ViT pytorch源码笔记

链接

源码地址
本文只列出了一些比较重要的部分。

patch embedding

先将大小为224$\times$ 224$\times$ 3的图像分割成16$\times$ 16$\times$ 3的patches，再展开做线性映射将每个patches的维度变为768。

""" Image to Patch Embedding using Conv2d

A convolution based approach to patchifying a 2D image w/ embedding projection.

Based on the impl in https://github.com/google-research/vision_transformer

Hacked together by / Copyright 2020 Ross Wightman
"""
from torch import nn as nn

from .helpers import to_2tuple
from .trace_utils import _assert


class PatchEmbed(nn.Module):
    """ 2D Image to Patch Embedding
    """
    def __init__(
            self,
            img_size=224,
            patch_size=16,
            in_chans=3,
            embed_dim=768,
            norm_layer=None,
            flatten=True,
            bias=True,
    ):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        # 原图像大小为(224, 224, 3)
        self.img_size = img_size
        # 每个patch的大小为(16, 16, 3)
        self.patch_size = patch_size
        # 分割后总共有 (224/16)*(224/16)=14*14=196个patches
        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
        self.num_patches = self.grid_size[0] * self.grid_size[1]
        # patches分割完要展开成一维的
        self.flatten = flatten

		# 分割成patches并做patch embedding的线性映射：用大小和步长都为patch_size(16*16)的卷积核来做
        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
    	# 原图形状：(B, 3, 224, 224)
        B, C, H, W = x.shape
        _assert(H == self.img_size[0], f"Input image height ({H}) doesn't match model ({self.img_size[0]}).")
        _assert(W == self.img_size[1], f"Input image width ({W}) doesn't match model ({self.img_size[1]}).")
        # 分割后线性映射的形状为：(B, 768, 14, 14)
        x = self.proj(x)
        # 展开后转置变成：(B, 14*14, 768)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
        x = self.norm(x)
        return x

注意力机制

class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0., proj_drop=0.):
        super().__init__()
        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
        self.num_heads = num_heads
        # dim是所有头合起来的维度即dim=768，分给每个头就是维度为768/12=64
        head_dim = dim // num_heads
        # scale是计算attention scores的根号d_k
        self.scale = 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 reshape后形状为：(B, 序列长度=N=14*14+1, 3, head数目=12, 768/12=64)
        # 序列长度有个 +1 是加上了 class token
        # permute后形状：(3, B, head数=12, 序列长度=N=14*14+1, 768/12=64)
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        # q,k,v形状分别为：(B, head数=12, N=14*14+1, 768/12=64)
        q, k, v = qkv.unbind(0)   # make torchscript happy (cannot use tensor as tuple)

		# 计算完attn形状：(B, head数=12, N=14*14+1, N=14*14+1)
        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

		# attn与V相乘转置后形状：(B, N=14*14+1, head=12, 768/12)
		# 把12个头的都拼在一起reshape后：(B, N=14*14+1, 768)
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        # 再做一个维度不变的线性映射W_o：(B,N=14*14+1,768)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

encoder的block

class Mlp(nn.Module):
    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
    """
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, bias=True, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        bias = to_2tuple(bias)
        drop_probs = to_2tuple(drop)

        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
        self.act = act_layer()
        self.drop1 = nn.Dropout(drop_probs[0])
        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
        self.drop2 = nn.Dropout(drop_probs[1])

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x


class Block(nn.Module):

    def __init__(
            self,
            dim,
            num_heads,
            mlp_ratio=4.,
            qkv_bias=False,
            drop=0.,
            attn_drop=0.,
            init_values=None,
            drop_path=0.,
            act_layer=nn.GELU,
            norm_layer=nn.LayerNorm
    ):
        super().__init__()
        # 先做一次layerNorm
        self.norm1 = norm_layer(dim)
        # 计算attention
        self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
        # 别的文章提出的改进方法：layerscale
        self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        # dropout
        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

