开源库 timm | 基于 Pytorch 的28个视觉 Transformer 实现方法-技术圈

作者丨科技猛兽

转自丨极市平台

导读

本文将介绍一个优秀的PyTorch开源库——timm库，并对其中的vision transformer.py代码进行了详细解读。

Transformer 架构早已在自然语言处理任务中得到广泛应用，但在计算机视觉领域中仍然受到限制。在计算机视觉领域，目前已有大量工作表明模型对 CNN 的依赖不是必需的，当直接应用于图像块序列时，Transformer 也能很好地执行图像分类任务。

本文将简要介绍了优秀的 PyTorch Image Model 库：timm库。与此同时，将会为大家详细介绍其中的视觉Transformer代码以及一个优秀的视觉Transformer 的PyTorch实现，以帮助大家更快地开展相关实验。

什么是timm库？

PyTorchImageModels，简称timm，是一个巨大的PyTorch代码集合，包括了一系列：

image models
layers
utilities
optimizers
schedulers
data-loaders / augmentations
training / validation scripts

旨在将各种SOTA模型整合在一起，并具有复现ImageNet训练结果的能力。

timm库作者是来自加拿大温哥华的Ross Wightman。

作者github链接：
https://github.com/rwightman

timm库链接：
https://github.com/rwightman/pytorch-image-models

所有的PyTorch模型及其对应arxiv链接如下：

Big Transfer ResNetV2 (BiT) - https://arxiv.org/abs/1912.11370
CspNet (Cross-Stage Partial Networks) - https://arxiv.org/abs/1911.11929
DeiT (Vision Transformer) - https://arxiv.org/abs/2012.12877
DenseNet - https://arxiv.org/abs/1608.06993
DLA - https://arxiv.org/abs/1707.06484
DPN (Dual-Path Network) - https://arxiv.org/abs/1707.01629
EfficientNet (MBConvNet Family)
EfficientNet NoisyStudent (B0-B7, L2) - https://arxiv.org/abs/1911.04252
EfficientNet AdvProp (B0-B8) - https://arxiv.org/abs/1911.09665
EfficientNet (B0-B7) - https://arxiv.org/abs/1905.11946
EfficientNet-EdgeTPU (S, M, L) - https://ai.googleblog.com/2019/08/efficientnet-edgetpu-creating.html
FBNet-C - https://arxiv.org/abs/1812.03443
MixNet - https://arxiv.org/abs/1907.09595
MNASNet B1, A1 (Squeeze-Excite), and Small - https://arxiv.org/abs/1807.11626
MobileNet-V2 - https://arxiv.org/abs/1801.04381
Single-Path NAS - https://arxiv.org/abs/1904.02877
GPU-Efficient Networks - https://arxiv.org/abs/2006.14090
HRNet - https://arxiv.org/abs/1908.07919
Inception-V3 - https://arxiv.org/abs/1512.00567
Inception-ResNet-V2 and Inception-V4 - https://arxiv.org/abs/1602.07261
MobileNet-V3 (MBConvNet w/ Efficient Head) - https://arxiv.org/abs/1905.02244
NASNet-A - https://arxiv.org/abs/1707.07012
NFNet-F - https://arxiv.org/abs/2102.06171
NF-RegNet / NF-ResNet - https://arxiv.org/abs/2101.08692
PNasNet - https://arxiv.org/abs/1712.00559
RegNet - https://arxiv.org/abs/2003.13678
RepVGG - https://arxiv.org/abs/2101.03697
ResNet/ResNeXt
ResNet (v1b/v1.5) - https://arxiv.org/abs/1512.03385
ResNeXt - https://arxiv.org/abs/1611.05431
'Bag of Tricks' / Gluon C, D, E, S variations - https://arxiv.org/abs/1812.01187
Weakly-supervised (WSL) Instagram pretrained / ImageNet tuned ResNeXt101 - https://arxiv.org/abs/1805.00932
Semi-supervised (SSL) / Semi-weakly Supervised (SWSL) ResNet/ResNeXts - https://arxiv.org/abs/1905.00546
ECA-Net (ECAResNet) - https://arxiv.org/abs/1910.03151v4
Squeeze-and-Excitation Networks (SEResNet) - https://arxiv.org/abs/1709.01507
Res2Net - https://arxiv.org/abs/1904.01169
ResNeSt - https://arxiv.org/abs/2004.08955
ReXNet - https://arxiv.org/abs/2007.00992
SelecSLS - https://arxiv.org/abs/1907.00837
Selective Kernel Networks - https://arxiv.org/abs/1903.06586
TResNet - https://arxiv.org/abs/2003.13630
Vision Transformer - https://arxiv.org/abs/2010.11929
VovNet V2 and V1 - https://arxiv.org/abs/1911.06667
Xception - https://arxiv.org/abs/1610.02357
Xception (Modified Aligned, Gluon) - https://arxiv.org/abs/1802.02611
Xception (Modified Aligned, TF) - https://arxiv.org/abs/1802.02611

