视觉 Transformer 优秀开源工作：timm 库 vision transformer 代码解读

timm库 vision_transformer.py代码解读

https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py

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：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
import math
import logging
from functools import partial
from collections import OrderedDict

import torch
import torch.nn as nn
import torch.nn.functional as F

from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from .helpers import load_pretrained
from .layers import StdConv2dSame, DropPath, to_2tuple, trunc_normal_
from .resnet import resnet26d, resnet50d
from .resnetv2 import ResNetV2
from .registry import register_model

2 定义字典，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。

1
2
3
4
5
6
7
8
9
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

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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
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: http://np.dot/(A, x)。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
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

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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
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 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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
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 VisionTransformer 类的实现

8.1 传入的变量

1
2
3
4
5
6
7
8
9
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 获取Patch的数量

1
2
3
4
5
6
7
8
9
10
11
12
super().__init__()
self.num_classes = num_classes
self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
norm_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)。

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

8.4 定义位置编码

1
2
3
4
5
6
7
8
9
10
11
12
13
14
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))

8.5 把12个Block连接起来：

​```python
self.pos_drop = nn.Dropout(p=drop_rate)

dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
self.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。

1
2
3
4
5
6
7
8
9
10
11
12
# Representation layer
if 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 head
self.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：

1
2
3
4
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。

1
2
3
4
5
6
7
8
9
10
11
12
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 forward实现

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
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。

1
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：https://zhuanlan.zhihu.com/p/348593638。

1
2
3
4
self.dist_token = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
num_patches = self.patch_embed.num_patches
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()

初始化新定义的变量：

1
2
3
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：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
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。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
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：

@ register_model这个函数来自timm库model文件夹下的registry.py文件，它的作用是：
@ 指装饰器
@register_model代表注册器，注册这个新定义的模型。
存储到**_model_entrypoints**这个字典中，比如：

1
_model_entrypoints[vit_base_patch16_224] = _create_vision_transformer('vit_base_patch16_224', pretrained=pretrained, **model_kwargs)

然后在factory.py的create_model函数中的下面这几行真正创建模型，你以后想创建的任何模型都会使用create_model这个函数，这里说清楚了为什么要用它：

1
2
3
4
5
6
7
if is_model(model_name):
    create_fn = model_entrypoint(model_name)
else:
    raise RuntimeError('Unknown model (%s)' % model_name)

with set_layer_config(scriptable=scriptable, exportable=exportable, no_jit=no_jit):
    model = create_fn(pretrained=pretrained, **kwargs)

比如刚才在main.py里面用了create_model创建模型，如下面代码所示。而create_model就来自factory.py：

1
2
3
4
5
6
7
8
model = create_model(
    args.model,
    pretrained=False,
    num_classes=args.nb_classes,
    drop_rate=args.drop,
    drop_path_rate=args.drop_path,
    drop_block_rate=None,
)

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。

4 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：

1
2
3
4
5
6
__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_：

1
2
3
from timm.models.vision_transformer import VisionTransformer, _cfg
from timm.models.registry import register_model
from timm.models.layers import trunc_normal_

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

1
class DistilledVisionTransformer(VisionTransformer):

添加上自己模型独有的一些变量：

1
2
3
4
5
6
7
8
9
10
11
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：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
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：

1
2
3
4
5
6
7
8
9
10
11
12
13
@register_model
def 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

———————————————— 更新 2021.03.01———————————————

5 timm库 train.py代码解读：

timm库的训练使用 结合apex支持的分布式训练，同步bn，以及混合精度的训练方式，其train.py的写法很具有代表性，值得拿出来讨论。因此这篇文章再多加一段，来专门讨论这个train.py。

结合apex支持的分布式训练，同步bn，以及混合精度的训练方式的详细讲解可以参考下面这篇文章：

科技猛兽：PyTorch 77.结合apex支持的分布式训练，同步bn，以及混合精度在这篇文章中我们使用8步法结合apex支持的分布式训练，同步bn，以及混合精度：

1. 先罗列自己网络的参数：

1
2
3
4
5
6
7
def parse():
    parser = argparse.ArgumentParser()
    parser.add_argument('--local_rank', type=int, default=0)
    ...
    ...
    args = parser.parse_args()
    return args

local_rank指定了输出设备，默认为GPU可用列表中的第一个GPU。这里这个是必须加的。原因后面讲

2. 在主函数中开头写：

1
2
3
4
5
6
7
8
def main():
  args = parse()
  torch.cuda.set_device(args.local_rank)  # 必须写！，还必须在下一句的前面，
  #torch.utils.launch也需要set_device， 所以必须写
  torch.distributed.init_process_group(
        'nccl',
        init_method='env://'
    )

