【代码阅读】vision transformer
2021-10-25 本文已影响0人
Joyner2018
vit.py的模型文件
#encoding=utf-8
import torch
from torch import nn
from einops import rearrange, repeat
from einops.layers.torch import Rearrange
# helpers
'''
主要用于生成成对的输入
比如
卷积中的kennel size,当的size为x时，生成（x，y）
图像中resize中，输入只有一个参数x时，则生成（x,x）,即需要resize成(x,x)
输入两个参数(x,y),则不做任何修改，直接返回（x,y）
'''
def pair(t):
return t if isinstance(t, tuple) else (t, t)
# classes
'''
PreNorm表示
先归一化
再执行fn的操作
'''
class PreNorm(nn.Module):
def __init__(self, dim, fn):
super().__init__()
self.norm = nn.LayerNorm(dim)
self.fn = fn
def forward(self, x, **kwargs):
return self.fn(self.norm(x), **kwargs)
'''
前向连接
线性层
激活层
随机丢弃层
线性层
随机丢弃层
'''
class FeedForward(nn.Module):
def __init__(self, dim, hidden_dim, dropout = 0.):
super().__init__()
self.net = nn.Sequential(
nn.Linear(dim, hidden_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(hidden_dim, dim),
nn.Dropout(dropout)
)
def forward(self, x):
return self.net(x)
'''
注意力层
输入的x，
线性层
分块得到q,k,v
根据head的数量把q,k,v分成head个分支
rearrange(t, 'b n (h d) -> b h n d', h = self.heads)
'''
class Attention(nn.Module):
def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
super().__init__()
inner_dim = dim_head * heads
project_out = not (heads == 1 and dim_head == dim)
self.heads = heads
self.scale = dim_head ** -0.5
self.attend = nn.Softmax(dim = -1)
self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, dim),
nn.Dropout(dropout)
) if project_out else nn.Identity()
def forward(self, x):
qkv = self.to_qkv(x).chunk(3, dim = -1)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
attn = self.attend(dots)
out = torch.matmul(attn, v)
out = rearrange(out, 'b h n d -> b n (h d)')
return self.to_out(out)
class Transformer(nn.Module):
def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
super().__init__()
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(nn.ModuleList([
PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
]))
def forward(self, x):
for attn, ff in self.layers:
x = attn(x) + x
x = ff(x) + x
return x
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)
patch_height, patch_width = pair(patch_size)
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)
patch_dim = channels * patch_height * patch_width
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),
nn.Linear(patch_dim, dim),
)
self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
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)
)
def forward(self, img):
print(img.shape)
x = self.to_patch_embedding(img)
print(x.shape)
print(x.type())
b, n, _ = x.shape
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
x = torch.cat((cls_tokens, x), dim=1)
print("self.pos_embedding.size()",self.pos_embedding.size())
print("n+1",n+1)
x += self.pos_embedding[:, :(n + 1)]
x = self.dropout(x)
x = self.transformer(x)
x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]
x = self.to_latent(x)
return self.mlp_head(x)
if __name__=="__main__":
from torchsummary import summary
vit= ViT(
image_size = 256,
patch_size = 32,
num_classes = 1000,
dim = 1024,
depth = 6,
heads = 16,
mlp_dim = 2048,
dropout = 0.1,
emb_dropout = 0.1
)
img = torch.randn(1, 3, 256, 256)
vit=vit.cuda()
img=img.cuda()
preds = vit(img)
print(preds.size())
summary(vit,(3,256,256))
知识点
- 代码须在pytorch1.7以上版本运行，因为nn.GELU()在低版本得pytorch不支持
- 成对表达函数；简单有效pair(t)
- 函数也可以作为输入参数
- einops 是一个很棒的库函数，通过灵活而强大的张量操作符为你提供易读并可靠的代码。 支持 numpy、pytorch、tensorflow 等等。einops正在缓慢而有力地渗入我代码的每一个角落和缝隙。如果你发现自己困扰于一堆高维的张量，这可能会改变你的生活。
- 如何使用einops，参考https://zhuanlan.zhihu.com/p/342675997
- einops在vit的使用
rearrange(t, 'b n (h d) -> b h n d', h = self.heads)
rearrange(out, 'b h n d -> b n (h d)')
Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width) 支持梯度反传
repeat(self.cls_token, '() n d -> b n d', b = b)
