Keras构建用于分类任务的Transformer（Vision Transformer/VIT）
文章目录
- 一、Vision Transformer (ViT)详细信息
- 二、Vision Transformer结构
- 三、Keras实现
-
- 3.1 相关包
- 3.2 数据读取
- 3.3 声明超参数
- 3.4 使用数据增强方法
- 3.5 计算训练数据的平均值和方差进行归一化
- 3.6 定义multilayer perceptron (MLP)
- 3.7 定义块
- 3.8 数据可视化
- 3.9 实现Encoding Layer
- 3.10 构建ViT模型
- 3.11 训练+评估（AdamW可以换成Adam，效果可能还更好）
- 四、完整代码
- 五、可视化
-
- 5.1 待切块图和处理后的切块图
- 5.2 模型结构参数
一、Vision Transformer (ViT)详细信息
| VIT | 详细信息 |
|---|---|
| 文献←下载 | An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale |
| 案例数据集 | CIFAR-100 |
更多网络模型及出处可见： 各神经网络参考文献整理
二、Vision Transformer结构
Transformer已经基本取代RNNs(包括变体LSTM ,GRU)，成为自然语言处理（NLP）领域的主流模型。Dosovitskiy等人将该模型迁移到计算机视觉领域，并且尽量减少了对Transformer的更改，因为是分类，所以模型的输出用全连接层代替。由此，Vision Transformer(ViT)应运而生。这是一种用于分类任务的改进Transformer。
ViT结构如下图所示
原作者的文章可能不太详细。为了便于理解，可参考文献提供的结构：
Vision Transformers for Remote Sensing Image Classification
三、Keras实现
经过测试，多头自注意力机制需要Tensorflow2.4、2.5(2.1-2.3是没有的)等带有MultiHeadAttention包的版本，相关配置在这边→点我
代码按顺序输入即可
3.1 相关包
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
#下包非必须，可以注释，后续内容替换为adam梯度优化算法，经过测试效果比原码好
import tensorflow_addons as tfa
3.2 数据读取
# 类别数
num_classes = 100
# 数据大小
input_shape = (32, 32, 3)
# 读取cifar100数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.cifar100.load_data()
print(f"x_train 大小: {x_train.shape} - y_train 大小: {y_train.shape}")
print(f"x_test 大小: {x_test.shape} - y_test 大小: {y_test.shape}")
3.3 声明超参数
learning_rate = 0.001
weight_decay = 0.0001
batch_size = 256
num_epochs = 100
image_size = 72 # 改变图形大小
patch_size = 6 # 输入图片拆分的块大小
num_patches = (image_size // patch_size) ** 2
projection_dim = 64
num_heads = 4
transformer_units = [
projection_dim * 2,
projection_dim,
] # Transformer layers的大小
transformer_layers = 8
mlp_head_units = [2048, 1024] # 输出部分的MLP全连接层的大小
3.4 使用数据增强方法
data_augmentation = keras.Sequential(
[
layers.experimental.preprocessing.Normalization(),
layers.experimental.preprocessing.Resizing(image_size, image_size),
layers.experimental.preprocessing.RandomFlip("horizontal"),
layers.experimental.preprocessing.RandomRotation(factor=0.02),
layers.experimental.preprocessing.RandomZoom(
height_factor=0.2, width_factor=0.2
),
],
name="data_augmentation",
)
3.5 计算训练数据的平均值和方差进行归一化
data_augmentation.layers[0].adapt(x_train)
3.6 定义multilayer perceptron (MLP)
def mlp(x, hidden_units, dropout_rate):
for units in hidden_units:
x = layers.Dense(units, activation=tf.nn.gelu)(x)
x = layers.Dropout(dropout_rate)(x)
return x
3.7 定义块
class Patches(layers.Layer):
def __init__(self, patch_size):
super(Patches, self).__init__()
self.patch_size = patch_size
def call(self, images):
batch_size = tf.shape(images)[0]
patches = tf.image.extract_patches(
images=images,
sizes=[1, self.patch_size, self.patch_size, 1],
strides=[1, self.patch_size, self.patch_size, 1],
rates=[1, 1, 1, 1],
padding="VALID",
)
patch_dims = patches.shape[-1]
patches = tf.reshape(patches, [batch_size, -1, patch_dims])
return patches
3.8 数据可视化
import matplotlib.pyplot as plt
plt.figure(figsize=(4, 4))
image = x_train[np.random.choice(range(x_train.shape[0]))]
plt.imshow(image.astype("uint8"))
plt.axis("off")
resized_image = tf.image.resize(
tf.convert_to_tensor([image]), size=(image_size, image_size)
)
patches = Patches(patch_size)(resized_image)
print(f"图片大小: {image_size} X {image_size}")
print(f"切块大小e: {patch_size} X {patch_size}")
print(f"每个图对应的切块大小: {patches.shape[1]}")
print(f"每个块对应的元素: {patches.shape[-1]}")
n = int(np.sqrt(patches.shape[1]))
plt.figure(figsize=(4, 4))
for i, patch in enumerate(patches[0]):
ax = plt.subplot(n, n, i + 1)
patch_img = tf.reshape(patch, (patch_size, patch_size, 3))
plt.imshow(patch_img.numpy().astype("uint8"))
plt.axis("off")
3.9 实现Encoding Layer
class PatchEncoder(layers.Layer):
def __init__(self, num_patches, projection_dim):
super(PatchEncoder, self).__init__()
self.num_patches = num_patches
self.projection = layers.Dense(units=projection_dim)
self.position_embedding = layers.Embedding(
input_dim=num_patches, output_dim=projection_dim
)
def call(self, patch):
positions = tf.range(start=0, limit=self.num_patches, delta=1)
encoded = self.projection(patch) + self.position_embedding(positions)
return encoded
这里大概率会报错，方法需要声明，方法可见此文
3.10 构建ViT模型
def create_vit_classifier():
inputs = layers.Input(shape=input_shape)
# 数据增强
augmented = data_augmentation(inputs)
