https://github.com/JunkyByte/easy_ViTPose

backbone→ ViT

head(TopdownHeatmapSimpleHead)→ Decoder

model.py

class ViTPose(nn.Module):
    def __init__(self, cfg: dict) -> None:
        super(ViTPose, self).__init__()
        
        backbone_cfg = {k: v for k, v in cfg['backbone'].items() if k != 'type'}
        head_cfg = {k: v for k, v in cfg['keypoint_head'].items() if k != 'type'}
        
        self.backbone = ViT(**backbone_cfg)
        self.keypoint_head = TopdownHeatmapSimpleHead(**head_cfg)
    
    def forward_features(self, x):
        return self.backbone(x)
    
    def forward(self, x):
        return self.keypoint_head(self.backbone(x))

1. Patch embedding layer

vit.py

class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, ratio=1):
        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]) * (ratio ** 2)
        self.patch_shape = (int(img_size[0] // patch_size[0] * ratio), int(img_size[1] // patch_size[1] * ratio))
        self.origin_patch_shape = (int(img_size[0] // patch_size[0]), int(img_size[1] // patch_size[1]))
        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[0] // ratio), padding=4 + 2 * (ratio//2-1))

    def forward(self, x, **kwargs):
        B, C, H, W = x.shape
				#downsampling
        x = self.proj(x)
        Hp, Wp = x.shape[2], x.shape[3]
				#flatten, linear projection
        x = x.flatten(2).transpose(1, 2)

2. Transformer Encoder

Multiple-Head Self Attention

FFN(feed-forward network)

vit.py

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, attn_head_dim=None
                 ):
        super().__init__()
        
				# norm layer
        self.norm1 = norm_layer(dim)
				# attention
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
            attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim
            )

        # 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)
				# mlp
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

		# FFN
    def forward(self, x):
				# skip connection
				x = x + self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x

class Attention(nn.Module):
    def __init__(
            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
            proj_drop=0., attn_head_dim=None,):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.dim = dim

        self.scale = qk_scale or head_dim ** -0.5

				# query key value
        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=qkv_bias)

        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(all_head_dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x):
        B, N, C = x.shape
        qkv = self.qkv(x)
        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
				
				# scaling
        q = q * self.scale
				# query, key 행렬 내적하여 attention score
        attn = (q @ k.transpose(-2, -1))
				
				# softmax
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

				# attention score를 value 행렬과 내적
        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
        x = self.proj(x)
        x = self.proj_drop(x)

        return x

Decoder

topdown_heatmap_base_head.py