什么是Vision Transformer(VIT)
视觉Transformer最近非常的火热,从VIT开始,我先学学看。
Vision Transformer是Transformer的视觉版本,Transformer基本上已经成为了自然语言处理的标配,但是在视觉中的运用还受到限制。
Vision Transformer打破了这种NLP与CV的隔离,将Transformer应用于图像图块(patch)序列上,进一步完成图像分类任务。简单来理解,Vision Transformer就是将输入进来的图片,每隔一定的区域大小划分图片块。然后将划分后的图片块组合成序列,将组合后的结果传入Transformer特有的Multi-head Self-attention进行特征提取。最后利用Cls Token进行分类。
代码下载
Vision Transforme的实现思路
一、整体结构解析
与寻常的分类网络类似,整个Vision Transformer可以氛围两部分,一部分是特征提取部分,另一部分是分类部分。
- 在特征提取部分,VIT所做的工作是特征提取。特征提取部分在图片中的对应区域是Patch+Position Embedding和Transformer Encoder。
- Patch+Position Embedding的作用主要是对输入进来的图片进行分块处理,每隔一定的区域大小划分图片块。然后将划分后的图片块组合成序列。
- 在获得序列信息后,传入Transformer Encoder进行特征提取,这是Transformer特有的Multi-head Self-attention结构,通过自注意力机制,关注每个图片块的重要程度。
- 在分类部分,VIT所做的工作是利用提取到的特征进行分类。在进行特征提取的时候,我们会在图片序列中添加上Cls Token,该Token会作为一个单位的序列信息一起进行特征提取,提取的过程中,该Cls Token会与其它的特征进行特征交互,融合其它图片序列的特征。
- 最终,我们利用Multi-head Self-attention结构提取特征后的Cls Token进行全连接分类。
二、网络结构解析
1、特征提取部分介绍
a、Patch+Position Embedding
Patch+Position Embedding的作用主要是对输入进来的图片进行分块处理,每隔一定的区域大小划分图片块。然后将划分后的图片块组合成序列。
该部分首先对输入进来的图片进行分块处理,处理方式其实很简单,使用的是现成的卷积。由于卷积使用的是滑动窗口的思想,我们只需要设定特定的步长,就可以输入进来的图片进行分块处理了。
在VIT中,我们常设置这个卷积的卷积核大小为16x16,步长也为16x16,此时卷积就会每隔16个像素点进行一次特征提取,由于卷积核大小为16x16,两个图片区域的特征提取过程就不会有重叠。当我们输入的图片是224, 224, 3的时候,我们可以获得一个14, 14, 768的特征层。
下一步就是将这个特征层组合成序列,组合的方式非常简单,就是将高宽维度进行平铺,14, 14, 768在高宽维度平铺后,获得一个196, 768的特征层。
平铺完成后,我们会在图片序列中添加上Cls Token,该Token会作为一个单位的序列信息一起进行特征提取,图中的这个0*就是Cls Token,我们此时获得一个197, 768的特征层。
添加完成Cls Token后,再为所有特征添加上位置信息,这样网络才有区分不同区域的能力。添加方式其实也非常简单,我们生成一个197, 768的参数矩阵,这个参数矩阵是可训练的,把这个矩阵加上197, 768的特征层即可。
到这里,Patch+Position Embedding就构建完成了,构建代码如下:
#--------------------------------------------------------------------------------------------------------------------# # classtoken部分是transformer的分类特征。用于堆叠到序列化后的图片特征中,作为一个单位的序列特征进行特征提取。 # # 在利用步长为16x16的卷积将输入图片划分成14x14的部分后,将14x14部分的特征平铺,一幅图片会存在序列长度为196的特征。 # 此时生成一个classtoken,将classtoken堆叠到序列长度为196的特征上,获得一个序列长度为197的特征。 # 在特征提取的过程中,classtoken会与图片特征进行特征的交互。最终分类时,我们取出classtoken的特征,利用全连接分类。 #--------------------------------------------------------------------------------------------------------------------# class ClassToken(Layer): def __init__(self, cls_initializer='zeros', cls_regularizer=None, cls_constraint=None, **kwargs): super(ClassToken, self).__init__(**kwargs) self.cls_initializer = keras.initializers.get(cls_initializer) self.cls_regularizer = keras.regularizers.get(cls_regularizer) self.cls_constraint = keras.constraints.get(cls_constraint) def get_config(self): config = { 'cls_initializer': keras.initializers.serialize(self.cls_initializer), 'cls_regularizer': keras.regularizers.serialize(self.cls_regularizer), 'cls_constraint': keras.constraints.serialize(self.cls_constraint), } base_config = super(ClassToken, self).get_config() return dict(list(base_config.items()) + list(config.items())) def compute_output_shape(self, input_shape): return (input_shape[0], input_shape[1] + 1, input_shape[2]) def build(self, input_shape): self.num_features = input_shape[-1] self.cls = self.add_weight( shape = (1, 1, self.num_features), initializer = self.cls_initializer, regularizer = self.cls_regularizer, constraint = self.cls_constraint, name = 'cls', ) super(ClassToken, self).build(input_shape) def call(self, inputs): batch_size = tf.shape(inputs)[0] cls_broadcasted = tf.cast(tf.broadcast_to(self.cls, [batch_size, 1, self.num_features]), dtype = inputs.dtype) return tf.concat([cls_broadcasted, inputs], 1) #--------------------------------------------------------------------------------------------------------------------# # 为网络提取到的特征添加上位置信息。 # 以输入图片为224, 224, 3为例,我们获得的序列化后的图片特征为196, 768。加上classtoken后就是197, 768 # 此时生成的pos_Embedding的shape也为197, 768,代表每一个特征的位置信息。 #--------------------------------------------------------------------------------------------------------------------# class AddPositionEmbs(Layer): def __init__(self, image_shape, patch_size, pe_initializer='zeros', pe_regularizer=None, pe_constraint=None, **kwargs): super(AddPositionEmbs, self).