timm库特点

所有的模型都有默认的API：

accessing/changing the classifier - get_classifier and reset_classifier
只对features做前向传播 - forward_features

所有模型都支持多尺度特征提取 (feature pyramids) (通过create_model函数)：

create_model(name, features_only=True, out_indices=..., output_stride=...)

out_indices 指定返回哪个feature maps to return, 从0开始，out_indices[i]对应着 C(i + 1) feature level。

output_stride 通过dilated convolutions控制网络的output stride。大多数网络默认 stride 32 。

所有的模型都有一致的pretrained weight loader，adapts last linear if necessary。

训练方式支持：

NVIDIA DDP w/ a single GPU per process, multiple processes with APEX present (AMP mixed-precision optional)
PyTorch DistributedDataParallel w/ multi-gpu, single process (AMP disabled as it crashes when enabled)
PyTorch w/ single GPU single process (AMP optional)

动态的全局池化方式可以选择： average pooling, max pooling, average + max, or concat([average, max])，默认是adaptive average。

Schedulers：

Schedulers 包括step,cosinew/ restarts,tanhw/ restarts,plateau 。

Optimizer：

rmsprop_tf adapted from PyTorch RMSProp by myself. Reproduces much improved Tensorflow RMSProp behaviour.
radam by Liyuan Liu (https://arxiv.org/abs/1908.03265)
novograd by Masashi Kimura (https://arxiv.org/abs/1905.11286)
lookahead adapted from impl by Liam (https://arxiv.org/abs/1907.08610)
fused optimizers by name with NVIDIA Apex installed
adamp and sgdp by Naver ClovAI (https://arxiv.org/abs/2006.08217)
adafactor adapted from FAIRSeq impl (https://arxiv.org/abs/1804.04235)
adahessian by David Samuel (https://arxiv.org/abs/2006.00719)

timm库 vision_transformer.py代码解读

代码来自：

https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py

对应的论文是ViT，是除了官方开源的代码之外的又一个优秀的PyTorch implement。

An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

https://arxiv.org/abs/2010.11929

另一篇工作DeiT也大量借鉴了timm库这份代码的实现：

Training data-efficient image transformers & distillation through attention

Training data-efficient image transformers & distillation through attention

https://arxiv.org/abs/2012.12877

vision_transformer.py：

代码中定义的变量的含义如下：

img_size：tuple 类型，里面是int类型，代表输入的图片大小，默认是 224。
patch_size：tuple 类型，里面是int类型，代表Patch的大小，默认是 16。
in_chans：int 类型，代表输入图片的channel数，默认是3。
num_classes：int 类型classification head的分类数，比如CIFAR100就是100，默认是 1000。
embed_dim：int 类型Transformer的embedding dimension，默认是 768。
depth：int 类型，Transformer的Block的数量，默认是 12。
num_heads：int 类型，attention heads的数量，默认是12。
mlp_ratio：int 类型，mlp hidden dim/embedding dim的值，默认是 4。
qkv_bias：bool 类型，attention模块计算qkv时需要bias吗，默认是 True。
qk_scale： 一般设置成 None 就行。
drop_rate：float 类型，dropout rate，默认是 0。
attn_drop_rate：float 类型，attention模块的dropout rate，默认是 0。
drop_path_rate：float 类型，默认是 0。
hybrid_backbone：nn.Module 类型，在把图片转换成Patch之前，需要先通过一个Backbone吗？默认是 None。
如果是None，就直接把图片转化成Patch。
如果不是None，就先通过这个Backbone，再转化成Patch。
norm_layer：nn.Module 类型，归一化层类型，默认是 None。

