• vision transformer 详解

    文章链接：https://arxiv.org/abs/2010.11929

     代码地址：GitHub - google-research/vision_transformer

    Pytorch实现代码： https://github.com/WZMIAOMIAO/deep-learning-for-image-processing/tree/master/pytorch_classification/vision_transformer
    Tensorflow2实现代码：https://github.com/WZMIAOMIAO/deep-learning-for-image-processing/tree/master/tensorflow_classification/vision_transformer
    在bilibili上的视频讲解：11.1 Vision Transformer(vit)网络详解_哔哩哔哩_bilibili

     一.引言

    在 "An Image Is Worth 16x16 Words: Transformers For Image Recognition At Scale" 论文的实验部分，与其他方法相比，该方法取得了以下突破性的成果：

    1. 在大规模图像分类任务上超越传统方法：论文中的方法在 ImageNet-1K 数据集上进行了实验，并与传统的卷积神经网络（CNN）进行了比较。结果显示，该方法在准确性方面超越了传统的CNN模型，取得了更好的图像分类性能。

    2. 大规模预训练模型的有效性：论文中的方法使用大规模的图像数据集进行了预训练，并通过自监督学习的方式学习图像的特征表示。这种预训练方法使得Transformer模型能够捕捉到广泛的视觉特征，并在具体任务的微调中展现出优势。

    3. 有效处理大尺寸图像的能力：传统的CNN模型在处理大尺寸图像时可能面临计算和内存限制的问题。然而，论文中的方法通过将图像分割成较小的块，并利用Transformer模型对这些块进行编码，展示了对大尺寸图像的高效处理能力。这使得该方法在处理高分辨率图像时具有优势，并能够在保持准确性的同时提高计算效率。

     二.ViT模型架构

    其原始架构图如下所示，可以看到首先输入图片分为很多 patch，论文中为 16。将 patch 输入一个 Linear Projection of Flattened Patches 这个 Embedding 层，就会得到一个个向量，通常就称作 token。紧接着在一系列 token 的前面加上加上一个新的 token（类别token，有点像输入给 Transformer Decoder 的 START，就是对应着 * 那个位置），此外还需要加上位置的信息，对应着 0～9。然后输入到 Transformer Encoder 中，对应着右边的图，将 block 重复堆叠 L 次。Transformer Encoder 有多少个输入就有多少个输出。最后只进行分类，所以将 class 位置对应的输出输入 MLP Head 进行预测分类输出。

     

     2.1embedding层

    接下来对每个模块进行细讲，首先是 Embedding 层。对于标准的 Transformer 模块，要求的输入是 token 向量的序列，即二维矩阵 [num_token, token_dim]。在具体的代码实现过程中呢，我们实际是通过一个卷积层来实现以 ViT-B/16 为例，使用卷积核大小为 16 × 16 16 \times 1616×16, stride 为 16，卷积核个数为 768 来实现的，即 [224,224,3] --> [14,14,768] --> [196, 768]。即一共 196 个token，每个 token 向量长度为 768。此外我们还需要加上一个类别的 token，为此我们实际上是初始化了一个可训练的参数 [1, 768]，将其与 token 序列进行拼接得到 Cat([1, 768], [196,768]) --> [197, 768]。然后再叠加上位置编码 Position Embedding: [197,768] --> [197, 768]。

     再详细考虑下 Position Embedding，如果不是用 Position Embedding 得到的结果是 0.61382，使用一维的位置编码得到的结果是 0.64206，明显比不使用位置编码高了三个百分点。使用 2D 以及相对位置编码其实和 1D 差不多啊。论文中也提到说 the difference in how to encoder spatial information is less important，即位置编码的差异其实不是特别重要。1D 的话，简单效果好参数少，所以默认使用 1D 的位置编码。

     论文中有给这样一个图，我们训练得到的位置编码与其他位置编码之间的余弦相似度。这里的 patches 大小是 32 × 32 32 \times 3232×32 的，224 / 32 = 7 224/32=7224/32=7，所以这里的大小是 7 × 7 7 \times 77×7。这张图怎么理解呢？我们会在每个 token 上叠加一个位置编码，中间那个图的 49 个小图中，每个小图其实也是 7 × 7 7 \times 77×7 的。左上角第一行第一个 patch 的位置编码与自己的位置编码是一样的，所以余弦相似度是1，所以左上角是黄色。然后在与其他位置编码进行计算。就得到了左上角的小图。其他的也都是类似的规律。注意，这个是学出来的。