__init__(**kwargs) self.image_shape = image_shape self.patch_size = patch_size self.pe_initializer = keras.initializers.get(pe_initializer) self.pe_regularizer = keras.regularizers.get(pe_regularizer) self.pe_constraint = keras.constraints.get(pe_constraint) def get_config(self): config = { 'pe_initializer': keras.initializers.serialize(self.pe_initializer), 'pe_regularizer': keras.regularizers.serialize(self.pe_regularizer), 'pe_constraint': keras.constraints.serialize(self.pe_constraint), } base_config = super(AddPositionEmbs, self).get_config() return dict(list(base_config.items()) + list(config.items())) def compute_output_shape(self, input_shape): return input_shape def build(self, input_shape): assert (len(input_shape) == 3), f"Number of dimensions should be 3, got {len(input_shape)}" length = (224 // self.patch_size) * (224 // self.patch_size) + 1 self.pe = self.add_weight( # shape = [1, input_shape[1], input_shape[2]], shape = [1, length, input_shape[2]], initializer = self.pe_initializer, regularizer = self.pe_regularizer, constraint = self.pe_constraint, name = 'pos_embedding', ) super(AddPositionEmbs, self).build(input_shape) def call(self, inputs): num_features = tf.shape(inputs)[2] cls_token_pe = self.pe[:, 0:1, :] img_token_pe = self.pe[:, 1: , :] img_token_pe = tf.reshape(img_token_pe, [1, (224 // self.patch_size), (224 // self.patch_size), num_features]) img_token_pe = tf.image.resize_bicubic(img_token_pe, (self.image_shape[0] // self.patch_size, self.image_shape[1] // self.patch_size), align_corners=False) img_token_pe = tf.reshape(img_token_pe, [1, -1, num_features]) pe = tf.concat([cls_token_pe, img_token_pe], axis = 1) return inputs + tf.cast(pe, dtype=inputs.dtype) def VisionTransformer(input_shape = [224, 224], patch_size = 16, num_layers = 12, num_features = 768, num_heads = 12, mlp_dim = 3072, classes = 1000, dropout = 0.1): #-----------------------------------------------# # 224, 224, 3 #-----------------------------------------------# inputs = Input(shape = (input_shape[0], input_shape[1], 3)) #-----------------------------------------------# # 224, 224, 3 -> 14, 14, 768 #-----------------------------------------------# x = Conv2D(num_features, patch_size, strides = patch_size, padding = "valid", name = "patch_embed.proj")(inputs) #-----------------------------------------------# # 14, 14, 768 -> 196, 768 #-----------------------------------------------# x = Reshape(((input_shape[0] // patch_size) * (input_shape[1] // patch_size), num_features))(x) #-----------------------------------------------# # 196, 768 -> 197, 768 #-----------------------------------------------# x = ClassToken(name="cls_token")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 #-----------------------------------------------# x = AddPositionEmbs(input_shape, patch_size, name="pos_embed")(x)
b、Transformer Encoder
在上一步获得shape为197, 768的序列信息后,将序列信息传入Transformer Encoder进行特征提取,这是Transformer特有的Multi-head Self-attention结构,通过自注意力机制,关注每个图片块的重要程度。
I、Self-attention结构解析
看懂Self-attention结构,其实看懂下面这个动图就可以了,动图中存在一个序列的三个单位输入,每一个序列单位的输入都可以通过三个处理(比如全连接)获得Query、Key、Value,Query是查询向量、Key是键向量、Value值向量。
如果我们想要获得input-1的输出,那么我们进行如下几步:
1、利用input-1的查询向量,分别乘上input-1、input-2、input-3的键向量,此时我们获得了三个score。
2、然后对这三个score取softmax,获得了input-1、input-2、input-3各自的重要程度。
3、然后将这个重要程度乘上input-1、input-2、input-3的值向量,求和。
4、此时我们获得了input-1的输出。
如图所示,我们进行如下几步:
1、input-1的查询向量为[1, 0, 2],分别乘上input-1、input-2、input-3的键向量,获得三个score为2,4,4。
2、然后对这三个score取softmax,获得了input-1、input-2、input-3各自的重要程度,获得三个重要程度为0.0,0.5,0.5。
3、然后将这个重要程度乘上input-1、input-2、input-3的值向量,求和,即0.0 ∗ [ 1 , 2 , 3 ] + 0.5 ∗ [ 2 , 8 , 0 ] + 0.5 ∗ [ 2 , 6 , 3 ] = [ 2.0 , 7.0 , 1.5 ] 0.0 * [1, 2, 3] + 0.5 * [2, 8, 0] + 0.5 * [2, 6, 3] = [2.0, 7.0, 1.5] 0.0∗[1,2,3]+0.5∗[2,8,0]+0.5∗[2,6,3]=[2.0,7.0,1.5]。
4、此时我们获得了input-1的输出 [2.0, 7.0, 1.5]。
上述的例子中,序列长度仅为3,每个单位序列的特征长度仅为3,在VIT的Transformer Encoder中,序列长度为197,每个单位序列的特征长度为768 // num_heads。但计算过程是一样的。在实际运算时,我们采用矩阵进行运算。