3. 导入数据接口，这里有一点不一样。需要用一个DistributedSampler：

1
2
3
4
5
6
dataset = DAVIS2017(root, 'training')
num_workers = 4 if cuda else 0
# 多了一个DistributedSampler，作为dataloader的sampler
train_sampler  = torch.utils.data.distributed.DistributedSampler(dataset)
loader = DataLoader(dataset,batch_size=batchsize,shuffle=False, num_workers=num_workers,pin_memory=cuda,
                                     drop_last=True, sampler=train_sampler)

4. 之后定义模型：

1
2
3
4
5
6
net = XXXNet(using_amp=True)
net.train()
net = convert_syncbn_model(net)  # 用apex支持的方法，使得普通bn成为同步bn。
# 切记在网络实现中，不要使用torch自带的SyncBatchnorm。
device = torch.device('cuda:{}'.format(args.local_rank))
net = net.to(device)  # 把模型搬运到第一块GPU上

5. 定义优化器，损失函数，定义优化器一定要在把模型搬运到GPU之后：

1
2
3
opt = Adam([{'params': params_low_lr, 'lr': 4e-5},
         {'params': params_high_lr, 'lr': 1e-4}], weight_decay=settings.WEIGHT_DECAY)
crit = nn.BCELoss().to(device)

6. 多GPU设置：

1
2
3
net, opt = amp.initialize(net, opt, opt_level="O1")  # 字母小写o,不是零。
# 关于initialize用法，见上一篇博客。
net = DDP(net, delay_allreduce=True)  # 必须在initialze之后

7. 记得loss要这么用：

1
2
3
4
5
opt.zero_grad()
# loss.backward()
with amp.scale_loss(loss, opt) as scaled_loss:
     scaled_loss.backward()
opt.step()

8. 然后在代码底部加入：

1
2
if __name__ == '__main__':
    main()

那么这个train.py大体上依然遵循这8步：

https://github.com/rwightman/pytorch-image-models/blob/master/train.py1. 通过命令行解析定义各种超参数，包括：

Dataset / Model parameters，比如：data，–model，–pretrained等等。
Optimizer parameters，比如：–opt，–opt-eps，–momentum等等。
Learning rate schedule parameters，比如：–sched，–lr，–epochs，–start-epoch，–decay-epochs，–decay-rate等等。
Augmentation & regularization parameters，比如：–mixup，–hflip，–vflip，–cutmix，–drop等等。
Batch norm parameters，比如：–bn-tf，–bn-momentum，–sync-bn，–dist-bn，–split-bn等等。
Model Exponential Moving Average parameters，比如：–model-ema，–model-ema-force-cpu，–model-ema-decay等等。
Misc parameters，比如：–seed，–log-interval，–num-gpu，–save-images，amp，–apex-amp，–native-amp，–output，–local_rank等等。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
def _parse_args():
    # Do we have a config file to parse?
    args_config, remaining = config_parser.parse_known_args()
    if args_config.config:
        with open(args_config.config, 'r') as f:
            cfg = yaml.safe_load(f)
            parser.set_defaults(**cfg)

    # The main arg parser parses the rest of the args, the usual
    # defaults will have been overridden if config file specified.
    args = parser.parse_args(remaining)

    # Cache the args as a text string to save them in the output dir later
    args_text = yaml.safe_dump(args.__dict__, default_flow_style=False)
    return args, args_text

2. 分布式命令：

1
2
3
4
5
6
7
8
9
10
11
12
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0  # global rank
if args.distributed:
    args.num_gpu = 1
    args.device = 'cuda:%d' % args.local_rank
    torch.cuda.set_device(args.local_rank)
    args.world_size = int(os.environ['WORLD_SIZE'])
    args.rank = int(os.environ['RANK'])
    torch.distributed.init_process_group(backend='nccl', init_method='env://', rank=args.rank, world_size=args.world_size)
    args.world_size = torch.distributed.get_world_size()
    args.rank = torch.distributed.get_rank()

3. 导入数据接口，这里有一点不一样。需要用一个DistributedSampler：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
loader_train = create_loader(
    dataset_train,
    input_size=data_config['input_size'],
    batch_size=args.batch_size,
    is_training=True,
    use_prefetcher=args.prefetcher,
    no_aug=args.no_aug,
    re_prob=args.reprob,
    re_mode=args.remode,
    re_count=args.recount,
    re_split=args.resplit,
    scale=args.scale,
    ratio=args.ratio,
    hflip=args.hflip,
    vflip=args.vflip,
    color_jitter=args.color_jitter,
    auto_augment=args.aa,
    num_aug_splits=num_aug_splits,
    interpolation=train_interpolation,
    mean=data_config['mean'],
    std=data_config['std'],
    num_workers=args.workers,
    distributed=args.distributed,
    collate_fn=collate_fn,
    pin_memory=args.pin_mem,
    use_multi_epochs_loader=args.use_multi_epochs_loader,
    repeated_aug=args.repeated_aug
)

