代码示例 / 计算机视觉 / 使用 Vision Transformer 进行图像分类

使用 Vision Transformer 进行图像分类

作者: Khalid Salama
创建日期 2021/01/18
上次修改 2021/01/18
描述: 实现用于图像分类的 Vision Transformer (ViT) 模型。

ⓘ 此示例使用 Keras 3

在 Colab 中查看 GitHub 源代码


简介

此示例实现了 Alexey Dosovitskiy 等人提出的 Vision Transformer (ViT) 模型,用于图像分类,并在 CIFAR-100 数据集上进行了演示。ViT 模型将 Transformer 架构与自注意力机制应用于图像块序列,无需使用卷积层。


设置

import os

os.environ["KERAS_BACKEND"] = "jax"  # @param ["tensorflow", "jax", "torch"]

import keras
from keras import layers
from keras import ops

import numpy as np
import matplotlib.pyplot as plt

准备数据

num_classes = 100
input_shape = (32, 32, 3)

(x_train, y_train), (x_test, y_test) = keras.datasets.cifar100.load_data()

print(f"x_train shape: {x_train.shape} - y_train shape: {y_train.shape}")
print(f"x_test shape: {x_test.shape} - y_test shape: {y_test.shape}")
x_train shape: (50000, 32, 32, 3) - y_train shape: (50000, 1)
x_test shape: (10000, 32, 32, 3) - y_test shape: (10000, 1)

配置超参数

learning_rate = 0.001
weight_decay = 0.0001
batch_size = 256
num_epochs = 10  # For real training, use num_epochs=100. 10 is a test value
image_size = 72  # We'll resize input images to this size
patch_size = 6  # Size of the patches to be extract from the input images
num_patches = (image_size // patch_size) ** 2
projection_dim = 64
num_heads = 4
transformer_units = [
    projection_dim * 2,
    projection_dim,
]  # Size of the transformer layers
transformer_layers = 8
mlp_head_units = [
    2048,
    1024,
]  # Size of the dense layers of the final classifier

使用数据增强

data_augmentation = keras.Sequential(
    [
        layers.Normalization(),
        layers.Resizing(image_size, image_size),
        layers.RandomFlip("horizontal"),
        layers.RandomRotation(factor=0.02),
        layers.RandomZoom(height_factor=0.2, width_factor=0.2),
    ],
    name="data_augmentation",
)
# Compute the mean and the variance of the training data for normalization.
data_augmentation.layers[0].adapt(x_train)

实现多层感知器 (MLP)

def mlp(x, hidden_units, dropout_rate):
    for units in hidden_units:
        x = layers.Dense(units, activation=keras.activations.gelu)(x)
        x = layers.Dropout(dropout_rate)(x)
    return x

实现将图像块创建作为层

class Patches(layers.Layer):
    def __init__(self, patch_size):
        super().__init__()
        self.patch_size = patch_size

    def call(self, images):
        input_shape = ops.shape(images)
        batch_size = input_shape[0]
        height = input_shape[1]
        width = input_shape[2]
        channels = input_shape[3]
        num_patches_h = height // self.patch_size
        num_patches_w = width // self.patch_size
        patches = keras.ops.image.extract_patches(images, size=self.patch_size)
        patches = ops.reshape(
            patches,
            (
                batch_size,
                num_patches_h * num_patches_w,
                self.patch_size * self.patch_size * channels,
            ),
        )
        return patches

    def get_config(self):
        config = super().get_config()
        config.update({"patch_size": self.patch_size})
        return config

让我们显示样本图像的图像块

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 = ops.image.resize(
    ops.convert_to_tensor([image]), size=(image_size, image_size)
)
patches = Patches(patch_size)(resized_image)
print(f"Image size: {image_size} X {image_size}")
print(f"Patch size: {patch_size} X {patch_size}")
print(f"Patches per image: {patches.shape[1]}")
print(f"Elements per patch: {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 = ops.reshape(patch, (patch_size, patch_size, 3))
    plt.imshow(ops.convert_to_numpy(patch_img).astype("uint8"))
    plt.axis("off")
Image size: 72 X 72
Patch size: 6 X 6
Patches per image: 144
Elements per patch: 108

png

png


实现图像块编码层

PatchEncoder 层将通过将其投影到大小为 projection_dim 的向量中来线性变换图像块。此外,它还将可学习的位置嵌入添加到投影向量中。

class PatchEncoder(layers.Layer):
    def __init__(self, num_patches, projection_dim):
        super().__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 = ops.expand_dims(
            ops.arange(start=0, stop=self.num_patches, step=1), axis=0
        )
        projected_patches = self.projection(patch)
        encoded = projected_patches + self.position_embedding(positions)
        return encoded

    def get_config(self):
        config = super().get_config()
        config.update({"num_patches": self.num_patches})
        return config

