, out_features, kernel_size=2, stride=2)
),
*[BottleNeckBlock(out_features, out_features, **kwargs)
for_inrange(depth)
],
)
当初咱们失去了最终的 BottleNeckBlock 层代码:
classBottleNeckBlock(nn.Module):
def__init__(
self,
in_features: int,
out_features: int,
expansion: int = 4,
):
super().__init__()
expanded_features = out_features*expansion
self.block = nn.Sequential(# narrow -> wide (with depth-wise and bigger kernel)
nn.Conv2d(in_features, in_features, kernel_size=7, padding=3, bias=False, groups=in_features),
# GroupNorm with num_groups=1 is the same as LayerNorm but works for 2D data
nn.GroupNorm(num_groups=1, num_channels=in_features),
# wide -> wide
nn.Conv2d(in_features, expanded_features, kernel_size=1),
nn.GELU(),
# wide -> narrow
nn.Conv2d(expanded_features, out_features, kernel_size=1),
)
defforward(self, x: Tensor) ->Tensor:
res = x
x = self.block(x)
x += res
returnx
让咱们测试一下最终的 stage 代码
stage = ConvNexStage(32, 62, depth=1)
stage(torch.randn(1, 32, 14, 14)).shape
#torch.Size([1, 62, 7, 7])
最初的一些丢该
论文中还增加了 Stochastic Depth,也称为 Drop Path 还有 Layer Scale。
fromtorchvision.opsimportStochasticDepth
classLayerScaler(nn.Module):
def__init__(self, init_value: float, dimensions: int):
super().__init__()
self.gamma = nn.Parameter(init_value*torch.ones((dimensions)),
requires_grad=True)
defforward(self, x):
returnself.gamma[None,...,None,None] *x
classBottleNeckBlock(nn.Module):
def__init__(
self,
in_features: int,
out_features: int,
expansion: int = 4,
drop_p: float = .0,
layer_scaler_init_value: float = 1e-6,
):
super().__init__()
expanded_features = out_features*expansion
self.block = nn.Sequential(# narrow -> wide (with depth-wise and bigger kernel)
nn.Conv2d(in_features, in_features, kernel_size=7, padding=3, bias=False, groups=in_features),
# GroupNorm with num_groups=1 is the same as LayerNorm but works for 2D data
nn.GroupNorm(num_groups=1, num_channels=in_features),
# wide -> wide
nn.Conv2d(in_features, expanded_features, kernel_size=1),
nn.GELU(),
# wide -> narrow
nn.Conv2d(expanded_features, out_features, kernel_size=1),
)
self.layer_scaler = LayerScaler(layer_scaler_init_value, out_features)
self.drop_path = StochasticDepth(drop_p, mode="batch")
defforward(self, x: Tensor) ->Tensor:
res = x
x = self.block(x)
x = self.layer_scaler(x)
x = self.drop_path(x)
x += res
returnx
好了,当初咱们看看最终后果
stage = ConvNexStage(32, 62, depth=1)
stage(torch.randn(1, 32, 14, 14)).shape
#torch.Size([1, 62, 7, 7])
最初咱们批改一下 Drop Path 的概率
classConvNextEncoder(nn.Module):
def__init__(
self,
in_channels: int,
stem_features: int,
depths: List[int],
widths: List[int],
drop_p: float = .0,
):
super().__init__()
self.stem = ConvNextStem(in_channels, stem_features)
in_out_widths = list(zip(widths, widths[1:]))
# create drop paths probabilities (one for each stage)
drop_probs = [x.item() forxintorch.linspace(0, drop_p, sum(depths))]
self.stages = nn.ModuleList(
[ConvNexStage(stem_features, widths[0], depths[0], drop_p=drop_probs[0]),
*[ConvNexStage(in_features, out_features, depth, drop_p=drop_p)
for (in_features, out_features), depth, drop_pinzip(in_out_widths, depths[1:], drop_probs[1:]
)
],
]
)
defforward(self, x):
x = self.stem(x)
forstageinself.stages:
x = stage(x)
returnx
测试:
image = torch.rand(1, 3, 224, 224)
encoder = ConvNextEncoder(in_channels=3, stem_features=64, depths=[3,4,6,4], widths=[256, 512, 1024, 2048])
encoder(image).shape
#torch.Size([1, 2048, 3, 3])
ConvNext 的特色,咱们须要在编码器顶部利用分类头。咱们还在最初一个线性层之前增加了一个 LayerNorm。
classClassificationHead(nn.Sequential):
def__init__(self, num_channels: int, num_classes: int = 1000):
super().__init__(nn.AdaptiveAvgPool2d((1, 1)),
nn.Flatten(1),
nn.LayerNorm(num_channels),
nn.Linear(num_channels, num_classes)
)
classConvNextForImageClassification(nn.Sequential):
def__init__(self,
in_channels: int,
stem_features: int,
depths: List[int],
widths: List[int],
drop_p: float = .0,
num_classes: int = 1000):
super().__init__()
self.encoder = ConvNextEncoder(in_channels, stem_features, depths, widths, drop_p)
self.head = ClassificationHead(widths[-1], num_classes)
最终模型测试:
image = torch.rand(1, 3, 224, 224)
classifier = ConvNextForImageClassification(in_channels=3, stem_features=64, depths=[3,4,6,4], widths=[256, 512, 1024, 2048])
classifier(image).shape
#torch.Size([1, 1000])
最初总结
在本文中复现了作者应用 ResNet 创立 ConvNext 的所有过程。如果你想须要残缺代码,能够查看这个地址:
https://avoid.overfit.cn/post/1fd17e7520134996b532ecd50de9672f
作者:Francesco Zuppichini