1. 导入必要的库和模型：

import mathimport loggingfrom functools import partialfrom collections import OrderedDict
import torchimport torch.nn as nnimport torch.nn.functional as F
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STDfrom .helpers import load_pretrainedfrom .layers import StdConv2dSame, DropPath, to_2tuple, trunc_normal_from .resnet import resnet26d, resnet50dfrom .resnetv2 import ResNetV2from .registry import register_model

2. 定义一个字典，代表标准的模型，如果需要更改模型超参数只需要改变_cfg

的传入的参数即可。

def _cfg(url='', **kwargs):    return {        'url': url,        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,        'crop_pct': .9, 'interpolation': 'bicubic',        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,        'first_conv': 'patch_embed.proj', 'classifier': 'head',        **kwargs    }

3. default_cfgs代表支持的所有模型，也定义成字典的形式：

vit_small_patch16_224里面的small代表小模型。
ViT的第一步要把图片分成一个个patch，然后把这些patch组合在一起作为对图像的序列化操作，比如一张224 × 224的图片分成大小为16 × 16的patch，那一共可以分成196个。所以这个图片就序列化成了(196, 256)的tensor。所以这里的：
16： 就代表patch的大小。
224： 就代表输入图片的大小。
按照这个命名方式，支持的模型有：vit_base_patch16_224，vit_base_patch16_384等等。
后面的vit_deit_base_patch16_224等等模型代表DeiT这篇论文的模型。


default_cfgs = {    # patch models (my experiments)    'vit_small_patch16_224': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_small_p16_224-15ec54c9.pth',    ),
    # patch models (weights ported from official Google JAX impl)    'vit_base_patch16_224': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth',        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5),    ),    'vit_base_patch32_224': _cfg(        url='',  # no official model weights for this combo, only for in21k        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_base_patch16_384': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_384-83fb41ba.pth',        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0),    'vit_base_patch32_384': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p32_384-830016f5.pth',        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0),    'vit_large_patch16_224': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_224-4ee7a4dc.pth',        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_large_patch32_224': _cfg(        url='',  # no official model weights for this combo, only for in21k        mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_large_patch16_384': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p16_384-b3be5167.pth',        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0),    'vit_large_patch32_384': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p32_384-9b920ba8.pth',        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0),
    # patch models, imagenet21k (weights ported from official Google JAX impl)    'vit_base_patch16_224_in21k': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_patch16_224_in21k-e5005f0a.pth',        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_base_patch32_224_in21k': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_patch32_224_in21k-8db57226.pth',        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_large_patch16_224_in21k': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch16_224_in21k-606da67d.pth',        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_large_patch32_224_in21k': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch32_224_in21k-9046d2e7.pth',        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),    'vit_huge_patch14_224_in21k': _cfg(        url='',  # FIXME I have weights for this but > 2GB limit for github release binaries        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
    # hybrid models (weights ported from official Google JAX impl)    'vit_base_resnet50_224_in21k': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_224_in21k-6f7c7740.pth',        num_classes=21843, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=0.9, first_conv='patch_embed.backbone.stem.conv'),    'vit_base_resnet50_384': _cfg(        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_resnet50_384-9fd3c705.pth',        input_size=(3, 384, 384), mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5), crop_pct=1.0, first_conv='patch_embed.backbone.stem.conv'),
    # hybrid models (my experiments)    'vit_small_resnet26d_224': _cfg(),    'vit_small_resnet50d_s3_224': _cfg(),    'vit_base_resnet26d_224': _cfg(),    'vit_base_resnet50d_224': _cfg(),
    # deit models (FB weights)    'vit_deit_tiny_patch16_224': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth'),    'vit_deit_small_patch16_224': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_small_patch16_224-cd65a155.pth'),    'vit_deit_base_patch16_224': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth',),    'vit_deit_base_patch16_384': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_384-8de9b5d1.pth',        input_size=(3, 384, 384), crop_pct=1.0),    'vit_deit_tiny_distilled_patch16_224': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_tiny_distilled_patch16_224-b40b3cf7.pth'),    'vit_deit_small_distilled_patch16_224': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_small_distilled_patch16_224-649709d9.pth'),    'vit_deit_base_distilled_patch16_224': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_224-df68dfff.pth', ),    'vit_deit_base_distilled_patch16_384': _cfg(        url='https://dl.fbaipublicfiles.com/deit/deit_base_distilled_patch16_384-d0272ac0.pth',        input_size=(3, 384, 384), crop_pct=1.0),}

