PyTorch笔记 - Vision Transformer

Transformer包含Encoder和Decoder，核心是Multi-Head Self-Attention(空间融合)，FeedForward Nerual Network(通道融合)。

Encoder和Decoder的交互信息：Memory-base Multi-Head Cross-Attention

注入位置信息Position Embedding

数据量的要求与 归纳偏置(Inductive Bias) 的引入成反比，上限很高，数据量要求也很高。

归纳法、演绎法，归纳偏置(Inductive Bias)，将人类的经验带入模型的设计当中。

Transformer的使用场景：

• Encoder Only：BERT、分类任务、非流式任务
• Decoder Only：GPT系列、语言建模、自回归生成任务、流式任务
• Encoder-Decoder：机器翻译、语言识别

Vision Transformer(ViT)：

• DNN perspective(视角): Image2Patch、Patch2Embedding
• CNN perspective(视角): 2D Convolution over image
• Class Token Embedding，占位符
• Position Embedding: Interpolation(插入) when inference
• Transformer Encoder
• Classification Head

Paper: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Classification Token：起到Query的作用

Linear Projection of Flattened Patches -> Patch + Position Embedding -> Transformer Encoder -> MLP Head

Patch + Position Embedding，先从左到右，再从上到下，拉成序列形状

实现Image2Embedding，TransformerEncoder由PyTorch封装

• torch.nn.functional.unfold: 取出patch卷积区域
• torch.nn.TransformerEncoder

ViT：

import torch
import torch.nn as nn
import torch.nn.functional as F


# step1 convert image to embedding vector sequence
def image2emb_naive(image, patch_size, weight):
    """
    使用unfold生成patch
    """
    # image shape: bs*channel*h*w
    # 没有交叠，stride=patch_size，直接生成patch
    patch = F.unfold(image, kernel_size=patch_size, stride=patch_size)
    patch = patch.transpose(2, 1)
    # (bs, patch_depth(patch_size*patch_size*ic), num_patch)
    print(f'patch: {patch.shape}')
    patch_embedding = patch @ weight  # 输出的embeding
    print(f'patch_embedding: {patch_embedding.shape}')
    return patch_embedding


def image2emb_conv(image, kernel, stride):
    """
    使用conv生成patch
    """
    conv_output = F.conv2d(image, kernel, stride=stride)  # bs*oc*oh*ow
    bs, oc, oh, ow = conv_output.shape
    patch_embedding = conv_output.reshape((bs, oc, oh*ow)).transpose(2, 1)
    print(f'patch_embedding: {patch_embedding.shape}')
    return patch_embedding


# test code for image2emb
bs, ic, image_h, image_w = 1, 3, 8, 8
patch_size = 4
model_dim = 8  # embedding dim
max_num_token = 16 
num_classes = 10
label = torch.randint(10, (bs,))
patch_depth = patch_size*patch_size*ic

# 分块方法得到embedding
torch.manual_seed(42)
image = torch.randn((bs, ic, image_h, image_w))  # 生成图像
weight = torch.randn((patch_depth, model_dim))  # patch_depth -> model_dim, model_dim是输出通道数目
print(f'weight: {weight.shape}')
patch_embedding_naive = image2emb_naive(image, patch_size, weight)
print(f'patch_embedding_naive: \n{patch_embedding_naive}')

# 二维卷积方法得到embedding
# kernel的形状，oc*ic*k_h*k_w
kernel = weight.transpose(1, 0).reshape((model_dim, ic, patch_size, patch_size))
patch_embedding_conv = image2emb_conv(image, kernel, stride=patch_size)
print(f'patch_embedding_conv: \n{patch_embedding_conv}')


# step2 prepend CLS token embedding
cls_token_embedding = torch.randn((bs, 1, model_dim), requires_grad=True)
token_embedding = torch.cat([cls_token_embedding, patch_embedding_conv], dim=1)
print(f'token_embedding: {token_embedding.shape}')


# step3 add position embedding
position_embedding_table = torch.randn((max_num_token, model_dim), requires_grad=True)
seq_len = token_embedding.shape[1]
# 复制 position_embedding 操作
position_embedding = torch.tile(position_embedding_table[:seq_len], [token_embedding.shape[0], 1, 1])   
token_embedding += position_embedding
print(f'token_embedding: {token_embedding.shape}')


# step4 pass embedding to Transformer Encoder
encoder_layer = nn.TransformerEncoderLayer(d_model=model_dim, nhead=8)
transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=6)
encoder_output = transformer_encoder(token_embedding)


# step5 do classification
cls_token_output = encoder_output[:, 0, :]
linear_layer = nn.Linear(model_dim, num_classes)
logits = linear_layer(cls_token_output)
loss_fn = nn.CrossEntropyLoss()
loss = loss_fn(logits, label)
print(f'loss: {loss}')
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84