     2.2 Transformer Encoder 层

    Transformer Encoder 就是将 Encoder Block 重复堆叠 L 次。我们来看看单个 Encoder Block。首先输入一个 Norm 层，这里的 Norm 指的是 Layer Normalization 层（有论文比较了 BN 在 transformer 中为什么不好，不如 LN ｜这里先 Norm 再 Multihead Attention 也是有论文研究的，原始的 Transformer 先 Attention 再 Norm，此外这个先 Norm 再操作和 DenseNet 的先 BN 再 Conv 异曲同工）。经过 LN 后经过 Multi-Head Attention，然后源码经过 Dropout 层，有些复现大神使用的是 DropPath 方法，根据以往的经验可能使用后者会更好一点。然后残差之后经过 LN，MLP Block，Dropout/DropPath 之后残差即可。

    MLP Block 其实也很简单，就是一个全连接，GELU 激活函数，Dropout，全连接，Dropout。需要注意第一个全连接层的节点个数是输入向量长度的 4 倍，第二个全连接层会还原会原来的大小。

     有一个地方要注意，在 Transformer Encoder 前有个 Dropout 层，在之后有一个 Layer Norm 层，这些在图中还没有画出来的。在 Transformer Encoder 前有个 Dropout 层，对此我的理解是在原图上随机加 Mask 遮挡，然后依然要进行分类。


    2.3 MLP Head 层

    在训练 ImageNet21K 时候是由 Linear + tanh 激活函数 + Linear 构成的。但是迁移到 ImageNet1k 之后或者做迁移学习时，其实只需要一个 Linear 就足够了。（获得类别概率需要一个 softmax）

     2.4 ViT B/16

    我们来从头梳理一次 ViT B/16 的结构，假设输入图为 224 × 224 × 3 224 \times 224 \times 3224×224×3，首先经过一个卷积层，然后进行高度和宽度方向的展平处理。紧接着 concat 一个 class token，再加上 Position Embedding 的相加操作，这里的 Position Embedding 也是可训练的参数。经过 Dropout 之后输入 12 个堆叠的 Encoder Block。Encoder 输出经过 LN 得到的输出为 197 × 768 197 \times 768197×768，即是不变的。然后我们提取第一个 class token 对应的输出，切片之后即变成了 1 × 768 1 \times 7681×768，将其输入 MLP Head 中。如果在 ImageNet21K 预训练的时候，Pre-Logits 就是一个全连接层，tanh 激活函数。如果是在 ImageNet1k 或者自己的数据集上的时候训练的时候，可以不要这个 Pre-Logits。

    2.5 ViT 模型参数

     

    我们来看看论文给出的 ViT 模型的参数。ViT B 对应的就是 ViT-Base，ViT L 对应的是 ViT-Large，ViT H 对应的是 ViT-Huge。patch size 是图片切片大小（源码中还有 32 × 32 32 \times 3232×32 的）；layers 则是 encoder block 堆叠的次数；Hidden size 是 token 向量的长度；MLP size 是 Hidden size 的四倍，即 Encoder block 中 MLP block 第一个全连接层节点个数；Heads 则是 Multi-head Attention 中 heads 的个数。

    三.hybrid混合模型 

    我们来看看 CNN 和 Transformer 的混合模型。首先用传统的神经网络 backbone 来提取特征，然后再通过 ViT 模型进一步得到最终的结果。这里的特征提取部分采用的是 ResNet50 网络，但是和原来的有所不同，第一点是采用 stdConv2d，第二点则是使用GN而非BN，第三点是将 stage4 中的 3 个 block 移动到 stage3 中。R50 backbone 的输出为 14 × 14 × 1024 14 \times 14 \times 102414×14×1024，然后通过 1 × 1 1 \times 11×1 卷积变为 14 × 14 × 768 14 \times 14 \times 76814×14×768，然后进行展平处理就得到 token 了。之后就是和 ViT 一摸一样的了。