运行结果
torch.Size([1, 3, 256, 256])
torch.Size([1, 64, 1024])
torch.cuda.FloatTensor
self.pos_embedding.size() torch.Size([1, 65, 1024])
n+1 65
torch.Size([1, 1000])
torch.Size([2, 3, 256, 256])
torch.Size([2, 64, 1024])
torch.cuda.FloatTensor
self.pos_embedding.size() torch.Size([1, 65, 1024])
n+1 65
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Rearrange-1 [-1, 64, 3072] 0
Linear-2 [-1, 64, 1024] 3,146,752
Dropout-3 [-1, 65, 1024] 0
LayerNorm-4 [-1, 65, 1024] 2,048
Linear-5 [-1, 65, 3072] 3,145,728
Softmax-6 [-1, 16, 65, 65] 0
Linear-7 [-1, 65, 1024] 1,049,600
Dropout-8 [-1, 65, 1024] 0
Attention-9 [-1, 65, 1024] 0
PreNorm-10 [-1, 65, 1024] 0
LayerNorm-11 [-1, 65, 1024] 2,048
Linear-12 [-1, 65, 2048] 2,099,200
GELU-13 [-1, 65, 2048] 0
Dropout-14 [-1, 65, 2048] 0
Linear-15 [-1, 65, 1024] 2,098,176
Dropout-16 [-1, 65, 1024] 0
FeedForward-17 [-1, 65, 1024] 0
PreNorm-18 [-1, 65, 1024] 0
LayerNorm-19 [-1, 65, 1024] 2,048
Linear-20 [-1, 65, 3072] 3,145,728
Softmax-21 [-1, 16, 65, 65] 0
Linear-22 [-1, 65, 1024] 1,049,600
Dropout-23 [-1, 65, 1024] 0
Attention-24 [-1, 65, 1024] 0
PreNorm-25 [-1, 65, 1024] 0
LayerNorm-26 [-1, 65, 1024] 2,048
Linear-27 [-1, 65, 2048] 2,099,200
GELU-28 [-1, 65, 2048] 0
Dropout-29 [-1, 65, 2048] 0
Linear-30 [-1, 65, 1024] 2,098,176
Dropout-31 [-1, 65, 1024] 0
FeedForward-32 [-1, 65, 1024] 0
PreNorm-33 [-1, 65, 1024] 0
LayerNorm-34 [-1, 65, 1024] 2,048
Linear-35 [-1, 65, 3072] 3,145,728
Softmax-36 [-1, 16, 65, 65] 0
Linear-37 [-1, 65, 1024] 1,049,600
Dropout-38 [-1, 65, 1024] 0
Attention-39 [-1, 65, 1024] 0
PreNorm-40 [-1, 65, 1024] 0
LayerNorm-41 [-1, 65, 1024] 2,048
Linear-42 [-1, 65, 2048] 2,099,200
GELU-43 [-1, 65, 2048] 0
Dropout-44 [-1, 65, 2048] 0
Linear-45 [-1, 65, 1024] 2,098,176
Dropout-46 [-1, 65, 1024] 0
FeedForward-47 [-1, 65, 1024] 0
PreNorm-48 [-1, 65, 1024] 0
LayerNorm-49 [-1, 65, 1024] 2,048
Linear-50 [-1, 65, 3072] 3,145,728
Softmax-51 [-1, 16, 65, 65] 0
Linear-52 [-1, 65, 1024] 1,049,600
Dropout-53 [-1, 65, 1024] 0
Attention-54 [-1, 65, 1024] 0
PreNorm-55 [-1, 65, 1024] 0
LayerNorm-56 [-1, 65, 1024] 2,048
Linear-57 [-1, 65, 2048] 2,099,200
GELU-58 [-1, 65, 2048] 0
Dropout-59 [-1, 65, 2048] 0
Linear-60 [-1, 65, 1024] 2,098,176
Dropout-61 [-1, 65, 1024] 0
FeedForward-62 [-1, 65, 1024] 0
PreNorm-63 [-1, 65, 1024] 0
LayerNorm-64 [-1, 65, 1024] 2,048
Linear-65 [-1, 65, 3072] 3,145,728
Softmax-66 [-1, 16, 65, 65] 0
Linear-67 [-1, 65, 1024] 1,049,600
Dropout-68 [-1, 65, 1024] 0
Attention-69 [-1, 65, 1024] 0
PreNorm-70 [-1, 65, 1024] 0
LayerNorm-71 [-1, 65, 1024] 2,048
Linear-72 [-1, 65, 2048] 2,099,200
GELU-73 [-1, 65, 2048] 0
Dropout-74 [-1, 65, 2048] 0
Linear-75 [-1, 65, 1024] 2,098,176
Dropout-76 [-1, 65, 1024] 0
FeedForward-77 [-1, 65, 1024] 0
PreNorm-78 [-1, 65, 1024] 0
LayerNorm-79 [-1, 65, 1024] 2,048
Linear-80 [-1, 65, 3072] 3,145,728
Softmax-81 [-1, 16, 65, 65] 0
Linear-82 [-1, 65, 1024] 1,049,600
Dropout-83 [-1, 65, 1024] 0
Attention-84 [-1, 65, 1024] 0
PreNorm-85 [-1, 65, 1024] 0
LayerNorm-86 [-1, 65, 1024] 2,048
Linear-87 [-1, 65, 2048] 2,099,200
GELU-88 [-1, 65, 2048] 0
Dropout-89 [-1, 65, 2048] 0
Linear-90 [-1, 65, 1024] 2,098,176
Dropout-91 [-1, 65, 1024] 0
FeedForward-92 [-1, 65, 1024] 0
PreNorm-93 [-1, 65, 1024] 0
Transformer-94 [-1, 65, 1024] 0
Identity-95 [-1, 1024] 0
LayerNorm-96 [-1, 1024] 2,048
Linear-97 [-1, 1000] 1,025,000
================================================================
Total params: 54,554,600
Trainable params: 54,554,600
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.75
Forward/backward pass size (MB): 64.02
Params size (MB): 208.11
Estimated Total Size (MB): 272.88
----------------------------------------------------------------