4. FFN实现：


class Mlp(nn.Module):    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):        super().__init__()        out_features = out_features or in_features        hidden_features = hidden_features or in_features        self.fc1 = nn.Linear(in_features, hidden_features)        self.act = act_layer()        self.fc2 = nn.Linear(hidden_features, out_features)        self.drop = nn.Dropout(drop)
    def forward(self, x):        x = self.fc1(x)        x = self.act(x)        x = self.drop(x)        x = self.fc2(x)        x = self.drop(x)        return x

5. Attention实现：

在python 3.5以后，@是一个操作符，表示矩阵-向量乘法
A@x 就是矩阵-向量乘法A*x: np.dot(A, x)。


class Attention(nn.Module):    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):        super().__init__()        self.num_heads = num_heads        head_dim = dim // num_heads        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights        self.scale = qk_scale or head_dim ** -0.5
        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)        self.attn_drop = nn.Dropout(attn_drop)        self.proj = nn.Linear(dim, dim)        self.proj_drop = nn.Dropout(proj_drop)
    def forward(self, x):        B, N, C = x.shape        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
        attn = (q @ k.transpose(-2, -1)) * self.scale        attn = attn.softmax(dim=-1)        attn = self.attn_drop(attn)
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)        x = self.proj(x)        x = self.proj_drop(x)
        # x: (B, N, C)        return x

6. 包含Attention和Add & Norm的Block实现：

图1：Block类对应结构

不同之处是：
先进行Norm，再Attention；先进行Norm，再通过FFN (MLP)。


class Block(nn.Module):    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):        super().__init__()        self.norm1 = norm_layer(dim)        self.attn = Attention(            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()        self.norm2 = norm_layer(dim)        mlp_hidden_dim = int(dim * mlp_ratio)        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
    def forward(self, x):        x = x + self.drop_path(self.attn(self.norm1(x)))        x = x + self.drop_path(self.mlp(self.norm2(x)))        return x

7. 接下来要把图片转换成Patch，一种做法是直接把Image转化成Patch，另一种做法是把Backbone输出的特征转化成Patch。

1) 直接把Image转化成Patch：

输入的x的维度是：(B, C, H, W)
输出的PatchEmbedding的维度是：(B, 14*14, 768)，768表示embed_dim，14*14表示一共有196个Patches。


class PatchEmbed(nn.Module):    """ Image to Patch Embedding    """    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):        super().__init__()        img_size = to_2tuple(img_size)        patch_size = to_2tuple(patch_size)        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])        self.img_size = img_size        self.patch_size = patch_size        self.num_patches = num_patches
        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
    def forward(self, x):        B, C, H, W = x.shape        # FIXME look at relaxing size constraints        assert H == self.img_size[0] and W == self.img_size[1], \            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."        x = self.proj(x).flatten(2).transpose(1, 2)
        # x: (B, 14*14, 768)        return x

2) 把Backbone输出的特征转化成Patch：

输入的x的维度是：(B, C, H, W)
得到Backbone输出的维度是：(B, feature_size, feature_size, feature_dim)
输出的PatchEmbedding的维度是：(B, feature_size, feature_size, embed_dim)，一共有feature_size * feature_size个Patches。


class HybridEmbed(nn.Module):    """ CNN Feature Map Embedding    Extract feature map from CNN, flatten, project to embedding dim.    """    def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768):        super().__init__()        assert isinstance(backbone, nn.Module)        img_size = to_2tuple(img_size)        self.img_size = img_size        self.backbone = backbone        if feature_size is None:            with torch.no_grad():                # FIXME this is hacky, but most reliable way of determining the exact dim of the output feature                # map for all networks, the feature metadata has reliable channel and stride info, but using                # stride to calc feature dim requires info about padding of each stage that isn't captured.                training = backbone.training                if training:                    backbone.eval()                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))                if isinstance(o, (list, tuple)):                    o = o[-1]  # last feature if backbone outputs list/tuple of features                feature_size = o.shape[-2:]                feature_dim = o.shape[1]                backbone.train(training)        else:            feature_size = to_2tuple(feature_size)            if hasattr(self.backbone, 'feature_info'):                feature_dim = self.backbone.feature_info.channels()[-1]            else:                feature_dim = self.backbone.num_features        self.num_patches = feature_size[0] * feature_size[1]        self.proj = nn.Conv2d(feature_dim, embed_dim, 1)
    def forward(self, x):        x = self.backbone(x)        if isinstance(x, (list, tuple)):            x = x[-1]  # last feature if backbone outputs list/tuple of features        x = self.proj(x).flatten(2).transpose(1, 2)        return x