4. 之后定义模型：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
    model = create_model(
        args.model,
        pretrained=args.pretrained,
        num_classes=args.num_classes,
        drop_rate=args.drop,
        drop_connect_rate=args.drop_connect,  # DEPRECATED, use drop_path
        drop_path_rate=args.drop_path,
        drop_block_rate=args.drop_block,
        global_pool=args.gp,
        bn_tf=args.bn_tf,
        bn_momentum=args.bn_momentum,
        bn_eps=args.bn_eps,
        scriptable=args.torchscript,
        checkpoint_path=args.initial_checkpoint)
--------------------------------------------------------------------------
    # move model to GPU, enable channels last layout if set
    model.cuda()
    if args.channels_last:
        model = model.to(memory_format=torch.channels_last)
--------------------------------------------------------------------------
    # setup synchronized BatchNorm for distributed training
    if args.distributed and args.sync_bn:
        assert not args.split_bn
        if has_apex and use_amp != 'native':
            # Apex SyncBN preferred unless native amp is activated
            model = convert_syncbn_model(model)
        else:
            model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
        if args.local_rank == 0:
            _logger.info(
                'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
                'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')



5. 定义优化器，损失函数，定义优化器一定要在把模型搬运到GPU之后：

1
optimizer = create_optimizer(args, model)

6. 多GPU设置：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
    # setup automatic mixed-precision (AMP) loss scaling and op casting
    amp_autocast = suppress  # do nothing
    loss_scaler = None
    if use_amp == 'apex':
        model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
        loss_scaler = ApexScaler()
        if args.local_rank == 0:
            _logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
    elif use_amp == 'native':
        amp_autocast = torch.cuda.amp.autocast
        loss_scaler = NativeScaler()
        if args.local_rank == 0:
            _logger.info('Using native Torch AMP. Training in mixed precision.')
    else:
        if args.local_rank == 0:
            _logger.info('AMP not enabled. Training in float32.')
--------------------------------------------------------------------------
    # setup distributed training
    if args.distributed:
        if has_apex and use_amp != 'native':
            # Apex DDP preferred unless native amp is activated
            if args.local_rank == 0:
                _logger.info("Using NVIDIA APEX DistributedDataParallel.")
            model = ApexDDP(model, delay_allreduce=True)
        else:
            if args.local_rank == 0:
                _logger.info("Using native Torch DistributedDataParallel.")
            model = NativeDDP(model, device_ids=[args.local_rank])  # can use device str in Torch >= 1.1
        # NOTE: EMA model does not need to be wrapped by DDP

7. 记得loss要这么用：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
    # setup automatic mixed-precision (AMP) loss scaling and op casting
    amp_autocast = suppress  # do nothing
    loss_scaler = None
    if use_amp == 'apex':
        model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
        loss_scaler = ApexScaler()
        if args.local_rank == 0:
            _logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
    elif use_amp == 'native':
        amp_autocast = torch.cuda.amp.autocast
        loss_scaler = NativeScaler()
        if args.local_rank == 0:
            _logger.info('Using native Torch AMP. Training in mixed precision.')
    else:
        if args.local_rank == 0:
            _logger.info('AMP not enabled. Training in float32.')
--------------------------------------------------------------------------
        optimizer.zero_grad()
        if loss_scaler is not None:
            loss_scaler(
                loss, optimizer,
                clip_grad=args.clip_grad, clip_mode=args.clip_mode,
                parameters=model_parameters(model, exclude_head='agc' in args.clip_mode),
                create_graph=second_order)
        else:
            loss.backward(create_graph=second_order)
            if args.clip_grad is not None:
                dispatch_clip_grad(
                    model_parameters(model, exclude_head='agc' in args.clip_mode),
                    value=args.clip_grad, mode=args.clip_mode)
            optimizer.step()

8. 然后在代码底部加入：

1
2
if __name__ == '__main__':
    main()

总结

本文简要介绍了优秀的PyTorch Image Model 库：timm库以及其中的 vision transformer 代码和训练代码。 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。本文的目的是为学者介绍一个优秀的 vision transformer 的PyTorch实现，以便更快地开展相关实验。

References