II、Self-attention的矩阵运算
实际的矩阵运算过程如下图所示。我以实际矩阵为例子给大家解析:
输入的Query、Key、Value如下图所示:
首先利用 查询向量query 叉乘 转置后的键向量key,这一步可以通俗的理解为,利用查询向量去查询序列的特征,获得序列每个部分的重要程度score。
输出的每一行,都代表input-1、input-2、input-3,对当前input的贡献,我们对这个贡献值取一个softmax。
然后利用 score 叉乘 value,这一步可以通俗的理解为,将序列每个部分的重要程度重新施加到序列的值上去。
这个矩阵运算的代码如下所示,各位同学可以自己试试。
import numpy as np def soft_max(z): t = np.exp(z) a = np.exp(z) / np.expand_dims(np.sum(t, axis=1), 1) return a Query = np.array([ [1,0,2], [2,2,2], [2,1,3] ]) Key = np.array([ [0,1,1], [4,4,0], [2,3,1] ]) Value = np.array([ [1,2,3], [2,8,0], [2,6,3] ]) scores = Query @ Key.T print(scores) scores = soft_max(scores) print(scores) out = scores @ Value print(out)
III、MultiHead多头注意力机制
多头注意力机制的示意图如图所示:
这幅图给人的感觉略显迷茫,我们跳脱出这个图,直接从矩阵的shape入手会清晰很多。
在第一步进行图像的分割后,我们获得的特征层为197, 768。
在施加多头的时候,我们直接对196, 768的最后一维度进行分割,比如我们想分割成12个头,那么矩阵的shepe就变成了196, 12, 64。
然后我们将196, 12, 64进行转置,将12放到前面去,获得的特征层为12, 196, 64。之后我们忽略这个12,把它和batch维度同等对待,只对196, 64进行处理,其实也就是上面的注意力机制的过程了。
#--------------------------------------------------------------------------------------------------------------------# # Attention机制 # 将输入的特征qkv特征进行划分,首先生成query, key, value。query是查询向量、key是键向量、v是值向量。 # 然后利用 查询向量query 叉乘 转置后的键向量key,这一步可以通俗的理解为,利用查询向量去查询序列的特征,获得序列每个部分的重要程度score。 # 然后利用 score 叉乘 value,这一步可以通俗的理解为,将序列每个部分的重要程度重新施加到序列的值上去。 #--------------------------------------------------------------------------------------------------------------------# class Attention(Layer): def __init__(self, num_features, num_heads, **kwargs): super(Attention, self).__init__(**kwargs) self.num_features = num_features self.num_heads = num_heads self.projection_dim = num_features // num_heads def compute_output_shape(self, input_shape): return (input_shape[0], input_shape[1], input_shape[2] // 3) def call(self, inputs): #-----------------------------------------------# # 获得batch_size #-----------------------------------------------# bs = tf.shape(inputs)[0] #-----------------------------------------------# # b, 197, 3 * 768 -> b, 197, 3, 12, 64 #-----------------------------------------------# inputs = tf.reshape(inputs, [bs, -1, 3, self.num_heads, self.projection_dim]) #-----------------------------------------------# # b, 197, 3, 12, 64 -> 3, b, 12, 197, 64 #-----------------------------------------------# inputs = tf.transpose(inputs, [2, 0, 3, 1, 4]) #-----------------------------------------------# # 将query, key, value划分开 # query b, 12, 197, 64 # key b, 12, 197, 64 # value b, 12, 197, 64 #-----------------------------------------------# query, key, value = inputs[0], inputs[1], inputs[2] #-----------------------------------------------# # b, 12, 197, 64 @ b, 12, 197, 64 = b, 12, 197, 197 #-----------------------------------------------# score = tf.matmul(query, key, transpose_b=True) #-----------------------------------------------# # 进行数量级的缩放 #-----------------------------------------------# scaled_score = score / tf.math.sqrt(tf.cast(self.projection_dim, score.dtype)) #-----------------------------------------------# # b, 12, 197, 197 -> b, 12, 197, 197 #-----------------------------------------------# weights = tf.nn.softmax(scaled_score, axis=-1) #-----------------------------------------------# # b, 12, 197, 197 @ b, 12, 197, 64 = b, 12, 197, 64 #-----------------------------------------------# value = tf.matmul(weights, value) #-----------------------------------------------# # b, 12, 197, 64 -> b, 197, 12, 64 #-----------------------------------------------# value = tf.transpose(value, perm=[0, 2, 1, 3]) #-----------------------------------------------# # b, 197, 12, 64 -> b, 197, 768 #-----------------------------------------------# output = tf.reshape(value, (tf.shape(value)[0], tf.shape(value)[1], -1)) return output def MultiHeadSelfAttention(inputs, num_features, num_heads, dropout, name): #-----------------------------------------------# # qkv b, 197, 768 -> b, 197, 3 * 768 #-----------------------------------------------# qkv = Dense(int(num_features * 3), name = name + "qkv")(inputs) #-----------------------------------------------# # b, 197, 3 * 768 -> b, 197, 768 #-----------------------------------------------# x = Attention(num_features, num_heads)(qkv) #-----------------------------------------------# # 197, 768 -> 197, 768 #-----------------------------------------------# x = Dense(num_features, name = name + "proj")(x) x = Dropout(dropout)(x) return x
