【深度学习】Vit视觉Transformer，数据经过网络的维度变化

源码网址 ：
github

网络模型定义：

model = ViT(					# 对此处设置断点1 进行调试观察
    image_size = 256,
    patch_size = 32,
    num_classes = 20,
    dim = 1024,
    depth = 6,
    heads = 16,
    mlp_dim = 2048,
    dropout = 0.1,
    emb_dropout = 0.1
)
img = torch.randn(5, 3, 256, 256)		# 相当于batchsize为5 即输入5张256*256*3的图片
preds =model(img)				# 对此处设置断点2 进行调试观察

分阶段调试查看数据维度变化

断点1调试

数据大小或维度以注释给出

class ViT(nn.Module):
    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.):
        super().__init__()
        image_height, image_width = pair(image_size)		# 256 256
        patch_height, patch_width = pair(patch_size)		# 32 32

        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'

        num_patches = (image_height // patch_height) * (image_width // patch_width)	# 64
        patch_dim = channels * patch_height * patch_width		# 3072
        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'

        self.to_patch_embedding = nn.Sequential(
            Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width),		# 5*3*256*256 to 5*64*3072
            nn.Linear(patch_dim, dim),	# 5*64*3072 to 5*64*1024
        )

        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))	# 1 64+1 1024
        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))		# 1 1 1024
        self.dropout = nn.Dropout(emb_dropout)

        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)

        self.pool = pool
        self.to_latent = nn.Identity()

        self.mlp_head = nn.Sequential(
            nn.LayerNorm(dim),
            nn.Linear(dim, num_classes)
        )

断点2调试

重点

输入为：img = torch.randn(5, 3, 256, 256)
==ViT的forward函数==
def forward(self, img):		# 5 3 256 256
    x = self.to_patch_embedding(img)	# 5 64 1024
    b, n, _ = x.shape	# 5 64 1024

    cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)	# 5 1 1024
    x = torch.cat((cls_tokens, x), dim=1)		# 5 65 1024
   												#self.pos_embedding 1 65 1024 
    x += self.pos_embedding[:, :(n + 1)]		# 5 65 1024
    x = self.dropout(x)

    x = self.transformer(x)		# 看transformer的forward函数
    ==transformer的forward==
	def forward(self, x):		# 5 65 1024
	    for attn, ff in self.layers:
	        x = attn(x) + x		# 5 65 1024	Attention网络
        	#  ==attention的forward==
            def forward(self, x):	# # 5 65 1024	
            												# self.to_qkv(x) 5 65 3072
		        qkv = self.to_qkv(x).chunk(3, dim = -1)		# 含3个张量的tuple 每个张量为5 65 1024
		       												# self.heads为16 
		        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)	#将每个张量由 5*65*1024 变为 5*16*65*64
															# self.scale 0.125   k.transpose(-1, -2) 5 16 64 65
		        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale	# 5 16 65 65
		        attn = self.attend(dots)		# 5 16 65 65
		        out = torch.matmul(attn, v)		# 5 16 65 64
		        out = rearrange(out, 'b h n d -> b n (h d)')	#  5 65 1024
		        return self.to_out(out)	# 5 65 1024
	        
	        x = ff(x) + x		# 5 65 1024 全连接网络
	    return x		# 5 65 1024

    x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]	#  5 1024
    x = self.to_latent(x)		# 5 1024
    return self.mlp_head(x)		# 5 20 类别数