		# 再做一次layerNorm
        self.norm2 = norm_layer(dim)
        # 做MLP，由两个全连接层和一个激活函数组成，先将维度768扩大4倍，再变回768
        self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
        # layerscale
        self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
        # dropout
        self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        x = x + self.drop_path1(self.ls1(self.attn(self.norm1(x))))
        x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
        return x

Transformer组装

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,
            global_pool='token',
            embed_dim=768,
            depth=12,
            num_heads=12,
            mlp_ratio=4.,
            qkv_bias=True,
            init_values=None,
            class_token=True,
            no_embed_class=False,
            pre_norm=False,
            fc_norm=None,
            drop_rate=0.,
            attn_drop_rate=0.,
            drop_path_rate=0.,
            weight_init='',
            embed_layer=PatchEmbed,
            norm_layer=None,
            act_layer=None,
            block_fn=Block,
    ):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            in_chans (int): number of input channels
            num_classes (int): number of classes for classification head
            global_pool (str): type of global pooling for final sequence (default: 'token')
            embed_dim (int): embedding dimension
            depth (int): depth of transformer
            num_heads (int): number of attention heads
            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
            qkv_bias (bool): enable bias for qkv if True
            init_values: (float): layer-scale init values
            class_token (bool): use class token
            fc_norm (Optional[bool]): pre-fc norm after pool, set if global_pool == 'avg' if None (default: None)
            drop_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            weight_init (str): weight init scheme
            embed_layer (nn.Module): patch embedding layer
            norm_layer: (nn.Module): normalization layer
            act_layer: (nn.Module): MLP activation layer
        """
        super().__init__()
        # 选择最后分类用cls token或者全局平均池化
        assert global_pool in ('', 'avg', 'token')
        assert class_token or global_pool != 'token'
        use_fc_norm = global_pool == 'avg' if fc_norm is None else fc_norm
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        act_layer = act_layer or nn.GELU

        self.num_classes = num_classes
        self.global_pool = global_pool
        # num_features, embed_dim都是768
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
        # 如果使用cls token进行分类，原始序列前面就要加上1个cls token
        self.num_prefix_tokens = 1 if class_token else 0
        self.no_embed_class = no_embed_class
        self.grad_checkpointing = False
		
		# 进行patch embdding的
        self.patch_embed = embed_layer(
            img_size=img_size,
            patch_size=patch_size,
            in_chans=in_chans,
            embed_dim=embed_dim,
            bias=not pre_norm,  # disable bias if pre-norm is used (e.g. CLIP)
        )
        num_patches = self.patch_embed.num_patches
		
		# cls token形状为(1,1,768)，为可学习的参数
        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if class_token else None
        # 序列长度等于patches数加上cls token数
        embed_len = num_patches if no_embed_class else num_patches + self.num_prefix_tokens
        # pos_embedding也使用可学习的参数
        self.pos_embed = nn.Parameter(torch.randn(1, embed_len, embed_dim) * .02)
        self.pos_drop = nn.Dropout(p=drop_rate)
        self.norm_pre = norm_layer(embed_dim) if pre_norm else nn.Identity()

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        # depth默认为12，每个encoder有12个blocks
        self.blocks = nn.Sequential(*[
            block_fn(
                dim=embed_dim,
                num_heads=num_heads,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                init_values=init_values,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[i],
                norm_layer=norm_layer,
                act_layer=act_layer
            )
            for i in range(depth)])
        self.norm = norm_layer(embed_dim) if not use_fc_norm else nn.Identity()

        # Classifier Head
        # 分类最后是对 cls token对应的输出 或者 global average pooling的输出 用一个线性映射完成的
        self.fc_norm = norm_layer(embed_dim) if use_fc_norm else nn.Identity()
        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()

        if weight_init != 'skip':
            self.init_weights(weight_init)

    def init_weights(self, mode=''):
        assert mode in ('jax', 'jax_nlhb', 'moco', '')
        head_bias = -math.log(self.num_classes) if 'nlhb' in mode else 0.
        trunc_normal_(self.pos_embed, std=.02)
        if self.cls_token is not None:
            nn.init.normal_(self.cls_token, std=1e-6)
        named_apply(get_init_weights_vit(mode, head_bias), self)

    def _init_weights(self, m):
        # this fn left here for compat with downstream users
        init_weights_vit_timm(m)

