政安晨：【Keras机器学习实践要点】（二十六）—

简介

设置

卷积

演变

测试逆卷积层

图像分类

获取 CIFAR10 数据集

数据可视化

卷积神经网络

逆向传播神经网络

比较

损失图和准确率图

可视化卷积核

结论

政安晨的个人主页：政安晨

欢迎 👍点赞✍评论⭐收藏

收录专栏: TensorFlow与Keras机器学习实战

希望政安晨的博客能够对您有所裨益，如有不足之处，欢迎在评论区提出指正！

本文目标：深入研究特定位置和渠道无关的“内卷”核函数。

简介

卷积是大多数现代计算机视觉神经网络的基础。卷积核具有空间无关性和通道特定性。因此，它无法适应不同空间位置的不同视觉模式。除了与位置相关的问题，卷积的作用范围也给捕捉远距离空间相互作用带来了挑战。

为了解决上述问题，Li 等人在《卷积》一书中重新思考了卷积的特性：Inverting the Inherence of Convolution for VisualRecognition》一书中重新思考了卷积的特性。

作者提出了 "卷积内核"（involution kernel），它具有位置特定性和通道无关性。由于这种操作具有特定位置的性质，作者认为自我注意属于内卷的设计范例。

本示例介绍了内卷核，比较了两种图像分类模型（一种是卷积模型，另一种是内卷模型），并尝试将其与自我注意层相提并论。

设置

import osos.environ["KERAS_BACKEND"] = "tensorflow"import tensorflow as tf
import keras
import matplotlib.pyplot as plt# Set seed for reproducibility.
tf.random.set_seed(42)

卷积

在计算机视觉中，卷积是一种重要的操作，用于图像处理和特征提取。卷积通过将一个滤波器或核与输入图像进行逐像素的计算，生成一个新的输出图像。

卷积操作与图像处理中常用的滤波操作类似，滤波器在图像上滑动并与图像的局部区域进行点乘，然后将乘积相加得到新的像素值。这个过程可以看作是将滤波器与图像进行卷积计算。

卷积操作的一个重要特性是它具有局部感知性。即卷积运算只计算滤波器与图像窗口内像素的乘积和，并将结果相加。这样的操作可以有效地提取图像的局部特征，例如边缘和纹理等。

卷积在计算机视觉中有广泛的应用。它可以用于图像增强、边缘检测、特征提取、目标检测和图像分类等任务。卷积神经网络（Convolutional Neural Network，CNN）是计算机视觉领域中最重要的算法之一，它借鉴了卷积操作的思想，并在深度学习中取得了很大的成功。

总之，卷积是计算机视觉中的一种重要操作，它能够提取图像的局部特征，并被广泛应用于图像处理和特征提取任务中。

卷积是计算机视觉深度神经网络的支柱。要理解卷积，就必须先谈谈卷积操作。

考虑一个维数为 H、W 和 C_in 的输入张量 X。我们取一组 C_out 卷积核，每个核的形状为 K、K、C_in。通过对输入张量和卷积核进行乘加运算，我们得到一个维数为 H、W、C_out 的输出张量 Y。

在上图中，C_out=3。这使得输出张量的形状为 H、W 和 3。我们可以注意到，卷积核并不依赖于输入张量的空间位置，因此它与位置无关。另一方面，输出张量中的每个通道都基于特定的卷积滤波器，这使得它具有特定通道的特性。

演变

我们的想法是让操作既能针对特定位置，又能与通道特定性无关。要实现这些特定属性是一项挑战。如果使用固定数量的卷积核（针对每个空间位置），我们将无法处理可变分辨率的输入张量。