IV、TransformerBlock的构建。
在完成MultiHeadSelfAttention的构建后,我们需要在其后加上两个全连接。就构建了整个TransformerBlock。
def MLP(y, num_features, mlp_dim, dropout, name): y = Dense(mlp_dim, name = name + "fc1")(y) y = Gelu()(y) y = Dropout(dropout)(y) y = Dense(num_features, name = name + "fc2")(y) return y def TransformerBlock(inputs, num_features, num_heads, mlp_dim, dropout, name): #-----------------------------------------------# # 施加层标准化 #-----------------------------------------------# x = LayerNormalization(epsilon=1e-6, name = name + "norm1")(inputs) #-----------------------------------------------# # 施加多头注意力机制 #-----------------------------------------------# x = MultiHeadSelfAttention(x, num_features, num_heads, dropout, name = name + "attn.") x = Dropout(dropout)(x) #-----------------------------------------------# # 施加残差结构 #-----------------------------------------------# x = Add()([x, inputs]) #-----------------------------------------------# # 施加层标准化 #-----------------------------------------------# y = LayerNormalization(epsilon=1e-6, name = name + "norm2")(x) #-----------------------------------------------# # 施加两次全连接 #-----------------------------------------------# y = MLP(y, num_features, mlp_dim, dropout, name = name + "mlp.") y = Dropout(dropout)(y) #-----------------------------------------------# # 施加残差结构 #-----------------------------------------------# y = Add()([x, y]) return y
c、整个VIT模型的构建
整个VIT模型由一个Patch+Position Embedding加上多个TransformerBlock组成。典型的TransforerBlock的数量为12个。
def VisionTransformer(input_shape = [224, 224], patch_size = 16, num_layers = 12, num_features = 768, num_heads = 12, mlp_dim = 3072, classes = 1000, dropout = 0.1): #-----------------------------------------------# # 224, 224, 3 #-----------------------------------------------# inputs = Input(shape = (input_shape[0], input_shape[1], 3)) #-----------------------------------------------# # 224, 224, 3 -> 14, 14, 768 #-----------------------------------------------# x = Conv2D(num_features, patch_size, strides = patch_size, padding = "valid", name = "patch_embed.proj")(inputs) #-----------------------------------------------# # 14, 14, 768 -> 196, 768 #-----------------------------------------------# x = Reshape(((input_shape[0] // patch_size) * (input_shape[1] // patch_size), num_features))(x) #-----------------------------------------------# # 196, 768 -> 197, 768 #-----------------------------------------------# x = ClassToken(name="cls_token")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 #-----------------------------------------------# x = AddPositionEmbs(input_shape, patch_size, name="pos_embed")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 12次 #-----------------------------------------------# for n in range(num_layers): x = TransformerBlock( x, num_features= num_features, num_heads = num_heads, mlp_dim = mlp_dim, dropout = dropout, name = "blocks." + str(n) + ".", ) x = LayerNormalization( epsilon=1e-6, name="norm" )(x)
2、分类部分
在分类部分,VIT所做的工作是利用提取到的特征进行分类。
在进行特征提取的时候,我们会在图片序列中添加上Cls Token,该Token会作为一个单位的序列信息一起进行特征提取,提取的过程中,该Cls Token会与其它的特征进行特征交互,融合其它图片序列的特征。
最终,我们利用Multi-head Self-attention结构提取特征后的Cls Token进行全连接分类。