# 创建块.
patches = Patches(patch_size)(augmented)
# Encode patches.
encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)
# 创建多个Transformer encoding 块
for _ in range(transformer_layers):
# Layer normalization 1.
x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
# 创建多头自注意力机制 multi-head attention layer，这里经过测试Tensorflow2.5可用
attention_output = layers.MultiHeadAttention(
num_heads=num_heads, key_dim=projection_dim, dropout=0.1
)(x1, x1)
# Skip connection.
x2 = layers.Add()([attention_output, encoded_patches])
x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
# MLP.
x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)
# Skip connection 2.
encoded_patches = layers.Add()([x3, x2])
representation = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
representation = layers.Flatten()(representation)
representation = layers.Dropout(0.5)(representation)
# 增加MLP.
features = mlp(representation, hidden_units=mlp_head_units, dropout_rate=0.5)
# 输出分类.
logits = layers.Dense(num_classes)(features)
# 构建
model = keras.Model(inputs=inputs, outputs=logits)
model.summary()
return model
3.11 训练+评估（AdamW可以换成Adam，效果可能还更好）
#tfa.方法可替换为adam
def run_experiment(model):
optimizer = tfa.optimizers.AdamW(
learning_rate=learning_rate, weight_decay=weight_decay
)
model.compile(
# 下述可直接替换为 optimizer='adam',
optimizer=optimizer,
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[
keras.metrics.SparseCategoricalAccuracy(name="accuracy"),
keras.metrics.SparseTopKCategoricalAccuracy(5, name="top-5-accuracy"),
],
)
checkpoint_filepath = "/tmp/checkpoint"
checkpoint_callback = keras.callbacks.ModelCheckpoint(
checkpoint_filepath,
monitor="val_accuracy",
save_best_only=True,
save_weights_only=True,
)
history = model.fit(
x=x_train,
y=y_train,
batch_size=batch_size,
epochs=num_epochs,
validation_split=0.1,
callbacks=[checkpoint_callback],
)
model.load_weights(checkpoint_filepath)
_, accuracy, top_5_accuracy = model.evaluate(x_test, y_test)
print(f"Test accuracy: {round(accuracy * 100, 2)}%")
print(f"Test top 5 accuracy: {round(top_5_accuracy * 100, 2)}%")
return history
vit_classifier = create_vit_classifier()
history = run_experiment(vit_classifier)
四、完整代码
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import matplotlib.pyplot as plt
#非必须包↓若用adamw则加，实测用keras自带的adam效果更好
import tensorflow_addons as tfa
# 类别数
num_classes = 100
# 数据大小
input_shape = (32, 32, 3)
# 读取cifar100数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.cifar100.load_data()
print(f"x_train 大小: {x_train.shape} - y_train 大小: {y_train.shape}")
print(f"x_test 大小: {x_test.shape} - y_test 大小: {y_test.shape}")
learning_rate = 0.001
weight_decay = 0.0001
batch_size = 256
num_epochs = 100
image_size = 72 # 改变图形大小
patch_size = 6 # 输入图片拆分的块大小
num_patches = (image_size // patch_size) ** 2
projection_dim = 64
num_heads = 4
transformer_units = [
projection_dim * 2,
projection_dim,
] # Transformer layers的大小
transformer_layers = 8
mlp_head_units = [2048, 1024] # 输出部分的MLP全连接层的大小
data_augmentation = keras.Sequential(
[
layers.experimental.preprocessing.Normalization(),
layers.experimental.preprocessing.Resizing(image_size, image_size),
layers.experimental.preprocessing.RandomFlip("horizontal"),
layers.experimental.preprocessing.RandomRotation(factor=0.02),
layers.experimental.preprocessing.RandomZoom(
height_factor=0.2, width_factor=0.2
),
],
name="data_augmentation",
)
data_augmentation.layers[0].adapt(x_train)
def mlp(x, hidden_units, dropout_rate):
for units in hidden_units:
x = layers.Dense(units, activation=tf.nn.gelu)(x)
x = layers.Dropout(dropout_rate)(x)
return x
class Patches(layers.Layer):
def __init__(self, patch_size):
super(Patches, self).__init__()
self.patch_size = patch_size
def call(self, images):
batch_size = tf.shape(images)[0]
patches = tf.image.extract_patches(
images=images,