     结果可见，混合模型比纯 transformer 模型的效果会好一些，这也是迁移学习之后的结果。在少量微调中混合模型占有，但是随着迭代次数的上升，纯 transformer 也能达到混合模型的效果，例如 14 个 epoches 时 ViT-L/16 和 Res50x1+ViT-L/16 就基本一样了。

    代码

    1. """
    2. original code from rwightman:
    3. https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
    4. """
    5. from functools import partial
    6. from collections import OrderedDict
    7.  
    8. import torch
    9. import torch.nn as nn
    10.  
    11.  
    12. def drop_path(x, drop_prob: float = 0., training: bool = False):
    13.     """
    14.     Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
    15.     This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
    16.     the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    17.     See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
    18.     changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
    19.     'survival rate' as the argument.
    20.     """
    21.     if drop_prob == 0. or not training:
    22.         return x
    23.     keep_prob = 1 - drop_prob
    24.     shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    25.     random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
    26.     random_tensor.floor_()  # binarize
    27.     output = x.div(keep_prob) * random_tensor
    28.     return output
    29.  
    30.  
    31. class DropPath(nn.Module):
    32.     """
    33.     Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    34.     """
    35.     def __init__(self, drop_prob=None):
    36.         super(DropPath, self).__init__()
    37.         self.drop_prob = drop_prob
    38.  
    39.     def forward(self, x):
    40.         return drop_path(x, self.drop_prob, self.training)
    41.  
    42.  
    43. class PatchEmbed(nn.Module):
    44.     """
    45.     2D Image to Patch Embedding
    46.     """
    47.     def __init__(self, img_size=224, patch_size=16, in_c=3, embed_dim=768, norm_layer=None):
    48.         super().__init__()
    49.         img_size = (img_size, img_size)
    50.         patch_size = (patch_size, patch_size)
    51.         self.img_size = img_size
    52.         self.patch_size = patch_size
    53.         self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
    54.         self.num_patches = self.grid_size[0] * self.grid_size[1]
    55.  
    56.         self.proj = nn.Conv2d(in_c, embed_dim, kernel_size=patch_size, stride=patch_size)
    57.         self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
    58.  
    59.     def forward(self, x):
    60.         B, C, H, W = x.shape
    61.         assert H == self.img_size[0] and W == self.img_size[1], \
    62.             f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
    63.  
    64.         # flatten: [B, C, H, W] -> [B, C, HW]
    65.         # transpose: [B, C, HW] -> [B, HW, C]
    66.         x = self.proj(x).flatten(2).transpose(1, 2)
    67.         x = self.norm(x)
    68.         return x
    69.  
    70.  
    71. class Attention(nn.Module):
    72.     def __init__(self,
    73.                  dim,   # 输入token的dim
    74.                  num_heads=8,
    75.                  qkv_bias=False,
    76.                  qk_scale=None,
    77.                  attn_drop_ratio=0.,
    78.                  proj_drop_ratio=0.):
    79.         super(Attention, self).__init__()
    80.         self.num_heads = num_heads
    81.         head_dim = dim // num_heads
    82.         self.scale = qk_scale or head_dim ** -0.5
    83.         self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
    84.         self.attn_drop = nn.Dropout(attn_drop_ratio)
    85.         self.proj = nn.Linear(dim, dim)
    86.         self.proj_drop = nn.Dropout(proj_drop_ratio)
    87.  
    88.     def forward(self, x):
    89.         # [batch_size, num_patches + 1, total_embed_dim]
    90.         B, N, C = x.shape
    91.  
    92.         # qkv(): -> [batch_size, num_patches + 1, 3 * total_embed_dim]
    93.         # reshape: -> [batch_size, num_patches + 1, 3, num_heads, embed_dim_per_head]
    94.         # permute: -> [3, batch_size, num_heads, num_patches + 1, embed_dim_per_head]
    95.         qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
    96.         # [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
    97.         q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
    98.  
    99.         # transpose: -> [batch_size, num_heads, embed_dim_per_head, num_patches + 1]
    100.         # @: multiply -> [batch_size, num_heads, num_patches + 1, num_patches + 1]
    101.         attn = (q @ k.transpose(-2, -1)) * self.scale
    102.         attn = attn.softmax(dim=-1)
