Vit 实战营 Class2：图像与Transformer基础

数组图像：图像与像素

什么是数字图像？在计算机图像的图像格式。每一个点叫pixel。type是uint8。

左边比右边清楚。

105可以用8位的二进制表示。24位的话，可以表示1600万个不同的颜色。HDR和杜比视界，用10位二进制表示一个颜色，所以是30位，1 billion个不同的颜色。


注意flatten是维度的展平，而Expand是一个维度的复制。

图像分类：机器如何学习？

NMT：Neuron Machine Translation

sequetial to sequential的。一个句子就是一个sequetial，一个序列，S2S，又是一个Encoder -> Decoder的过程。

Transformer

Transformer本质是一个Encoder+Decoder的过程。

Encoder和Decoder都不是一层，可能是多层。

FFN：Feed Forward Network
MSA：Multi head self-Attention

Vision Transformer

Vit 是没有Decoder的。Encoder就是实现每一层，然后resnet去跳连。

映射就是linear层，或者说可学习的层。
把图像分成块，就相当于NLP领域把一个句子分成分词。
Patch Embedding过程：一张[h, w, 3] img 按照window的大小，分成不同的patch，他们的维度变成[num_patch, Wh, Ww, 3]. 每一个patch我们讲他flatten成1维，然后过一个linear层，最后输出的就是一个token，所有token就是patch embedding。每一个token拉平就是embed_dim。

关于Linear Prejection是怎么操作的，可以看下图。

从卷积CNN的角度理解patch embedding。卷积里的参数值，其实就是Linear Project中的可学习参数w。embed_di的长度，其实就是卷积核的个数，比如我们有32个filter，那么embed_dim=32。

代码实战

import paddle
t = paddle.zeros([3, 3])
print(t)

t = paddle.randn([5, 3])
print(t)

from PIL import Image
import numpy as np
img = np.array(Image.open("img.png"))
print(img.shape)
# print(img)
t = paddle.to_tensor(img, dtype='float32')
print(type(t))
print(t.dtype)
t = t.transpose([2, 1, 0])
print(t.shape)

# chunk
t = paddle.randint(0, 10, [5, 15])
print(t)
qkv = t.chunk(3, -1)
print(qkv)
q, k, v = qkv
print(q)

# ViT Online Class
# Author: Dr. Zhu
# Project: PaddleViT (https://github.com/BR-IDL/PaddleViT)
# 2021.11
import paddle
import paddle.nn as nn
import numpy as np
from PIL import Image

# paddle.set_device('cpu')

class Identity(nn.Layer):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        return x


class Mlp(nn.Layer):
    def __init__(self, embed_dim, mlp_ratio=4.0, dropout=0.):
        super().__init__()
        self.fc1 = nn.Linear(embed_dim, int(embed_dim*mlp_ratio))
        self.fc2 = nn.Linear(int(embed_dim*mlp_ratio), embed_dim)
        self.act = nn.GELU()
        self.dropout = nn.Dropout(dropout)

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



class PatchEmbedding(nn.Layer):
    def __init__(self, image_size, patch_size, in_channels, embed_dim, dropout=0.):
        super().__init__()
        self.patch_embedding = nn.Conv2D(in_channels,
                                        embed_dim,
                                        kernel_size=patch_size,
                                        bias_attr=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        # [n, c, h, w] [1, 1, 28, 28]
        x = self.patch_embedding(x) # [n, c', h', w']
        x = x.flatten(2) # [n, c', h'*w']
        x = x.transpose([0, 2, 1]) # [n, h'*w', c']  h'*w'=num_patches
        x = self.dropout(x)
        return x


class Attention(nn.Layer):
    def __init__(self):
        super().__init__()

    def forward(self, x):
        return x


class EncoderLayer(nn.Layer):
    def __init__(self, embed_dim):
        super().__init__()
        self.attn_norm = nn.LayerNorm(embed_dim)
        self.attn = Attention()
        self.mlp_norm = nn.LayerNorm(embed_dim)
        self.mlp = Mlp(embed_dim)

    def forward(self, x):
        h = x 
        x = self.attn_norm(x)
        x = self.attn(x)
        x = x + h

        h = x
        x = self.mlp_norm(x)
        x = self.mlp(x)
        x = x + h
        return x


class ViT(nn.Layer):
    def __init__(self):
        super().__init__()
        self.patch_embed = PatchEmbedding(224, 7, 3, 16)
        layer_list = [EncoderLayer(16) for i in range(5)]
        self.encoders = nn.LayerList(layer_list)
        self.head = nn.Linear(16, 10)
        self.avgpool = nn.AdaptiveAvgPool1D(1)
        self.norm = nn.LayerNorm(16)

    def forward(self, x):
        x = self.patch_embed(x) # [n, h*w, c]: 4, 1024, 16
        for encoder in self.encoders:
            x = encoder(x)
        # avg
        x = self.norm(x)
        x = x.transpose([0, 2, 1])
        x = self.avgpool(x)
        x = x.flatten(1)
        x = self.head(x)
        
        return x


def main():
    t = paddle.randn([4, 3, 224, 224])
    model = ViT()
    out = model(t)
    print(out.shape)


if __name__ == "__main__":
    main()