为了解决这个问题，咱们考虑根据特定的空间位置生成每个内核。通过这种方法，我们应该可以轻松处理变分辨率输入张量。下图提供了这种内核生成方法的直观图。

class Involution(keras.layers.Layer):def __init__(self, channel, group_number, kernel_size, stride, reduction_ratio, name):super().__init__(name=name)# Initialize the parameters.self.channel = channelself.group_number = group_numberself.kernel_size = kernel_sizeself.stride = strideself.reduction_ratio = reduction_ratiodef build(self, input_shape):# Get the shape of the input.(_, height, width, num_channels) = input_shape# Scale the height and width with respect to the strides.height = height // self.stridewidth = width // self.stride# Define a layer that average pools the input tensor# if stride is more than 1.self.stride_layer = (keras.layers.AveragePooling2D(pool_size=self.stride, strides=self.stride, padding="same")if self.stride > 1else tf.identity)# Define the kernel generation layer.self.kernel_gen = keras.Sequential([keras.layers.Conv2D(filters=self.channel // self.reduction_ratio, kernel_size=1),keras.layers.BatchNormalization(),keras.layers.ReLU(),keras.layers.Conv2D(filters=self.kernel_size * self.kernel_size * self.group_number,kernel_size=1,),])# Define reshape layersself.kernel_reshape = keras.layers.Reshape(target_shape=(height,width,self.kernel_size * self.kernel_size,1,self.group_number,))self.input_patches_reshape = keras.layers.Reshape(target_shape=(height,width,self.kernel_size * self.kernel_size,num_channels // self.group_number,self.group_number,))self.output_reshape = keras.layers.Reshape(target_shape=(height, width, num_channels))def call(self, x):# Generate the kernel with respect to the input tensor.# B, H, W, K*K*Gkernel_input = self.stride_layer(x)kernel = self.kernel_gen(kernel_input)# reshape the kerenl# B, H, W, K*K, 1, Gkernel = self.kernel_reshape(kernel)# Extract input patches.# B, H, W, K*K*Cinput_patches = tf.image.extract_patches(images=x,sizes=[1, self.kernel_size, self.kernel_size, 1],strides=[1, self.stride, self.stride, 1],rates=[1, 1, 1, 1],padding="SAME",)# Reshape the input patches to align with later operations.# B, H, W, K*K, C//G, Ginput_patches = self.input_patches_reshape(input_patches)# Compute the multiply-add operation of kernels and patches.# B, H, W, K*K, C//G, Goutput = tf.multiply(kernel, input_patches)# B, H, W, C//G, Goutput = tf.reduce_sum(output, axis=3)# Reshape the output kernel.# B, H, W, Coutput = self.output_reshape(output)# Return the output tensor and the kernel.return output, kernel

测试逆卷积层

# Define the input tensor.
input_tensor = tf.random.normal((32, 256, 256, 3))# Compute involution with stride 1.
output_tensor, _ = Involution(channel=3, group_number=1, kernel_size=5, stride=1, reduction_ratio=1, name="inv_1"
)(input_tensor)
print(f"with stride 1 ouput shape: {output_tensor.shape}")# Compute involution with stride 2.
output_tensor, _ = Involution(channel=3, group_number=1, kernel_size=5, stride=2, reduction_ratio=1, name="inv_2"
)(input_tensor)
print(f"with stride 2 ouput shape: {output_tensor.shape}")# Compute involution with stride 1, channel 16 and reduction ratio 2.
output_tensor, _ = Involution(channel=16, group_number=1, kernel_size=5, stride=1, reduction_ratio=2, name="inv_3"
)(input_tensor)
print("with channel 16 and reduction ratio 2 ouput shape: {}".format(output_tensor.shape)
)

演绎展示：

with stride 1 ouput shape: (32, 256, 256, 3)
with stride 2 ouput shape: (32, 128, 128, 3)
with channel 16 and reduction ratio 2 ouput shape: (32, 256, 256, 3)

图像分类

在本文中，我们将建立一个图像分类器模型。将有两个模型，一个是卷积模型，另一个是渐开线模型。

图像分类模型的灵感主要来自谷歌的卷积神经网络（CNN）教程。

获取 CIFAR10 数据集

# Load the CIFAR10 dataset.
print("loading the CIFAR10 dataset...")
((train_images, train_labels),(test_images,test_labels,),
) = keras.datasets.cifar10.load_data()# Normalize pixel values to be between 0 and 1.
(train_images, test_images) = (train_images / 255.0, test_images / 255.0)# Shuffle and batch the dataset.
train_ds = (tf.data.Dataset.from_tensor_slices((train_images, train_labels)).shuffle(256).batch(256)
)
test_ds = tf.data.Dataset.from_tensor_slices((test_images, test_labels)).batch(256)

加载 CIFAR10 数据集...