    103.         attn = self.attn_drop(attn)
    104.  
    105.         # @: multiply -> [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
    106.         # transpose: -> [batch_size, num_patches + 1, num_heads, embed_dim_per_head]
    107.         # reshape: -> [batch_size, num_patches + 1, total_embed_dim]
    108.         x = (attn @ v).transpose(1, 2).reshape(B, N, C)
    109.         x = self.proj(x)
    110.         x = self.proj_drop(x)
    111.         return x
    112.  
    113.  
    114. class Mlp(nn.Module):
    115.     """
    116.     MLP as used in Vision Transformer, MLP-Mixer and related networks
    117.     """
    118.     def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
    119.         super().__init__()
    120.         out_features = out_features or in_features
    121.         hidden_features = hidden_features or in_features
    122.         self.fc1 = nn.Linear(in_features, hidden_features)
    123.         self.act = act_layer()
    124.         self.fc2 = nn.Linear(hidden_features, out_features)
    125.         self.drop = nn.Dropout(drop)
    126.  
    127.     def forward(self, x):
    128.         x = self.fc1(x)
    129.         x = self.act(x)
    130.         x = self.drop(x)
    131.         x = self.fc2(x)
    132.         x = self.drop(x)
    133.         return x
    134.  
    135.  
    136. class Block(nn.Module):
    137.     def __init__(self,
    138.                  dim,
    139.                  num_heads,
    140.                  mlp_ratio=4.,
    141.                  qkv_bias=False,
    142.                  qk_scale=None,
    143.                  drop_ratio=0.,
    144.                  attn_drop_ratio=0.,
    145.                  drop_path_ratio=0.,
    146.                  act_layer=nn.GELU,
    147.                  norm_layer=nn.LayerNorm):
    148.         super(Block, self).__init__()
    149.         self.norm1 = norm_layer(dim)
    150.         self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
    151.                               attn_drop_ratio=attn_drop_ratio, proj_drop_ratio=drop_ratio)
    152.         # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
    153.         self.drop_path = DropPath(drop_path_ratio) if drop_path_ratio > 0. else nn.Identity()
    154.         self.norm2 = norm_layer(dim)
    155.         mlp_hidden_dim = int(dim * mlp_ratio)
    156.         self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop_ratio)
    157.  
    158.     def forward(self, x):
    159.         x = x + self.drop_path(self.attn(self.norm1(x)))
    160.         x = x + self.drop_path(self.mlp(self.norm2(x)))
    161.         return x
    162.  
    163.  
    164. class VisionTransformer(nn.Module):
    165.     def __init__(self, img_size=224, patch_size=16, in_c=3, num_classes=1000,
    166.                  embed_dim=768, depth=12, num_heads=12, mlp_ratio=4.0, qkv_bias=True,
    167.                  qk_scale=None, representation_size=None, distilled=False, drop_ratio=0.,
    168.                  attn_drop_ratio=0., drop_path_ratio=0., embed_layer=PatchEmbed, norm_layer=None,
    169.                  act_layer=None):
    170.         """
    171.         Args:
    172.             img_size (int, tuple): input image size
    173.             patch_size (int, tuple): patch size
    174.             in_c (int): number of input channels
    175.             num_classes (int): number of classes for classification head
    176.             embed_dim (int): embedding dimension
    177.             depth (int): depth of transformer
    178.             num_heads (int): number of attention heads
    179.             mlp_ratio (int): ratio of mlp hidden dim to embedding dim
    180.             qkv_bias (bool): enable bias for qkv if True
    181.             qk_scale (float): override default qk scale of head_dim ** -0.5 if set
    182.             representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set
    183.             distilled (bool): model includes a distillation token and head as in DeiT models
    184.             drop_ratio (float): dropout rate
    185.             attn_drop_ratio (float): attention dropout rate
    186.             drop_path_ratio (float): stochastic depth rate
    187.             embed_layer (nn.Module): patch embedding layer
    188.             norm_layer: (nn.Module): normalization layer
    189.         """
    190.         super(VisionTransformer, self).__init__()
    191.         self.num_classes = num_classes
    192.         self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
    193.         self.num_tokens = 2 if distilled else 1