def VisionTransformer(input_shape = [224, 224], patch_size = 16, num_layers = 12, num_features = 768, num_heads = 12, mlp_dim = 3072, classes = 1000, dropout = 0.1): #-----------------------------------------------# # 224, 224, 3 #-----------------------------------------------# inputs = Input(shape = (input_shape[0], input_shape[1], 3)) #-----------------------------------------------# # 224, 224, 3 -> 14, 14, 768 #-----------------------------------------------# x = Conv2D(num_features, patch_size, strides = patch_size, padding = "valid", name = "patch_embed.proj")(inputs) #-----------------------------------------------# # 14, 14, 768 -> 196, 768 #-----------------------------------------------# x = Reshape(((input_shape[0] // patch_size) * (input_shape[1] // patch_size), num_features))(x) #-----------------------------------------------# # 196, 768 -> 197, 768 #-----------------------------------------------# x = ClassToken(name="cls_token")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 #-----------------------------------------------# x = AddPositionEmbs(input_shape, patch_size, name="pos_embed")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 12次 #-----------------------------------------------# for n in range(num_layers): x = TransformerBlock( x, num_features= num_features, num_heads = num_heads, mlp_dim = mlp_dim, dropout = dropout, name = "blocks." + str(n) + ".", ) x = LayerNormalization( epsilon=1e-6, name="norm" )(x) x = Lambda(lambda v: v[:, 0], name="ExtractToken")(x) x = Dense(classes, name="head")(x) x = Softmax()(x) return keras.models.Model(inputs, x)
Vision Transforme的构建代码
import math import keras import tensorflow as tf from keras import backend as K from keras.layers import (Add, Conv2D, Dense, Dropout, Input, Lambda, Layer, Reshape, Softmax) #--------------------------------------# # LayerNormalization # 层标准化的实现 #--------------------------------------# class LayerNormalization(keras.layers.Layer): def __init__(self, center=True, scale=True, epsilon=None, gamma_initializer='ones', beta_initializer='zeros', gamma_regularizer=None, beta_regularizer=None, gamma_constraint=None, beta_constraint=None, **kwargs): """Layer normalization layer See: [Layer Normalization](https://arxiv.org/pdf/1607.06450.pdf) :param center: Add an offset parameter if it is True. :param scale: Add a scale parameter if it is True. :param epsilon: Epsilon for calculating variance. :param gamma_initializer: Initializer for the gamma weight. :param beta_initializer: Initializer for the beta weight. :param gamma_regularizer: Optional regularizer for the gamma weight. :param beta_regularizer: Optional regularizer for the beta weight. :param gamma_constraint: Optional constraint for the gamma weight. :param beta_constraint: Optional constraint for the beta weight. :param kwargs: """ super(LayerNormalization, self).__init__(**kwargs) self.supports_masking = True self.center = center self.scale = scale if epsilon is None: epsilon = K.epsilon() * K.epsilon() self.epsilon = epsilon self.gamma_initializer = keras.initializers.get(gamma_initializer) self.beta_initializer = keras.initializers.get(beta_initializer) self.gamma_regularizer = keras.regularizers.get(gamma_regularizer) self.beta_regularizer = keras.regularizers.get(beta_regularizer) self.gamma_constraint = keras.constraints.get(gamma_constraint) self.beta_constraint = keras.constraints.get(beta_constraint) self.gamma, self.beta = None, None def get_config(self): config = { 'center': self.center, 'scale': self.scale, 'epsilon': self.epsilon, 'gamma_initializer': keras.initializers.serialize(self.gamma_initializer), 'beta_initializer': keras.initializers.serialize(self.beta_initializer), 'gamma_regularizer': keras.regularizers.serialize(self.gamma_regularizer), 'beta_regularizer': keras.regularizers.serialize(self.beta_regularizer), 'gamma_constraint': keras.constraints.serialize(self.gamma_constraint), 'beta_constraint': keras.constraints.serialize(self.beta_constraint), } base_config = super(LayerNormalization, self).get_config() return dict(list(base_config.items()) + list(config.items())) def compute_output_shape(self, input_shape): return input_shape def compute_mask(self, inputs, input_mask=None): return