数据可视化

class_names = ["airplane","automobile","bird","cat","deer","dog","frog","horse","ship","truck",
]plt.figure(figsize=(10, 10))
for i in range(25):plt.subplot(5, 5, i + 1)plt.xticks([])plt.yticks([])plt.grid(False)plt.imshow(train_images[i])plt.xlabel(class_names[train_labels[i][0]])
plt.show()

演绎展示：

卷积神经网络

# Build the conv model.
print("building the convolution model...")
conv_model = keras.Sequential([keras.layers.Conv2D(32, (3, 3), input_shape=(32, 32, 3), padding="same"),keras.layers.ReLU(name="relu1"),keras.layers.MaxPooling2D((2, 2)),keras.layers.Conv2D(64, (3, 3), padding="same"),keras.layers.ReLU(name="relu2"),keras.layers.MaxPooling2D((2, 2)),keras.layers.Conv2D(64, (3, 3), padding="same"),keras.layers.ReLU(name="relu3"),keras.layers.Flatten(),keras.layers.Dense(64, activation="relu"),keras.layers.Dense(10),]
)# Compile the mode with the necessary loss function and optimizer.
print("compiling the convolution model...")
conv_model.compile(optimizer="adam",loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),metrics=["accuracy"],
)# Train the model.
print("conv model training...")
conv_hist = conv_model.fit(train_ds, epochs=20, validation_data=test_ds)

演绎展示：

building the convolution model...
compiling the convolution model...
conv model training...
Epoch 1/20196/196 ━━━━━━━━━━━━━━━━━━━━ 6s 15ms/step - accuracy: 0.3068 - loss: 1.9000 - val_accuracy: 0.4861 - val_loss: 1.4593
Epoch 2/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.5153 - loss: 1.3603 - val_accuracy: 0.5741 - val_loss: 1.1913
Epoch 3/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.5949 - loss: 1.1517 - val_accuracy: 0.6095 - val_loss: 1.0965
Epoch 4/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.6414 - loss: 1.0330 - val_accuracy: 0.6260 - val_loss: 1.0635
Epoch 5/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.6690 - loss: 0.9485 - val_accuracy: 0.6622 - val_loss: 0.9833
Epoch 6/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.6951 - loss: 0.8764 - val_accuracy: 0.6783 - val_loss: 0.9413
Epoch 7/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.7122 - loss: 0.8167 - val_accuracy: 0.6856 - val_loss: 0.9134
Epoch 8/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.7299 - loss: 0.7709 - val_accuracy: 0.7001 - val_loss: 0.8792
Epoch 9/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.7467 - loss: 0.7288 - val_accuracy: 0.6992 - val_loss: 0.8821
Epoch 10/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.7591 - loss: 0.6982 - val_accuracy: 0.7235 - val_loss: 0.8237
Epoch 11/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.7725 - loss: 0.6550 - val_accuracy: 0.7115 - val_loss: 0.8521
Epoch 12/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.7808 - loss: 0.6302 - val_accuracy: 0.7051 - val_loss: 0.8823
Epoch 13/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.7860 - loss: 0.6101 - val_accuracy: 0.7122 - val_loss: 0.8635
Epoch 14/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.7998 - loss: 0.5786 - val_accuracy: 0.7214 - val_loss: 0.8348
Epoch 15/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.8117 - loss: 0.5473 - val_accuracy: 0.7139 - val_loss: 0.8835
Epoch 16/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.8168 - loss: 0.5267 - val_accuracy: 0.7155 - val_loss: 0.8840
Epoch 17/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.8266 - loss: 0.5022 - val_accuracy: 0.7239 - val_loss: 0.8576
Epoch 18/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.8374 - loss: 0.4750 - val_accuracy: 0.7262 - val_loss: 0.8756
Epoch 19/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.8452 - loss: 0.4505 - val_accuracy: 0.7235 - val_loss: 0.9049
Epoch 20/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.8531 - loss: 0.4283 - val_accuracy: 0.7304 - val_loss: 0.8962

