Transformer主干网络——TNT保姆级解析

前言

论文地址：arxiv
代码地址：github
接收单位：NeurIPS 2021

系列文章

Transformer主干网络——ViT保姆级解析
Transformer主干网络——DeiT保姆级解析
Transformer主干网络——T2T-ViT保姆级解析
Transformer主干网络——TNT保姆级解析
Transformer主干网络——PVT_V1保姆级解析
Transformer主干网络——PVT_V2保姆级解析
Transformer主干网络——Swin保姆级解析
Transformer主干网络——PatchConvNet保姆级解析
持续更新…

动机

作者出发点也是ViT的Patch embed的不足(ViT留的这个坑真的是给后人留了发论文的机会…)，作者将一张图片比作文章，一篇文章由句子（patchs）构成，一个句子由单词（pixels）构成。以往的attention只对句子和句子间的关系进行建模，作者认为句子内单词和单词间的关系也不可以忽视。

• 基于以上的动机作者设计了transformer in transformer的结构，相当于transformer嵌套，内层的transformer用来建模句子内单词和单词间的attention，外层的transformer用来建模句子和句子间的attention。
• 因为引入了TNT嵌套的transformer结构所以作者设计了针对TNT的位置编码。

网络分析

首先看下完整的结构示意图，源自作者论文：

为了方便分析，本文给网络输入的数据大小为(1,3,224,224)

1、对于输入的224 * 224的图像，作者先执行了nn.Unfold的操作：
self.unfold = nn.Unfold(kernel_size=patch_size=16, stride=patch_size=16)
# input_shape (1,3,224,224)
x = self.unfold(x)
# output_shape (1,768,196)
这里196=14*14，196表示用16*16的patch可以遍历224*224次数（224/16 * 224/16）
然后768=16*16*3，前两个16是一个patch的长宽，3是彩色的三通道。相当于每个patch的pixel数。
这一步Unfold相当于卷积的“卷”过程用滑窗取出所有的块。

那么对应到图片上就是第一步将原图用分割成很多个patch了：

2、图片转单词级别的tokens：
self.proj = nn.Conv2d(in_chans, inner_dim, kernel_size=7, padding=3, stride=inner_stride)
# input_shape (1,196，768)
x = x.transpose(1, 2).reshape(B * self.num_patches, C, *self.patch_size)
# shape (196,3,16,16) 
x = self.proj(x)
# shape (196,40,4,4)
x = x.reshape(B * self.num_patches, self.inner_dim, -1).transpose(1, 2)
# output_shape (196，16，40)
(196,3,16,16) 这个很好理解这是196个patch，每个patch是彩色的所以3通道，长宽是16
(196,40,4,4)这是卷积后的结果
(196，16，40)是变形后的结构，可以理解一张图有196个句子级别的tokens，每个句子级别的token有16个单词级别的tokens，每个单词级别的token的dim是40.

以上的操作得到了单词级别的tokens，对应图中所有的圈起来部分：

3、单词级别的tokens矩阵加上位置编码：
inner_tokens = self.patch_embed(x) + self.inner_pos
按位相加shape不变!
这个inner_pos和ViT一样的，硬训一个位置编码。

4、图片转句子级别的tokens：
由于已经有单词级别的tokens了所以句子级别的tokens可以直接由单词级别的tokens组成：
self.proj_norm1 = norm_layer(num_words * inner_dim)
self.proj = nn.Linear(num_words * inner_dim, outer_dim)
self.proj_norm2 = norm_layer(outer_dim)
# input_shape (196, 16, 40)
outer_tokens = self.proj_norm2(self.proj(self.proj_norm1(inner_tokens.reshape(B, self.num_patches, -1))))
# output_shape (1, 196, 640)
(1, 196, 640)就是一张图由196个句子级别的tokens构成，每个tokens的dim是640

那么上述就完成了本文最重要的部分，组成句子级别的tokens和单词级别的tokens！对应图示：

5、句子级别的tokens拼接分类头用于分类：
# input_shape (1,196,640)
outer_tokens = torch.cat((self.cls_token.expand(B, -1, -1), outer_tokens), dim=1)
# output_shape (1,197,640)

6、句子级别的tokens矩阵加上位置编码：
outer_tokens = outer_tokens + self.outer_pos
按位相加shape不变!
这个outer_pos和ViT一样的，硬训一个位置编码。
outer_tokens = self.pos_drop(outer_tokens)
加drop path