8. 以上是ViT所需的所有模块的定义，下面是VisionTransformer 这个类的实现：

8.1 使用这个类时需要传入的变量，其含义已经在本小节一开始介绍。

class VisionTransformer(nn.Module):    """ Vision Transformer
    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`  -        https://arxiv.org/abs/2010.11929    """    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,                 num_heads=12, mlp_ratio=4., qkv_bias=True, qk_scale=None, representation_size=None,                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None):

8.2 得到分块后的Patch的数量：

super().__init__()self.num_classes = num_classesself.num_features = self.embed_dim = embed_dim  # num_features for consistency with other modelsnorm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
if hybrid_backbone is not None:    self.patch_embed = HybridEmbed(        hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)else:    self.patch_embed = PatchEmbed(        img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)num_patches = self.patch_embed.num_patches

8.3 class token：

一开始定义成(1, 1, 768)，之后再变成(B, 1, 768)。


self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))

8.4 定义位置编码：

self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))

8.5 把12个Block连接起来：


self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay ruleself.blocks = nn.ModuleList([    Block(        dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,        drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)    for i in range(depth)])self.norm = norm_layer(embed_dim)

8.6 表示层和分类头：

表示层输出维度是representation_size，分类头输出维度是num_classes。

# Representation layerif representation_size:    self.num_features = representation_size    self.pre_logits = nn.Sequential(OrderedDict([        ('fc', nn.Linear(embed_dim, representation_size)),        ('act', nn.Tanh())    ]))else:    self.pre_logits = nn.Identity()
# Classifier headself.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

8.7 初始化各个模块：

函数trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.)的目的是用截断的正态分布绘制的值填充输入张量，我们只需要输入均值mean，标准差std，下界a，上界b即可。
self.apply(self._init_weights)表示对各个模块的权重进行初始化。apply函数的代码是：


        for module in self.children():            module.apply(fn)        fn(self)        return self

递归地将fn应用于每个子模块，相当于在递归调用fn，即_init_weights这个函数。
也就是把模型的所有子模块的nn.Linear和nn.LayerNorm层都初始化掉。


trunc_normal_(self.pos_embed, std=.02)trunc_normal_(self.cls_token, std=.02)self.apply(self._init_weights)
def _init_weights(self, m):if isinstance(m, nn.Linear):    trunc_normal_(m.weight, std=.02)    if isinstance(m, nn.Linear) and m.bias is not None:        nn.init.constant_(m.bias, 0)elif isinstance(m, nn.LayerNorm):    nn.init.constant_(m.bias, 0)    nn.init.constant_(m.weight, 1.0)

8.8 最后就是整个ViT模型的forward实现：


def forward_features(self, x):    B = x.shape[0]    x = self.patch_embed(x)
    cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks    x = torch.cat((cls_tokens, x), dim=1)    x = x + self.pos_embed    x = self.pos_drop(x)
    for blk in self.blocks:        x = blk(x)
    x = self.norm(x)[:, 0]    x = self.pre_logits(x)    return x
def forward(self, x):    x = self.forward_features(x)    x = self.head(x)    return x

9. 下面是Training data-efficient image transformers & distillation through attention这篇论文的DeiT这个类的实现：

整体结构与ViT相似，继承了上面的VisionTransformer类。


class DistilledVisionTransformer(VisionTransformer):

再额外定义以下3个变量：

distillation token：dist_token
新的位置编码：pos_embed
蒸馏分类头：head_dist

DeiT相关介绍可以参考：Vision Transformer 超详细解读 (原理分析+代码解读) (三)。

self.dist_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))num_patches = self.patch_embed.num_patchesself.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 2, self.embed_dim))self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if self.num_classes > 0 else nn.Identity()