    194.         norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
    195.         act_layer = act_layer or nn.GELU
    196.  
    197.         self.patch_embed = embed_layer(img_size=img_size, patch_size=patch_size, in_c=in_c, embed_dim=embed_dim)
    198.         num_patches = self.patch_embed.num_patches
    199.  
    200.         self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
    201.         self.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None
    202.         self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
    203.         self.pos_drop = nn.Dropout(p=drop_ratio)
    204.  
    205.         dpr = [x.item() for x in torch.linspace(0, drop_path_ratio, depth)]  # stochastic depth decay rule
    206.         self.blocks = nn.Sequential(*[
    207.             Block(dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
    208.                   drop_ratio=drop_ratio, attn_drop_ratio=attn_drop_ratio, drop_path_ratio=dpr[i],
    209.                   norm_layer=norm_layer, act_layer=act_layer)
    210.             for i in range(depth)
    211.         ])
    212.         self.norm = norm_layer(embed_dim)
    213.  
    214.         # Representation layer
    215.         if representation_size and not distilled:
    216.             self.has_logits = True
    217.             self.num_features = representation_size
    218.             self.pre_logits = nn.Sequential(OrderedDict([
    219.                 ("fc", nn.Linear(embed_dim, representation_size)),
    220.                 ("act", nn.Tanh())
    221.             ]))
    222.         else:
    223.             self.has_logits = False
    224.             self.pre_logits = nn.Identity()
    225.  
    226.         # Classifier head(s)
    227.         self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
    228.         self.head_dist = None
    229.         if distilled:
    230.             self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()
    231.  
    232.         # Weight init
    233.         nn.init.trunc_normal_(self.pos_embed, std=0.02)
    234.         if self.dist_token is not None:
    235.             nn.init.trunc_normal_(self.dist_token, std=0.02)
    236.  
    237.         nn.init.trunc_normal_(self.cls_token, std=0.02)
    238.         self.apply(_init_vit_weights)
    239.  
    240.     def forward_features(self, x):
    241.         # [B, C, H, W] -> [B, num_patches, embed_dim]
    242.         x = self.patch_embed(x)  # [B, 196, 768]
    243.         # [1, 1, 768] -> [B, 1, 768]
    244.         cls_token = self.cls_token.expand(x.shape[0], -1, -1)
    245.         if self.dist_token is None:
    246.             x = torch.cat((cls_token, x), dim=1)  # [B, 197, 768]
    247.         else:
    248.             x = torch.cat((cls_token, self.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
    249.  
    250.         x = self.pos_drop(x + self.pos_embed)
    251.         x = self.blocks(x)
    252.         x = self.norm(x)
    253.         if self.dist_token is None:
    254.             return self.pre_logits(x[:, 0])
    255.         else:
    256.             return x[:, 0], x[:, 1]
    257.  
    258.     def forward(self, x):
    259.         x = self.forward_features(x)
    260.         if self.head_dist is not None:
    261.             x, x_dist = self.head(x[0]), self.head_dist(x[1])
    262.             if self.training and not torch.jit.is_scripting():
    263.                 # during inference, return the average of both classifier predictions
    264.                 return x, x_dist
    265.             else:
    266.                 return (x + x_dist) / 2
    267.         else:
    268.             x = self.head(x)
    269.         return x
    270.  
    271.  
    272. def _init_vit_weights(m):
    273.     """
    274.     ViT weight initialization
    275.     :param m: module
    276.     """
    277.     if isinstance(m, nn.Linear):
    278.         nn.init.trunc_normal_(m.weight, std=.01)
    279.         if m.bias is not None:
    280.             nn.init.zeros_(m.bias)
    281.     elif isinstance(m, nn.Conv2d):
    282.         nn.init.kaiming_normal_(m.weight, mode="fan_out")
    283.         if m.bias is not None:
    284.             nn.init.zeros_(m.bias)
    285.     elif isinstance(m, nn.LayerNorm):
    286.         nn.init.zeros_(m.bias)
    287.         nn.init.ones_(m.weight)
    288.  
    289.  
    290. def vit_base_patch16_224_in21k(num_classes: int = 21843, has_logits: bool = True):
    291.     """
    292.     ViT-Base model (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929).
    293.     ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
