Siamese ネットワークを使用してサンプルの不均衡なデータセットを処理する方法 (サンプルコード付き)

from keras.layers import Input, Conv2D, Lambda, Dense, Flatten, MaxPooling2D
from keras.models import Model
from keras import backend as K
import numpy as np

# 定义输入维度和卷积核大小
input_shape = (224, 224, 3)
kernel_size = 3

# 定义共享的卷积层
conv1 = Conv2D(64, kernel_size, activation='relu', padding='same')
pool1 = MaxPooling2D(pool_size=(2, 2))
conv2 = Conv2D(128, kernel_size, activation='relu', padding='same')
pool2 = MaxPooling2D(pool_size=(2, 2))
conv3 = Conv2D(256, kernel_size, activation='relu', padding='same')
pool3 = MaxPooling2D(pool_size=(2, 2))
conv4 = Conv2D(512, kernel_size, activation='relu', padding='same')
flatten = Flatten()

# 定义共享的全连接层
dense1 = Dense(512, activation='relu')
dense2 = Dense(512, activation='relu')

# 定义距离度量层
def euclidean_distance(vects):
    x, y = vects
    sum_square = K.sum(K.square(x - y), axis=1, keepdims=True)
    return K.sqrt(K.maximum(sum_square, K.epsilon()))

# 定义Siamese网络
input_a = Input(shape=input_shape)
input_b = Input(shape=input_shape)

processed_a = conv1(input_a)
processed_a = pool1(processed_a)
processed_a = conv2(processed_a)
processed_a = pool2(processed_a)
processed_a = conv3(processed_a)
processed_a = pool3(processed_a)
processed_a = conv4(processed_a)
processed_a = flatten(processed_a)
processed_a = dense1(processed_a)
processed_a = dense2(processed_a)

processed_b = conv1(input_b)
processed_b = pool1(processed_b)
processed_b = conv2(processed_b)
processed_b = pool2(processed_b)
processed_b = conv3(processed_b)
processed_b = pool3(processed_b)
processed_b = conv4(processed_b)
processed_b = flatten(processed_b)
processed_b = dense1(processed_b)
processed_b = dense2(processed_b)

distance = Lambda(euclidean_distance)([processed_a, processed_b])

model = Model([input_a, input_b], distance)

# 定义加权损失函数
def weighted_binary_crossentropy(y_true, y_pred):
    class1_weight = K.variable(1.0)
    class2_weight = K.variable(1.0)
    class1_mask = K.cast(K.equal(y_true, 0), 'float32')
    class2_mask = K.cast(K.equal(y_true, 1), 'float32')
    class1_loss = class1_weight * K.binary_crossentropy(y_true, y_pred) * class1_mask
    class2_loss = class2_weight * K.binary_crossentropy(y_true, y_pred) * class2_mask
    return K.mean(class1_loss + class2_loss)

# 编译模型，使用加权损失函数和Adam优化器
model.compile(loss=weighted_binary_crossentropy, optimizer='adam')

# 训练模型
model.fit([X_train[:, 0], X_train[:, 1]], y_train, batch_size=32, epochs=10, validation_data=([X_val[:, 0], X_val[:, 1]], y_val))