7、接下来就是把word tokens和sentence tokens一起送到block中，
这个block和ViT的block一样，但是因为这里的输出有两种tokens，
所以有一下的变动，看图！整个block的返回是红色圈出来的两个部分！

完结撒花~
再好的模型，也得经得起业务的考验，自己数据集上的实验效果会在后续更新出…

完整测试代码

# 2021.06.15-Changed for implementation of TNT model
#            Huawei Technologies Co., Ltd. <foss@huawei.com>
""" Vision Transformer (ViT) in PyTorch
A PyTorch implement of Vision Transformers as described in
'An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale' - https://arxiv.org/abs/2010.11929
The official jax code is released and available at https://github.com/google-research/vision_transformer
Status/TODO:
* Models updated to be compatible with official impl. Args added to support backward compat for old PyTorch weights.
* Weights ported from official jax impl for 384x384 base and small models, 16x16 and 32x32 patches.
* Trained (supervised on ImageNet-1k) my custom 'small' patch model to 77.9, 'base' to 79.4 top-1 with this code.
* Hopefully find time and GPUs for SSL or unsupervised pretraining on OpenImages w/ ImageNet fine-tune in future.
Acknowledgments:
* The paper authors for releasing code and weights, thanks!
* I fixed my class token impl based on Phil Wang's https://github.com/lucidrains/vit-pytorch ... check it out
for some einops/einsum fun
* Simple transformer style inspired by Andrej Karpathy's https://github.com/karpathy/minGPT
* Bert reference code checks against Huggingface Transformers and Tensorflow Bert
Hacked together by / Copyright 2020 Ross Wightman
"""
import torch
import torch.nn as nn
from functools import partial
import mathfrom timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.models.helpers import load_pretrained
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from timm.models.resnet import resnet26d, resnet50d
from timm.models.registry import register_modeldef _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}default_cfgs = {'tnt_s_patch16_224': _cfg(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5),),'tnt_b_patch16_224': _cfg(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5),),
}def make_divisible(v, divisor=8, min_value=None):min_value = min_value or divisornew_v = max(min_value, int(v + divisor / 2) // divisor * divisor)# Make sure that round down does not go down by more than 10%.if new_v < 0.9 * v:new_v += divisorreturn new_vclass 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_featureshidden_features = hidden_features or in_featuresself.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 xclass SE(nn.Module):def __init__(self, dim, hidden_ratio=None):super().__init__()hidden_ratio = hidden_ratio or 1self.dim = dimhidden_dim = int(dim * hidden_ratio)self.fc = nn.Sequential(nn.LayerNorm(dim),nn.Linear(dim, hidden_dim),nn.ReLU(inplace=True),nn.Linear(hidden_dim, dim),nn.Tanh())def forward(self, x):a = x.mean(dim=1, keepdim=True)  # B, 1, Ca = self.fc(a)x = a * xreturn xclass Attention(nn.Module):def __init__(self, dim, hidden_dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):super().__init__()self.hidden_dim = hidden_dimself.num_heads = num_headshead_dim = hidden_dim // num_headsself.head_dim = head_dim# NOTE scale factor was wrong in my original version, can set manually to be compat with prev weightsself.scale = qk_scale or head_dim ** -0.5self.qk = nn.Linear(dim, hidden_dim * 2, bias=qkv_bias)self.v = nn.Linear(dim, dim, bias=qkv_bias)self.attn_drop = nn.Dropout(attn_drop, inplace=True)self.proj = nn.Linear(dim, dim)self.proj_drop = nn.Dropout(proj_drop, inplace=True)def forward(self, x):B, N, C = x.shapeqk = self.qk(x).reshape(B, N, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)q, k = qk[0], qk[1]  # make torchscript happy (cannot use tensor as tuple)v = self.v(x).reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)attn = (q @ k.transpose(-2, -1)) * self.scaleattn = attn.softmax(dim=-1)attn = self.attn_drop(attn)x = (attn @ v).transpose(1, 2).reshape(B, N, -1)x = self.proj(x)x = self.proj_drop(x)return xclass Block(nn.Module):""" TNT Block"""def __init__(self, outer_dim, inner_dim, outer_num_heads, inner_num_heads, num_words, 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, se=0):super().