逆向传播神经网络

# Build the involution model.
print("building the involution model...")inputs = keras.Input(shape=(32, 32, 3))
x, _ = Involution(channel=3, group_number=1, kernel_size=3, stride=1, reduction_ratio=2, name="inv_1"
)(inputs)
x = keras.layers.ReLU()(x)
x = keras.layers.MaxPooling2D((2, 2))(x)
x, _ = Involution(channel=3, group_number=1, kernel_size=3, stride=1, reduction_ratio=2, name="inv_2"
)(x)
x = keras.layers.ReLU()(x)
x = keras.layers.MaxPooling2D((2, 2))(x)
x, _ = Involution(channel=3, group_number=1, kernel_size=3, stride=1, reduction_ratio=2, name="inv_3"
)(x)
x = keras.layers.ReLU()(x)
x = keras.layers.Flatten()(x)
x = keras.layers.Dense(64, activation="relu")(x)
outputs = keras.layers.Dense(10)(x)inv_model = keras.Model(inputs=[inputs], outputs=[outputs], name="inv_model")# Compile the mode with the necessary loss function and optimizer.
print("compiling the involution model...")
inv_model.compile(optimizer="adam",loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),metrics=["accuracy"],
)# train the model
print("inv model training...")
inv_hist = inv_model.fit(train_ds, epochs=20, validation_data=test_ds)

building the involution model...
compiling the involution model...
inv model training...
Epoch 1/20196/196 ━━━━━━━━━━━━━━━━━━━━ 9s 25ms/step - accuracy: 0.1369 - loss: 2.2728 - val_accuracy: 0.2716 - val_loss: 2.1041
Epoch 2/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.2922 - loss: 1.9489 - val_accuracy: 0.3478 - val_loss: 1.8275
Epoch 3/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.3477 - loss: 1.8098 - val_accuracy: 0.3782 - val_loss: 1.7435
Epoch 4/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.3741 - loss: 1.7420 - val_accuracy: 0.3901 - val_loss: 1.6943
Epoch 5/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.3931 - loss: 1.6942 - val_accuracy: 0.4007 - val_loss: 1.6639
Epoch 6/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.4057 - loss: 1.6622 - val_accuracy: 0.4108 - val_loss: 1.6494
Epoch 7/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4134 - loss: 1.6374 - val_accuracy: 0.4202 - val_loss: 1.6363
Epoch 8/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4200 - loss: 1.6166 - val_accuracy: 0.4312 - val_loss: 1.6062
Epoch 9/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.4286 - loss: 1.5949 - val_accuracy: 0.4316 - val_loss: 1.6018
Epoch 10/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.4346 - loss: 1.5794 - val_accuracy: 0.4346 - val_loss: 1.5963
Epoch 11/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4395 - loss: 1.5641 - val_accuracy: 0.4388 - val_loss: 1.5831
Epoch 12/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 5ms/step - accuracy: 0.4445 - loss: 1.5502 - val_accuracy: 0.4443 - val_loss: 1.5826
Epoch 13/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4493 - loss: 1.5391 - val_accuracy: 0.4497 - val_loss: 1.5574
Epoch 14/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4528 - loss: 1.5255 - val_accuracy: 0.4547 - val_loss: 1.5433
Epoch 15/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 4ms/step - accuracy: 0.4575 - loss: 1.5148 - val_accuracy: 0.4548 - val_loss: 1.5438
Epoch 16/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4599 - loss: 1.5072 - val_accuracy: 0.4581 - val_loss: 1.5323
Epoch 17/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4664 - loss: 1.4957 - val_accuracy: 0.4598 - val_loss: 1.5321
Epoch 18/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4701 - loss: 1.4863 - val_accuracy: 0.4575 - val_loss: 1.5302
Epoch 19/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4737 - loss: 1.4790 - val_accuracy: 0.4676 - val_loss: 1.5233
Epoch 20/20196/196 ━━━━━━━━━━━━━━━━━━━━ 1s 6ms/step - accuracy: 0.4771 - loss: 1.4740 - val_accuracy: 0.4719 - val_loss: 1.5096