    @torch.jit.ignore()
    def load_pretrained(self, checkpoint_path, prefix=''):
        _load_weights(self, checkpoint_path, prefix)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token', 'dist_token'}

    @torch.jit.ignore
    def group_matcher(self, coarse=False):
        return dict(
            stem=r'^cls_token|pos_embed|patch_embed',  # stem and embed
            blocks=[(r'^blocks\.(\d+)', None), (r'^norm', (99999,))]
        )

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.grad_checkpointing = enable

    @torch.jit.ignore
    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes: int, global_pool=None):
        self.num_classes = num_classes
        if global_pool is not None:
            assert global_pool in ('', 'avg', 'token')
            self.global_pool = global_pool
        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()

    def _pos_embed(self, x):
    # 可以选择是否对cls token做pos_embedding
        if self.no_embed_class:
            # deit-3, updated JAX (big vision)
            # position embedding does not overlap with class token, add then concat
            x = x + self.pos_embed
            if self.cls_token is not None:
                x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
        else:
            # original timm, JAX, and deit vit impl
            # pos_embed has entry for class token, concat then add
            if self.cls_token is not None:
                x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
            x = x + self.pos_embed
        return self.pos_drop(x)

    def forward_features(self, x):
        x = self.patch_embed(x)
        x = self._pos_embed(x)
        x = self.norm_pre(x)
        if self.grad_checkpointing and not torch.jit.is_scripting():
            x = checkpoint_seq(self.blocks, x)
        else:
            x = self.blocks(x)
        x = self.norm(x)
        return x

    def forward_head(self, x, pre_logits: bool = False):
    	# 如果global_pool=True，则使用全局平均池化的输出进行分类，否则使用encoder的输出的第一个向量进行分类
        if self.global_pool:
            x = x[:, self.num_prefix_tokens:].mean(dim=1) if self.global_pool == 'avg' else x[:, 0]
        x = self.fc_norm(x)
        return x if pre_logits else self.head(x)

    def forward(self, x):
        x = self.forward_features(x)
        x = self.forward_head(x)
        return x

pos embedding的插值

def resize_pos_embed(posemb, posemb_new, num_prefix_tokens=1, gs_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是新的序列的长度
    ntok_new = posemb_new.shape[1]
    if num_prefix_tokens:
    	# 如果有加入cls token，把cls token与原序列分开处理
        posemb_prefix, posemb_grid = posemb[:, :num_prefix_tokens], posemb[0, num_prefix_tokens:]
        ntok_new -= num_prefix_tokens
    else:
        posemb_prefix, posemb_grid = posemb[:, :0], posemb[0]
    # 原序列长度为14*14，则原始patches组成的正方形边长gs_old为14
    gs_old = int(math.sqrt(len(posemb_grid)))
    # 新的序列长度对应的patches组成的正方形边长为 gs_new开根号
    if not len(gs_new):  # backwards compatibility
        gs_new = [int(math.sqrt(ntok_new))] * 2
    assert len(gs_new) >= 2
    _logger.info('Position embedding grid-size from %s to %s', [gs_old, gs_old], gs_new)
    # pos embedding变回二维形状：(1, 14, 14, 768)
    # permute后：(1, 768, 14, 14)
    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
    # 进行2D插值：(1, 768, gs_new, gs_new)
    posemb_grid = F.interpolate(posemb_grid, size=gs_new, mode='bicubic', align_corners=False)
    # 重新变回pos embedding应有的形状：(1, ntok_new/新序列长度，768)
    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -1)
    # 如果有cls token，将cls token和新序列拼接
    posemb = torch.cat([posemb_prefix, posemb_grid], dim=1)
    return posemb