__init__()self.has_inner = inner_dim > 0if self.has_inner:# Innerself.inner_norm1 = norm_layer(inner_dim)self.inner_attn = Attention(inner_dim, inner_dim, num_heads=inner_num_heads, qkv_bias=qkv_bias,qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)self.inner_norm2 = norm_layer(inner_dim)self.inner_mlp = Mlp(in_features=inner_dim, hidden_features=int(inner_dim * mlp_ratio),out_features=inner_dim, act_layer=act_layer, drop=drop)self.proj_norm1 = norm_layer(num_words * inner_dim)self.proj = nn.Linear(num_words * inner_dim, outer_dim, bias=False)self.proj_norm2 = norm_layer(outer_dim)# Outerself.outer_norm1 = norm_layer(outer_dim)self.outer_attn = Attention(outer_dim, outer_dim, num_heads=outer_num_heads, qkv_bias=qkv_bias,qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()self.outer_norm2 = norm_layer(outer_dim)self.outer_mlp = Mlp(in_features=outer_dim, hidden_features=int(outer_dim * mlp_ratio),out_features=outer_dim, act_layer=act_layer, drop=drop)# SEself.se = seself.se_layer = Noneif self.se > 0:self.se_layer = SE(outer_dim, 0.25)def forward(self, inner_tokens, outer_tokens):if self.has_inner:print(inner_tokens.shape)inner_tokens = inner_tokens + self.drop_path(self.inner_attn(self.inner_norm1(inner_tokens)))  # B*N, k*k, cprint(inner_tokens.shape)inner_tokens = inner_tokens + self.drop_path(self.inner_mlp(self.inner_norm2(inner_tokens)))  # B*N, k*k, cprint(inner_tokens.shape)B, N, C = outer_tokens.size()print(inner_tokens.reshape(B, N - 1, -1).shape)outer_tokens[:, 1:] = outer_tokens[:, 1:] + self.proj_norm2(self.proj(self.proj_norm1(inner_tokens.reshape(B, N - 1, -1))))  # B, N, Cif self.se > 0:outer_tokens = outer_tokens + self.drop_path(self.outer_attn(self.outer_norm1(outer_tokens)))tmp_ = self.outer_mlp(self.outer_norm2(outer_tokens))outer_tokens = outer_tokens + self.drop_path(tmp_ + self.se_layer(tmp_))else:outer_tokens = outer_tokens + self.drop_path(self.outer_attn(self.outer_norm1(outer_tokens)))outer_tokens = outer_tokens + self.drop_path(self.outer_mlp(self.outer_norm2(outer_tokens)))return inner_tokens, outer_tokensclass PatchEmbed(nn.Module):""" Image to Visual Word Embedding"""def __init__(self, img_size=224, patch_size=16, in_chans=3, outer_dim=768, inner_dim=24, inner_stride=4):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_sizeself.patch_size = patch_sizeself.num_patches = num_patchesself.inner_dim = inner_dimself.num_words = math.ceil(patch_size[0] / inner_stride) * math.ceil(patch_size[1] / inner_stride)self.unfold = nn.Unfold(kernel_size=patch_size, stride=patch_size)self.proj = nn.Conv2d(in_chans, inner_dim, kernel_size=7, padding=3, stride=inner_stride)def forward(self, x):B, C, H, W = x.shape# FIXME look at relaxing size constraintsassert 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]})."print(x.shape)x = self.unfold(x)  # B, Ck2, Nprint(x.shape)x = x.transpose(1, 2).reshape(B * self.num_patches, C, *self.patch_size)  # B*N, C, 16, 16print(x.shape)x = self.proj(x)  # B*N, C, 8, 8print(x.shape)x = x.reshape(B * self.num_patches, self.inner_dim, -1).transpose(1, 2)  # B*N, 8*8, Cprint(x.shape)return xclass TNT(nn.Module):""" TNT (Transformer in Transformer) for computer vision"""def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, outer_dim=768, inner_dim=48,depth=12, outer_num_heads=12, inner_num_heads=4, mlp_ratio=4., qkv_bias=False, qk_scale=None,drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, inner_stride=4, se=0):super().__init__()self.num_classes = num_classesself.num_features = self.outer_dim = outer_dim  # num_features for consistency with other modelsself.