初始化新定义的变量：


trunc_normal_(self.dist_token, std=.02)trunc_normal_(self.pos_embed, std=.02)self.head_dist.apply(self._init_weights)

前向函数：


def forward_features(self, x):    B = x.shape[0]    x = self.patch_embed(x)
    cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks    dist_token = self.dist_token.expand(B, -1, -1)    x = torch.cat((cls_tokens, dist_token, x), dim=1)
    x = x + self.pos_embed    x = self.pos_drop(x)
    for blk in self.blocks:        x = blk(x)
    x = self.norm(x)    return x[:, 0], x[:, 1]
def forward(self, x):    x, x_dist = self.forward_features(x)    x = self.head(x)    x_dist = self.head_dist(x_dist)    if self.training:        return x, x_dist    else:        # during inference, return the average of both classifier predictions        return (x + x_dist) / 2

10. 对位置编码进行插值：

posemb代表未插值的位置编码权值，posemb_tok为位置编码的token部分，posemb_grid为位置编码的插值部分。
首先把要插值部分posemb_grid给reshape成(1, gs_old, gs_old, -1)的形式，再插值成(1, gs_new, gs_new, -1)的形式，最后与token部分在第1维度拼接在一起，得到插值后的位置编码posemb。

def resize_pos_embed(posemb, posemb_new):    # Rescale the grid of position embeddings when loading from state_dict. Adapted from    # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224    _logger.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape)    ntok_new = posemb_new.shape[1]    if True:        posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:]        ntok_new -= 1    else:        posemb_tok, posemb_grid = posemb[:, :0], posemb[0]    gs_old = int(math.sqrt(len(posemb_grid)))    gs_new = int(math.sqrt(ntok_new))    _logger.info('Position embedding grid-size from %s to %s', gs_old, gs_new)    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)    posemb_grid = F.interpolate(posemb_grid, size=(gs_new, gs_new), mode='bilinear')    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new * gs_new, -1)    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)    return posemb

11. _create_vision_transformer函数用于创建vision transformer：

checkpoint_filter_fn的作用是加载预训练权重。


def checkpoint_filter_fn(state_dict, model):    """ convert patch embedding weight from manual patchify + linear proj to conv"""    out_dict = {}    if 'model' in state_dict:        # For deit models        state_dict = state_dict['model']    for k, v in state_dict.items():        if 'patch_embed.proj.weight' in k and len(v.shape) < 4:            # For old models that I trained prior to conv based patchification            O, I, H, W = model.patch_embed.proj.weight.shape            v = v.reshape(O, -1, H, W)        elif k == 'pos_embed' and v.shape != model.pos_embed.shape:            # To resize pos embedding when using model at different size from pretrained weights            v = resize_pos_embed(v, model.pos_embed)        out_dict[k] = v    return out_dict

def _create_vision_transformer(variant, pretrained=False, distilled=False, **kwargs):    default_cfg = default_cfgs[variant]    default_num_classes = default_cfg['num_classes']    default_img_size = default_cfg['input_size'][-1]
    num_classes = kwargs.pop('num_classes', default_num_classes)    img_size = kwargs.pop('img_size', default_img_size)    repr_size = kwargs.pop('representation_size', None)    if repr_size is not None and num_classes != default_num_classes:        # Remove representation layer if fine-tuning. This may not always be the desired action,        # but I feel better than doing nothing by default for fine-tuning. Perhaps a better interface?        _logger.warning("Removing representation layer for fine-tuning.")        repr_size = None
    model_cls = DistilledVisionTransformer if distilled else VisionTransformer    model = model_cls(img_size=img_size, num_classes=num_classes, representation_size=repr_size, **kwargs)    model.default_cfg = default_cfg
    if pretrained:        load_pretrained(            model, num_classes=num_classes, in_chans=kwargs.get('in_chans', 3),            filter_fn=partial(checkpoint_filter_fn, model=model))    return model

12. 定义和注册vision transformer模型：

@ 指装饰器。
@register_model代表注册器，注册这个新定义的模型。
model_kwargs是一个存有模型所有超参数的字典。
最后使用上面定义的_create_vision_transformer函数创建模型。