比较

在本文中，我们将研究这两种模式，并比较一些要点。

参数

我们可以看到，在类似的架构下，CNN 的参数要比 INN（卷积神经网络）的参数大得多。

conv_model.summary()inv_model.summary()

损失图和准确率图

这里的损失图和准确率图表明，INNs 是一种学习速度较慢的学习器（参数较低）。

plt.figure(figsize=(20, 5))plt.subplot(1, 2, 1)
plt.title("Convolution Loss")
plt.plot(conv_hist.history["loss"], label="loss")
plt.plot(conv_hist.history["val_loss"], label="val_loss")
plt.legend()plt.subplot(1, 2, 2)
plt.title("Involution Loss")
plt.plot(inv_hist.history["loss"], label="loss")
plt.plot(inv_hist.history["val_loss"], label="val_loss")
plt.legend()plt.show()plt.figure(figsize=(20, 5))plt.subplot(1, 2, 1)
plt.title("Convolution Accuracy")
plt.plot(conv_hist.history["accuracy"], label="accuracy")
plt.plot(conv_hist.history["val_accuracy"], label="val_accuracy")
plt.legend()plt.subplot(1, 2, 2)
plt.title("Involution Accuracy")
plt.plot(inv_hist.history["accuracy"], label="accuracy")
plt.plot(inv_hist.history["val_accuracy"], label="val_accuracy")
plt.legend()plt.show()

演绎展示：

可视化卷积核

为了使内核可视化，我们取每个内卷内核的 K×K 值之和。不同空间位置上的所有代表都构成了相应的热图。

有人说，"我们提出的内卷让人联想到自我注意，本质上可以成为自我注意的通用版本"。

通过核的可视化，我们确实可以获得图像的注意力图谱。

学习到的内卷化内核为输入张量的各个空间位置提供了注意力。

这种针对特定位置的特性使得内卷成为自我注意模型的通用空间。

layer_names = ["inv_1", "inv_2", "inv_3"]
outputs = [inv_model.get_layer(name).output[1] for name in layer_names]
vis_model = keras.Model(inv_model.input, outputs)fig, axes = plt.subplots(nrows=10, ncols=4, figsize=(10, 30))for ax, test_image in zip(axes, test_images[:10]):(inv1_kernel, inv2_kernel, inv3_kernel) = vis_model.predict(test_image[None, ...])inv1_kernel = tf.reduce_sum(inv1_kernel, axis=[-1, -2, -3])inv2_kernel = tf.reduce_sum(inv2_kernel, axis=[-1, -2, -3])inv3_kernel = tf.reduce_sum(inv3_kernel, axis=[-1, -2, -3])ax[0].imshow(keras.utils.array_to_img(test_image))ax[0].set_title("Input Image")ax[1].imshow(keras.utils.array_to_img(inv1_kernel[0, ..., None]))ax[1].set_title("Involution Kernel 1")ax[2].imshow(keras.utils.array_to_img(inv2_kernel[0, ..., None]))ax[2].set_title("Involution Kernel 2")ax[3].imshow(keras.utils.array_to_img(inv3_kernel[0, ..., None]))ax[3].set_title("Involution Kernel 3")

1/1 ━━━━━━━━━━━━━━━━━━━━ 1s 503ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 11ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 11ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 11ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 10ms/step1/1 ━━━━━━━━━━━━━━━━━━━━ 0s 9ms/step

演绎展示：