构建 ViT 模型

ViT 模型由多个 Transformer 块组成,这些块使用 layers.MultiHeadAttention 层作为应用于图像块序列的自注意力机制。Transformer 块生成一个 [batch_size, num_patches, projection_dim] 张量,该张量通过具有 softmax 的分类器头部进行处理,以生成最终的类别概率输出。

论文 中描述的技术不同,该技术将可学习的嵌入添加到编码图像块的序列的开头以用作图像表示,最终 Transformer 块的所有输出都使用 layers.Flatten() 重塑并用作分类器头的图像表示输入。请注意,layers.GlobalAveragePooling1D 层也可以用作替代方案来聚合 Transformer 块的输出,尤其是在图像块的数量和投影维度很大时。

def create_vit_classifier():
    inputs = keras.Input(shape=input_shape)
    # Augment data.
    augmented = data_augmentation(inputs)
    # Create patches.
    patches = Patches(patch_size)(augmented)
    # Encode patches.
    encoded_patches = PatchEncoder(num_patches, projection_dim)(patches)

    # Create multiple layers of the Transformer block.
    for _ in range(transformer_layers):
        # Layer normalization 1.
        x1 = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
        # Create a multi-head attention layer.
        attention_output = layers.MultiHeadAttention(
            num_heads=num_heads, key_dim=projection_dim, dropout=0.1
        )(x1, x1)
        # Skip connection 1.
        x2 = layers.Add()([attention_output, encoded_patches])
        # Layer normalization 2.
        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])

    # Create a [batch_size, projection_dim] tensor.
    representation = layers.LayerNormalization(epsilon=1e-6)(encoded_patches)
    representation = layers.Flatten()(representation)
    representation = layers.Dropout(0.5)(representation)
    # Add MLP.
    features = mlp(representation, hidden_units=mlp_head_units, dropout_rate=0.5)
    # Classify outputs.
    logits = layers.Dense(num_classes)(features)
    # Create the Keras model.
    model = keras.Model(inputs=inputs, outputs=logits)
    return model

编译、训练和评估模型

def run_experiment(model):
    optimizer = keras.optimizers.AdamW(
        learning_rate=learning_rate, weight_decay=weight_decay
    )

    model.compile(
        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.weights.h5"
    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)


def plot_history(item):
    plt.plot(history.history[item], label=item)
    plt.plot(history.history["val_" + item], label="val_" + item)
    plt.xlabel("Epochs")
    plt.ylabel(item)
    plt.title("Train and Validation {} Over Epochs".format(item), fontsize=14)
    plt.legend()
    plt.grid()
    plt.show()


plot_history("loss")
plot_history("top-5-accuracy")
Epoch 1/10
...
Epoch 10/10
 176/176 ━━━━━━━━━━━━━━━━━━━━ 449s 3s/step - accuracy: 0.0790 - loss: 3.9468 - top-5-accuracy: 0.2711 - val_accuracy: 0.0986 - val_loss: 3.8537 - val_top-5-accuracy: 0.3052

 313/313 ━━━━━━━━━━━━━━━━━━━━ 66s 198ms/step - accuracy: 0.1001 - loss: 3.8428 - top-5-accuracy: 0.3107
Test accuracy: 10.61%
Test top 5 accuracy: 31.51%

png

png

经过 100 个 epoch 后,ViT 模型在测试数据上实现了大约 55% 的准确率和 82% 的前 5 名准确率。这些在 CIFAR-100 数据集上并非具有竞争力的结果,因为从头开始在相同数据上训练的 ResNet50V2 可以达到 67% 的准确率。

请注意,论文 中报告的最先进的结果是通过使用 JFT-300M 数据集对 ViT 模型进行预训练,然后在目标数据集上对其进行微调来实现的。为了在不进行预训练的情况下提高模型质量,您可以尝试训练模型更多 epoch、使用更多 Transformer 层、调整输入图像大小、更改图像块大小或增加投影维度。此外,如论文中所述,模型的质量不仅受架构选择的影响,还受学习率调度、优化器、权重衰减等参数的影响。在实践中,建议对使用大型、高分辨率数据集预训练的 ViT 模型进行微调。