@register_modeldef vit_base_patch16_224(pretrained=False, **kwargs):    """ ViT-Base (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929).    ImageNet-1k weights fine-tuned from in21k @ 224x224, source https://github.com/google-research/vision_transformer.    """    model_kwargs = dict(patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs)    model = _create_vision_transformer('vit_base_patch16_224', pretrained=pretrained, **model_kwargs)    return model

一共可以选择的模型包括：

ViT系列：
vit_small_patch16_224
vit_base_patch16_224
vit_base_patch32_224
vit_base_patch16_384
vit_base_patch32_384
vit_large_patch16_224
vit_large_patch32_224
vit_large_patch16_384
vit_large_patch32_384
vit_base_patch16_224_in21k
vit_base_patch32_224_in21k
vit_large_patch16_224_in21k
vit_large_patch32_224_in21k
vit_huge_patch14_224_in21k
vit_base_resnet50_224_in21k
vit_base_resnet50_384
vit_small_resnet26d_224
vit_small_resnet50d_s3_224
vit_base_resnet26d_224
vit_base_resnet50d_224
DeiT系列：
vit_deit_tiny_patch16_224
vit_deit_small_patch16_224
vit_deit_base_patch16_224
vit_deit_base_patch16_384
vit_deit_tiny_distilled_patch16_224
vit_deit_small_distilled_patch16_224
vit_deit_base_distilled_patch16_224
vit_deit_base_distilled_patch16_384

以上就是对timm库 vision_transformer.py代码的分析。

如何使用timm库以及 vision_transformer.py代码搭建自己的模型？

在搭建我们自己的视觉Transformer模型时，我们可以按照下面的步骤操作：首先

继承timm库的VisionTransformer这个类。
添加上自己模型独有的一些变量。
重写forward函数。
通过timm库的注册器注册新模型。

我们以ViT模型的改进版DeiT为例：

首先，DeiT的所有模型列表如下：

__all__ = [    'deit_tiny_patch16_224', 'deit_small_patch16_224', 'deit_base_patch16_224',    'deit_tiny_distilled_patch16_224', 'deit_small_distilled_patch16_224',    'deit_base_distilled_patch16_224', 'deit_base_patch16_384',    'deit_base_distilled_patch16_384',]

导入VisionTransformer这个类，注册器register_model，以及初始化函数trunc_normal_：

from timm.models.vision_transformer import VisionTransformer, _cfgfrom timm.models.registry import register_modelfrom timm.models.layers import trunc_normal_

DeiT的class名称是DistilledVisionTransformer，它直接继承了VisionTransformer这个类：


class DistilledVisionTransformer(VisionTransformer):

添加上自己模型独有的一些变量：


def __init__(self, *args, **kwargs):    super().__init__(*args, **kwargs)    self.dist_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))    num_patches = self.patch_embed.num_patches    # 位置编码不是ViT中的(b, N, 256), 而变成了(b, N+2, 256), 原因是还有class token和distillation token.    self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 2, self.embed_dim))    self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if self.num_classes > 0 else nn.Identity()
    trunc_normal_(self.dist_token, std=.02)    trunc_normal_(self.pos_embed, std=.02)    self.head_dist.apply(self._init_weights)

重写forward函数：


def forward_features(self, x):    # taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py    # with slight modifications to add the dist_token    B = x.shape[0]
    x = self.patch_embed(x)
    cls_tokens = self.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks    dist_token = self.dist_token.expand(B, -1, -1)
    x = torch.cat((cls_tokens, dist_token, x), dim=1)
    x = x + self.pos_embed    x = self.pos_drop(x)
    for blk in self.blocks:        x = blk(x)
    x = self.norm(x)
    return x[:, 0], x[:, 1]
def forward(self, x):    x, x_dist = self.forward_features(x)    x = self.head(x)    x_dist = self.head_dist(x_dist)    if self.training:        return x, x_dist    else:        # during inference, return the average of both classifier predictions        return (x + x_dist) / 2

通过timm库的注册器注册新模型：


@register_modeldef deit_base_patch16_224(pretrained=False, **kwargs):    model = VisionTransformer(        patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,        norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)    model.default_cfg = _cfg()    if pretrained:        checkpoint = torch.hub.load_state_dict_from_url(            url="https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth",            map_location="cpu", check_hash=True        )        model.load_state_dict(checkpoint["model"])    return model