    294.     weights ported from official Google JAX impl:
    295.     https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_patch16_224_in21k-e5005f0a.pth
    296.     """
    297.     model = VisionTransformer(img_size=224,
    298.                               patch_size=16,
    299.                               embed_dim=768,
    300.                               depth=12,
    301.                               num_heads=12,
    302.                               representation_size=768 if has_logits else None,
    303.                               num_classes=num_classes)
    304.     return model
    305.  
    306.  
    307. def vit_base_patch32_224_in21k(num_classes: int = 21843, has_logits: bool = True):
    308.     """
    309.     ViT-Base model (ViT-B/32) from original paper (https://arxiv.org/abs/2010.11929).
    310.     ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
    311.     weights ported from official Google JAX impl:
    312.     https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_patch32_224_in21k-8db57226.pth
    313.     """
    314.     model = VisionTransformer(img_size=224,
    315.                               patch_size=32,
    316.                               embed_dim=768,
    317.                               depth=12,
    318.                               num_heads=12,
    319.                               representation_size=768 if has_logits else None,
    320.                               num_classes=num_classes)
    321.     return model
    322.  
    323.  
    324. def vit_large_patch16_224_in21k(num_classes: int = 21843, has_logits: bool = True):
    325.     """
    326.     ViT-Large model (ViT-L/16) from original paper (https://arxiv.org/abs/2010.11929).
    327.     ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
    328.     weights ported from official Google JAX impl:
    329.     https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch16_224_in21k-606da67d.pth
    330.     """
    331.     model = VisionTransformer(img_size=224,
    332.                               patch_size=16,
    333.                               embed_dim=1024,
    334.                               depth=24,
    335.                               num_heads=16,
    336.                               representation_size=1024 if has_logits else None,
    337.                               num_classes=num_classes)
    338.     return model
    339.  
    340.  
    341. def vit_large_patch32_224_in21k(num_classes: int = 21843, has_logits: bool = True):
    342.     """
    343.     ViT-Large model (ViT-L/32) from original paper (https://arxiv.org/abs/2010.11929).
    344.     ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
    345.     weights ported from official Google JAX impl:
    346.     https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch32_224_in21k-9046d2e7.pth
    347.     """
    348.     model = VisionTransformer(img_size=224,
    349.                               patch_size=32,
    350.                               embed_dim=1024,
    351.                               depth=24,
    352.                               num_heads=16,
    353.                               representation_size=1024 if has_logits else None,
    354.                               num_classes=num_classes)
    355.     return model
    356.  
    357.  
    358. def vit_huge_patch14_224_in21k(num_classes: int = 21843, has_logits: bool = True):
    359.     """
    360.     ViT-Huge model (ViT-H/14) from original paper (https://arxiv.org/abs/2010.11929).
    361.     ImageNet-21k weights @ 224x224, source https://github.com/google-research/vision_transformer.
    362.     NOTE: converted weights not currently available, too large for github release hosting.
    363.     """
    364.     model = VisionTransformer(img_size=224,
    365.                               patch_size=14,
    366.                               embed_dim=1280,
    367.                               depth=32,
    368.                               num_heads=16,
    369.                               representation_size=1280 if has_logits else None,
    370.                               num_classes=num_classes)
    371.     return model
    372.

     参考博文：深度学习之图像分类（十八）-- Vision Transformer(ViT)网络详解_vit网络_木卯_THU的博客-CSDN博客

  • 相关阅读:
    WuThreat身份安全云-TVD每日漏洞情报-2023-09-25
    WebDriver的工作原理及常用方法大全
    数据结构day7栈-顺序栈的实现
    Unity UI 框架相关的一些思考
    spark小笔记
    Cilium系列-9-主机路由切换为基于 BPF 的模式
    [附源码]计算机毕业设计springboot社区生活废品回收APP
    【Python】创建一个简易服务器并实现移动端快速访问电脑文件
    设计模式-单例模式
    电荷泵CP原理及在PLL/DLL中的使用
  • 原文地址：https://blog.csdn.net/weixin_48637535/article/details/132731583