input_mask def build(self, input_shape): shape = input_shape[-1:] if self.scale: self.gamma = self.add_weight( shape=shape, initializer=self.gamma_initializer, regularizer=self.gamma_regularizer, constraint=self.gamma_constraint, name='gamma', ) if self.center: self.beta = self.add_weight( shape=shape, initializer=self.beta_initializer, regularizer=self.beta_regularizer, constraint=self.beta_constraint, name='beta', ) super(LayerNormalization, self).build(input_shape) def call(self, inputs, training=None): mean = K.mean(inputs, axis=-1, keepdims=True) variance = K.mean(K.square(inputs - mean), axis=-1, keepdims=True) std = K.sqrt(variance + self.epsilon) outputs = (inputs - mean) / std if self.scale: outputs *= self.gamma if self.center: outputs += self.beta return outputs #--------------------------------------# # Gelu激活函数的实现 # 利用近似的数学公式 #--------------------------------------# class Gelu(Layer): def __init__(self, **kwargs): super(Gelu, self).__init__(**kwargs) self.supports_masking = True def call(self, inputs): return 0.5 * inputs * (1 + tf.tanh(tf.sqrt(2 / math.pi) * (inputs + 0.044715 * tf.pow(inputs, 3)))) def get_config(self): config = super(Gelu, self).get_config() return config def compute_output_shape(self, input_shape): return input_shape #--------------------------------------------------------------------------------------------------------------------# # classtoken部分是transformer的分类特征。用于堆叠到序列化后的图片特征中,作为一个单位的序列特征进行特征提取。 # # 在利用步长为16x16的卷积将输入图片划分成14x14的部分后,将14x14部分的特征平铺,一幅图片会存在序列长度为196的特征。 # 此时生成一个classtoken,将classtoken堆叠到序列长度为196的特征上,获得一个序列长度为197的特征。 # 在特征提取的过程中,classtoken会与图片特征进行特征的交互。最终分类时,我们取出classtoken的特征,利用全连接分类。 #--------------------------------------------------------------------------------------------------------------------# class ClassToken(Layer): def __init__(self, cls_initializer='zeros', cls_regularizer=None, cls_constraint=None, **kwargs): super(ClassToken, self).__init__(**kwargs) self.cls_initializer = keras.initializers.get(cls_initializer) self.cls_regularizer = keras.regularizers.get(cls_regularizer) self.cls_constraint = keras.constraints.get(cls_constraint) def get_config(self): config = { 'cls_initializer': keras.initializers.serialize(self.cls_initializer), 'cls_regularizer': keras.regularizers.serialize(self.cls_regularizer), 'cls_constraint': keras.constraints.serialize(self.cls_constraint), } base_config = super(ClassToken, self).get_config() return dict(list(base_config.items()) + list(config.items())) def compute_output_shape(self, input_shape): return (input_shape[0], input_shape[1] + 1, input_shape[2]) def build(self, input_shape): self.num_features = input_shape[-1] self.cls = self.add_weight( shape = (1, 1, self.num_features), initializer = self.cls_initializer, regularizer = self.cls_regularizer, constraint = self.cls_constraint, name = 'cls', ) super(ClassToken, self).build(input_shape) def call(self, inputs): batch_size = tf.shape(inputs)[0] cls_broadcasted = tf.cast(tf.broadcast_to(self.cls, [batch_size, 1, self.num_features]), dtype = inputs.dtype) return tf.concat([cls_broadcasted, inputs], 1) #--------------------------------------------------------------------------------------------------------------------# # 为网络提取到的特征添加上位置信息。 # 以输入图片为224, 224, 3为例,我们获得的序列化后的图片特征为196, 768。加上classtoken后就是197, 768 # 此时生成的pos_Embedding的shape也为197, 768,代表每一个特征的位置信息。 #--------------------------------------------------------------------------------------------------------------------# class AddPositionEmbs(Layer): def __init__(self, image_shape, patch_size, pe_initializer='zeros', pe_regularizer=None, pe_constraint=None, **kwargs): super(AddPositionEmbs, self).