sizes=[1, self.patch_size, self.patch_size, 1],
strides=[1, self.patch_size, self.patch_size, 1],
rates=[1, 1, 1, 1],
padding="VALID",
)
patch_dims = patches.shape[-1]
patches = tf.reshape(patches, [batch_size, -1, patch_dims])
return patches
plt.figure(figsize=(4, 4))
image = x_train[np.random.choice(range(x_train.shape[0]))]
plt.imshow(image.astype("uint8"))
plt.axis("off")
resized_image = tf.image.resize(
tf.convert_to_tensor([image]), size=(image_size, image_size)
)
patches = Patches(patch_size)(resized_image)
print(f"图片大小: {image_size} X {image_size}")
print(f"切块大小e: {patch_size} X {patch_size}")
print(f"每个图对应的切块大小: {patches.shape[1]}")
print(f"每个块对应的元素: {patches.shape[-1]}")
n = int(np.sqrt(patches.shape[1]))
plt.figure(figsize=(4, 4))
for i, patch in enumerate(patches[0]):
ax = plt.subplot(n, n, i + 1)
patch_img = tf.reshape(patch, (patch_size, patch_size, 3))
plt.imshow(patch_img.numpy().astype("uint8"))
plt.axis("off")
class PatchEncoder(layers.Layer):
def __init__(self, num_patches, projection_dim):
super(PatchEncoder, self).__init__()
self.num_patches = num_patches
self.projection = layers.Dense(units=projection_dim)
self.position_embedding = layers.Embedding(
input_dim=num_patches, output_dim=projection_dim
)
#这里call后需要定义get_config函数，命名自拟，文章3.9中给出
def call(self, patch):
positions = tf.range(start=0, limit=self.num_patches, delta=1)
encoded = self.projection(patch) + self.position_embedding(positions)
return encoded
def create_vit_classifier():
inputs = layers.Input(shape=input_shape)
# 数据增强
augmented = data_augmentation(inputs)
# Create patches.
patches = Patches(patch_size)(augmented)
# Encode patches.
encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)
# 创建多个Transformer encoding 块
for _ in range(transformer_layers):
# Layer normalization 1.
x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
# 创建多头自注意力机制 multi-head attention layer，这里经过测试Tensorflow2.5可用
attention_output = layers.MultiHeadAttention(
num_heads=num_heads, key_dim=projection_dim, dropout=0.1
)(x1, x1)
# Skip connection.
x2 = layers.Add()([attention_output, encoded_patches])
x3 = layers.LayerNormalization(epsilon=1e-6)(x2)
# MLP.
x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)
# Skip connection 2.
encoded_patches = layers.Add()([x3, x2])
representation = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
representation = layers.Flatten()(representation)
representation = layers.Dropout(0.5)(representation)
# 增加MLP.
features = mlp(representation, hidden_units=mlp_head_units, dropout_rate=0.5)
# 输出分类.
logits = layers.Dense(num_classes)(features)
# 构建
model = keras.Model(inputs=inputs, outputs=logits)
model.summary()
return model
#tfa.方法可替换为adam
def run_experiment(model):
optimizer = tfa.optimizers.AdamW(
learning_rate=learning_rate, weight_decay=weight_decay
)
model.compile(
# 下述可直接替换为 optimizer='adam',
optimizer=optimizer,
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[
keras.metrics.SparseCategoricalAccuracy(name="accuracy"),
keras.metrics.SparseTopKCategoricalAccuracy(5, name="top-5-accuracy"),
],
)
checkpoint_filepath = "/tmp/checkpoint"
checkpoint_callback = keras.callbacks.ModelCheckpoint(
checkpoint_filepath,
monitor="val_accuracy",
save_best_only=True,
save_weights_only=True,
)
history = model.fit(
x=x_train,
y=y_train,
batch_size=batch_size,
epochs=num_epochs,
validation_split=0.1,
callbacks=[checkpoint_callback],
)
model.load_weights(checkpoint_filepath)
_, accuracy, top_5_accuracy = model.evaluate(x_test, y_test)
print(f"Test accuracy: {round(accuracy * 100, 2)}%")
print(f"Test top 5 accuracy: {round(top_5_accuracy * 100, 2)}%")
return history
vit_classifier = create_vit_classifier()
history = run_experiment(vit_classifier)
五、可视化
5.1 待切块图和处理后的切块图
原始图
切块处理后的图（12*12个块，每个块大小为108个像素点）
5.2 模型结构参数
就参数量来说，Transformer对内存的占用率还是很大（相对一般的模型）