自定义 Conv2D 层的卷积操作
作者： lukewood
创建日期 11/03/2021
最后修改日期 11/03/2021
描述：此示例展示了如何使用 Conv.convolution_op() API 实现自定义卷积层。
简介
有时您可能需要实现 Conv1D 和 Conv2D 等卷积层的自定义版本。Keras 允许您在不从头开始实现整个层的情况下做到这一点：您可以重用大部分基本卷积层，并仅通过 convolution_op() 方法自定义卷积操作本身。
此方法是在 Keras 2.7 中引入的。因此，在使用 convolution_op() API 之前，请确保您正在运行 Keras 2.7.0 或更高版本。
简单的 StandardizedConv2D 实现
有两种方法可以使用 Conv.convolution_op() API。第一种方法是在卷积层子类上覆盖 convolution_op() 方法。使用这种方法，我们可以快速实现一个 StandardizedConv2D，如下所示。
import os
os.environ["KERAS_BACKEND"] = "tensorflow"
import tensorflow as tf
import keras
from keras import layers
import numpy as np
class StandardizedConv2DWithOverride(layers.Conv2D):
def convolution_op(self, inputs, kernel):
mean, var = tf.nn.moments(kernel, axes=[0, 1, 2], keepdims=True)
return tf.nn.conv2d(
inputs,
(kernel - mean) / tf.sqrt(var + 1e-10),
padding="VALID",
strides=list(self.strides),
name=self.__class__.__name__,
)
使用 Conv.convolution_op() API 的另一种方法是从卷积层子类的 call() 方法中直接调用 convolution_op() 方法。下面显示了使用此方法实现的可比类。
class StandardizedConv2DWithCall(layers.Conv2D):
def call(self, inputs):
mean, var = tf.nn.moments(self.kernel, axes=[0, 1, 2], keepdims=True)
result = self.convolution_op(
inputs, (self.kernel - mean) / tf.sqrt(var + 1e-10)
)
if self.use_bias:
result = result + self.bias
return result
示例用法
这两个层都可以作为 Conv2D 的替代品。以下演示在 MNIST 数据集上执行分类。
# Model / data parameters
num_classes = 10
input_shape = (28, 28, 1)
# the data, split between train and test sets
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
# Scale images to the [0, 1] range
x_train = x_train.astype("float32") / 255
x_test = x_test.astype("float32") / 255
# Make sure images have shape (28, 28, 1)
x_train = np.expand_dims(x_train, -1)
x_test = np.expand_dims(x_test, -1)
print("x_train shape:", x_train.shape)
print(x_train.shape[0], "train samples")
print(x_test.shape[0], "test samples")
# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
model = keras.Sequential(
[
keras.layers.Input(shape=input_shape),
StandardizedConv2DWithCall(32, kernel_size=(3, 3), activation="relu"),
layers.MaxPooling2D(pool_size=(2, 2)),
StandardizedConv2DWithOverride(64, kernel_size=(3, 3), activation="relu"),
layers.MaxPooling2D(pool_size=(2, 2)),
layers.Flatten(),
layers.Dropout(0.5),
layers.Dense(num_classes, activation="softmax"),
]
)
model.summary()
x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
Model: "sequential"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ ┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩ │ standardized_conv2d_with_call │ (None, 26, 26, 32) │ 320 │ │ (StandardizedConv2DWithCall) │ │ │ ├─────────────────────────────────┼───────────────────────────┼────────────┤ │ max_pooling2d (MaxPooling2D) │ (None, 13, 13, 32) │ 0 │ ├─────────────────────────────────┼───────────────────────────┼────────────┤ │ standardized_conv2d_with_overr… │ (None, 11, 11, 64) │ 18,496 │ │ (StandardizedConv2DWithOverrid… │ │ │ ├─────────────────────────────────┼───────────────────────────┼────────────┤ │ max_pooling2d_1 (MaxPooling2D) │ (None, 5, 5, 64) │ 0 │ ├─────────────────────────────────┼───────────────────────────┼────────────┤ │ flatten (Flatten) │ (None, 1600) │ 0 │ ├─────────────────────────────────┼───────────────────────────┼────────────┤ │ dropout (Dropout) │ (None, 1600) │ 0 │ ├─────────────────────────────────┼───────────────────────────┼────────────┤ │ dense (Dense) │ (None, 10) │ 16,010 │ └─────────────────────────────────┴───────────────────────────┴────────────┘
Total params: 34,826 (136.04 KB)
Trainable params: 34,826 (136.04 KB)
Non-trainable params: 0 (0.00 B)
batch_size = 128
epochs = 5
model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
model.fit(x_train, y_train, batch_size=batch_size, epochs=5, validation_split=0.1)
Epoch 1/5
64/422 ━━━[37m━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.4439 - loss: 13.1274
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1699557098.952525 26800 device_compiler.h:187] Compiled cluster using XLA! This line is logged at most once for the lifetime of the process.
422/422 ━━━━━━━━━━━━━━━━━━━━ 10s 14ms/step - accuracy: 0.7277 - loss: 4.5649 - val_accuracy: 0.9690 - val_loss: 0.1140
Epoch 2/5
422/422 ━━━━━━━━━━━━━━━━━━━━ 2s 3ms/step - accuracy: 0.9311 - loss: 0.2493 - val_accuracy: 0.9798 - val_loss: 0.0795
Epoch 3/5
422/422 ━━━━━━━━━━━━━━━━━━━━ 1s 3ms/step - accuracy: 0.9531 - loss: 0.1655 - val_accuracy: 0.9838 - val_loss: 0.0610
Epoch 4/5
422/422 ━━━━━━━━━━━━━━━━━━━━ 1s 3ms/step - accuracy: 0.9652 - loss: 0.1201 - val_accuracy: 0.9847 - val_loss: 0.0577
Epoch 5/5
422/422 ━━━━━━━━━━━━━━━━━━━━ 1s 3ms/step - accuracy: 0.9687 - loss: 0.1059 - val_accuracy: 0.9870 - val_loss: 0.0525
<keras.src.callbacks.history.History at 0x7fed258da200>
结论
Conv.convolution_op() API 提供了一种简单易读的方式来实现自定义卷积层。使用该 API 的 StandardizedConvolution 实现非常简洁，仅包含四行代码。