__init__(**kwargs) self.image_shape = image_shape self.patch_size = patch_size self.pe_initializer = keras.initializers.get(pe_initializer) self.pe_regularizer = keras.regularizers.get(pe_regularizer) self.pe_constraint = keras.constraints.get(pe_constraint) def get_config(self): config = { 'pe_initializer': keras.initializers.serialize(self.pe_initializer), 'pe_regularizer': keras.regularizers.serialize(self.pe_regularizer), 'pe_constraint': keras.constraints.serialize(self.pe_constraint), } base_config = super(AddPositionEmbs, self).get_config() return dict(list(base_config.items()) + list(config.items())) def compute_output_shape(self, input_shape): return input_shape def build(self, input_shape): assert (len(input_shape) == 3), f"Number of dimensions should be 3, got {len(input_shape)}" length = (224 // self.patch_size) * (224 // self.patch_size) + 1 self.pe = self.add_weight( # shape = [1, input_shape[1], input_shape[2]], shape = [1, length, input_shape[2]], initializer = self.pe_initializer, regularizer = self.pe_regularizer, constraint = self.pe_constraint, name = 'pos_embedding', ) super(AddPositionEmbs, self).build(input_shape) def call(self, inputs): num_features = tf.shape(inputs)[2] cls_token_pe = self.pe[:, 0:1, :] img_token_pe = self.pe[:, 1: , :] img_token_pe = tf.reshape(img_token_pe, [1, (224 // self.patch_size), (224 // self.patch_size), num_features]) img_token_pe = tf.image.resize_bicubic(img_token_pe, (self.image_shape[0] // self.patch_size, self.image_shape[1] // self.patch_size), align_corners=False) img_token_pe = tf.reshape(img_token_pe, [1, -1, num_features]) pe = tf.concat([cls_token_pe, img_token_pe], axis = 1) return inputs + tf.cast(pe, dtype=inputs.dtype) #--------------------------------------------------------------------------------------------------------------------# # Attention机制 # 将输入的特征qkv特征进行划分,首先生成query, key, value。query是查询向量、key是键向量、v是值向量。 # 然后利用 查询向量query 叉乘 转置后的键向量key,这一步可以通俗的理解为,利用查询向量去查询序列的特征,获得序列每个部分的重要程度score。 # 然后利用 score 叉乘 value,这一步可以通俗的理解为,将序列每个部分的重要程度重新施加到序列的值上去。 #--------------------------------------------------------------------------------------------------------------------# class Attention(Layer): def __init__(self, num_features, num_heads, **kwargs): super(Attention, self).__init__(**kwargs) self.num_features = num_features self.num_heads = num_heads self.projection_dim = num_features // num_heads def compute_output_shape(self, input_shape): return (input_shape[0], input_shape[1], input_shape[2] // 3) def call(self, inputs): #-----------------------------------------------# # 获得batch_size #-----------------------------------------------# bs = tf.shape(inputs)[0] #-----------------------------------------------# # b, 197, 3 * 768 -> b, 197, 3, 12, 64 #-----------------------------------------------# inputs = tf.reshape(inputs, [bs, -1, 3, self.num_heads, self.projection_dim]) #-----------------------------------------------# # b, 197, 3, 12, 64 -> 3, b, 12, 197, 64 #-----------------------------------------------# inputs = tf.transpose(inputs, [2, 0, 3, 1, 4]) #-----------------------------------------------# # 将query, key, value划分开 # query b, 12, 197, 64 # key b, 12, 197, 64 # value b, 12, 197, 64 #-----------------------------------------------# query, key, value = inputs[0], inputs[1], inputs[2] #-----------------------------------------------# # b, 12, 197, 64 @ b, 12, 197, 64 = b, 12, 197, 197 #-----------------------------------------------# score = tf.matmul(query, key, transpose_b=True) #-----------------------------------------------# # 进行数量级的缩放 #-----------------------------------------------# scaled_score = score / tf.math.sqrt(tf.cast(self.projection_dim, score.dtype)) #-----------------------------------------------# # b, 12, 197, 197 -> b, 12, 197, 197 #-----------------------------------------------# weights = tf.nn.softmax(scaled_score, axis=-1) #-----------------------------------------------# # b, 12, 197, 197 @ b, 12, 197, 64 = b, 12, 197, 64 #-----------------------------------------------# value = tf.matmul(weights, value) #-----------------------------------------------# # b, 12, 197, 64 -> b, 197, 12, 64 #-----------------------------------------------# value = tf.transpose(value, perm=[0, 2, 1, 3]) #-----------------------------------------------# # b, 197, 12, 64 -> b, 197, 768 #-----------------------------------------------# output = tf.reshape(value, (tf.shape(value)[0], tf.shape(value)[1], -1)) return output def MultiHeadSelfAttention(inputs, num_features, num_heads, dropout, name): #-----------------------------------------------# # qkv b, 197, 768 -> b, 197, 3 * 768 #-----------------------------------------------# qkv = Dense(int(num_features * 3), name = name + "qkv")(inputs) #-----------------------------------------------# # b, 197, 3 * 768 -> b, 197, 768 #-----------------------------------------------# x = Attention(num_features, num_heads)(qkv) #-----------------------------------------------# # 197, 768 -> 197, 768 #-----------------------------------------------# x = Dense(num_features, name = name + "proj")(x) x = Dropout(dropout)(x) return x def MLP(y, num_features, mlp_dim, dropout, name): y = Dense(mlp_dim, name = name + "fc1")(y) y = Gelu()(y) y = Dropout(dropout)(y) y = Dense(num_features, name = name + "fc2")(y) return y def TransformerBlock(inputs, num_features, num_heads, mlp_dim, dropout, name): #-----------------------------------------------# # 施加层标准化 #-----------------------------------------------# x = LayerNormalization(epsilon=1e-6, name = name + "norm1")(inputs) #-----------------------------------------------# # 施加多头注意力机制 #-----------------------------------------------# x = MultiHeadSelfAttention(x, num_features, num_heads, dropout, name = name + "attn.") x = Dropout(dropout)(x) #-----------------------------------------------# # 施加残差结构 #-----------------------------------------------# x = Add()([x, inputs]) #-----------------------------------------------# # 施加层标准化 #-----------------------------------------------# y = LayerNormalization(epsilon=1e-6, name = name + "norm2")(x) #-----------------------------------------------# # 施加两次全连接 #-----------------------------------------------# y = MLP(y, num_features, mlp_dim, dropout, name = name + "mlp.") y = Dropout(dropout)(y) #-----------------------------------------------# # 施加残差结构 #-----------------------------------------------# y = Add()([x, y]) return y def VisionTransformer(input_shape = [224, 224], patch_size = 16, num_layers = 12, num_features = 768, num_heads = 12, mlp_dim = 3072, classes = 1000, dropout = 0.1): #-----------------------------------------------# # 224, 224, 3 #-----------------------------------------------# inputs = Input(shape = (input_shape[0], input_shape[1], 3)) #-----------------------------------------------# # 224, 224, 3 -> 14, 14, 768 #-----------------------------------------------# x = Conv2D(num_features, patch_size, strides = patch_size, padding = "valid", name = "patch_embed.proj")(inputs) #-----------------------------------------------# # 14, 14, 768 -> 196, 768 #-----------------------------------------------# x = Reshape(((input_shape[0] // patch_size) * (input_shape[1] // patch_size), num_features))(x) #-----------------------------------------------# # 196, 768 -> 197, 768 #-----------------------------------------------# x = ClassToken(name="cls_token")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 #-----------------------------------------------# x = AddPositionEmbs(input_shape, patch_size, name="pos_embed")(x) #-----------------------------------------------# # 197, 768 -> 197, 768 12次 #-----------------------------------------------# for n in range(num_layers): x = TransformerBlock( x, num_features= num_features, num_heads = num_heads, mlp_dim = mlp_dim, dropout = dropout, name = "blocks." + str(n) + ".", ) x = LayerNormalization( epsilon=1e-6, name="norm" )(x) x = Lambda(lambda v: v[:, 0], name="ExtractToken")(x) x = Dense(classes, name="head")(x) x = Softmax()(x) return keras.models.Model(inputs, x)
以上就是Keras实现Vision Transformer VIT模型示例详解的详细内容,更多关于Keras实现VIT模型的资料请关注脚本之家其它相关文章!