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=in_chans, outer_dim=outer_dim,inner_dim=inner_dim, inner_stride=inner_stride)self.num_patches = num_patches = self.patch_embed.num_patchesnum_words = self.patch_embed.num_wordsself.proj_norm1 = norm_layer(num_words * inner_dim)self.proj = nn.Linear(num_words * inner_dim, outer_dim)self.proj_norm2 = norm_layer(outer_dim)self.cls_token = nn.Parameter(torch.zeros(1, 1, outer_dim))self.outer_tokens = nn.Parameter(torch.zeros(1, num_patches, outer_dim), requires_grad=False)self.outer_pos = nn.Parameter(torch.zeros(1, num_patches + 1, outer_dim))self.inner_pos = nn.Parameter(torch.zeros(1, num_words, inner_dim))self.pos_drop = nn.Dropout(p=drop_rate)dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rulevanilla_idxs = []blocks = []for i in range(depth):if i in vanilla_idxs:blocks.append(Block(outer_dim=outer_dim, inner_dim=-1, outer_num_heads=outer_num_heads, inner_num_heads=inner_num_heads,num_words=num_words, 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, se=se))else:blocks.append(Block(outer_dim=outer_dim, inner_dim=inner_dim, outer_num_heads=outer_num_heads,inner_num_heads=inner_num_heads,num_words=num_words, 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, se=se))self.blocks = nn.ModuleList(blocks)self.norm = norm_layer(outer_dim)# NOTE as per official impl, we could have a pre-logits representation dense layer + tanh here# self.repr = nn.Linear(outer_dim, representation_size)# self.repr_act = nn.Tanh()# Classifier headself.head = nn.Linear(outer_dim, num_classes) if num_classes > 0 else nn.Identity()trunc_normal_(self.cls_token, std=.02)trunc_normal_(self.outer_pos, std=.02)trunc_normal_(self.inner_pos, 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)@torch.jit.ignoredef no_weight_decay(self):return {'outer_pos', 'inner_pos', 'cls_token'}def get_classifier(self):return self.headdef reset_classifier(self, num_classes, global_pool=''):self.num_classes = num_classesself.head = nn.Linear(self.outer_dim, num_classes) if num_classes > 0 else nn.Identity()def forward_features(self, x):B = x.shape[0]inner_tokens = self.patch_embed(x) + self.inner_pos  # B*N, 8*8, Couter_tokens = self.proj_norm2(self.proj(self.proj_norm1(inner_tokens.reshape(B, self.num_patches, -1))))outer_tokens = torch.cat((self.cls_token.expand(B, -1, -1), outer_tokens), dim=1)outer_tokens = outer_tokens + self.outer_posouter_tokens = self.pos_drop(outer_tokens)for blk in self.blocks:inner_tokens, outer_tokens = blk(inner_tokens, outer_tokens)outer_tokens = self.norm(outer_tokens)return outer_tokens[:, 0]def forward(self, x):x = self.forward_features(x)x = self.head(x)return xdef _conv_filter(state_dict, patch_size=16):""" convert patch embedding weight from manual patchify + linear proj to conv"""out_dict = {}for k, v in state_dict.items():if 'patch_embed.proj.weight' in k:v = v.reshape((v.shape[0], 3, patch_size, patch_size))out_dict[k] = vreturn out_dict@register_model
def tnt_s_patch16_224(pretrained=False, **kwargs):patch_size = 16inner_stride = 4outer_dim = 384inner_dim = 24outer_num_heads = 6inner_num_heads = 4outer_dim = make_divisible(outer_dim, outer_num_heads)inner_dim = make_divisible(inner_dim, inner_num_heads)model = TNT(img_size=224, patch_size=patch_size, outer_dim=outer_dim, inner_dim=inner_dim, depth=12,outer_num_heads=outer_num_heads, inner_num_heads=inner_num_heads, qkv_bias=False,inner_stride=inner_stride, **kwargs)model.default_cfg = default_cfgs['tnt_s_patch16_224']if pretrained:load_pretrained(model, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3), filter_fn=_conv_filter)return model@register_model
def tnt_b_patch16_224(pretrained=False, **kwargs):patch_size = 16inner_stride = 4outer_dim = 640inner_dim = 40outer_num_heads = 10inner_num_heads = 4outer_dim = make_divisible(outer_dim, outer_num_heads)inner_dim = make_divisible(inner_dim, inner_num_heads)model = TNT(img_size=224, patch_size=patch_size, outer_dim=outer_dim, inner_dim=inner_dim, depth=12,outer_num_heads=outer_num_heads, inner_num_heads=inner_num_heads, qkv_bias=False,inner_stride=inner_stride, **kwargs)model.default_cfg = default_cfgs['tnt_b_patch16_224']if pretrained:load_pretrained(model, num_classes=model.num_classes, in_chans=kwargs.get('in_chans', 3), filter_fn=_conv_filter)return modelif __name__ == '__main__':model = tnt_b_patch16_224(pretrained=False)input = torch.randn((1,3,224,224))output = model(input)print(output.shape)