如何为Keras2（Theano后端）训练的样本分配差异化权重？

Great question! When working with time series data fed into a Conv1D layer—where you can’t just discard low-importance samples—adjusting sample weights during training is exactly the right approach. There are two practical ways to do this in Keras with the Theano backend, depending on your needs:

方法一：使用Keras内置的sample_weight参数（最简洁） #

If your sample weights are fixed (i.e., you know upfront which samples should have lower impact), Keras’ built-in sample_weight parameter is the easiest solution. This lets you pass an array where each value corresponds to a training sample’s weight—higher values mean the sample contributes more to loss calculation, while lower values reduce its influence.

Example Implementation #

import numpy as np
from keras.models import Model
from keras.layers import Input, Conv1D, Dense, Flatten

# 模拟时间序列数据：形状为 (样本数, 时间步长, 特征数)
X = np.random.rand(1000, 50, 10)
y = np.random.randint(0, 2, (1000, 1))  # 二分类标签

# 定义样本权重：前800个为重要样本（权重1），后200个为低重要性样本（权重0.2）
sample_weights = np.ones(1000)
sample_weights[800:] = 0.2

# 构建Conv1D模型
inputs = Input(shape=(50, 10))
x = Conv1D(filters=32, kernel_size=3, activation='relu')(inputs)
x = Flatten()(x)
outputs = Dense(1, activation='sigmoid')(x)

model = Model(inputs=inputs, outputs=outputs)
model.compile(optimizer='adam', loss='binary_crossentropy')

# 训练时传入sample_weight参数
model.fit(
    X, y,
    sample_weight=sample_weights,
    epochs=10,
    batch_size=32
)

This works seamlessly with the Theano backend—Keras handles the weight multiplication under the hood, so you don’t need to modify your loss function at all.

方法二：自定义带样本权重的损失函数（更灵活） #

If you need dynamic weights (e.g., weights calculated from input data itself, or changing during training), create a custom loss function that incorporates sample weights. You’ll add an extra input layer to pass in weights, then scale the loss by these values.

Example Implementation #

import numpy as np
from keras.models import Model
from keras.layers import Input, Conv1D, Dense, Flatten
from keras import backend as K

# 模拟数据
X = np.random.rand(1000, 50, 10)
y = np.random.randint(0, 2, (1000, 1))
sample_weights = np.ones(1000)
sample_weights[800:] = 0.2
sample_weights = sample_weights.reshape(-1, 1)  # 调整形状匹配输入层

# 自定义加权损失函数
def weighted_binary_crossentropy(y_true, y_pred):
    # 获取样本权重（从额外输入层传入）
    weights = K.placeholder(shape=(None, 1))
    # 计算基础二元交叉熵损失
    base_loss = K.binary_crossentropy(y_true, y_pred)
    # 乘以样本权重后取均值
    weighted_loss = K.mean(base_loss * weights)
    return weighted_loss

# 构建包含权重输入的模型
time_series_input = Input(shape=(50, 10))
weight_input = Input(shape=(1,))

x = Conv1D(filters=32, kernel_size=3, activation='relu')(time_series_input)
x = Flatten()(x)
outputs = Dense(1, activation='sigmoid')(x)

# 模型接受两个输入：时间序列数据和样本权重
model = Model(inputs=[time_series_input, weight_input], outputs=outputs)
model.compile(optimizer='adam', loss=weighted_binary_crossentropy)

# 训练时传入两个输入
model.fit(
    [X, sample_weights], y,
    epochs=10,
    batch_size=32
)

Key Notes for Theano Backend #

• Always use Keras backend functions (like K.mean, K.binary_crossentropy) instead of raw NumPy operations—this ensures compatibility with Theano’s computation graph.
• For fully dynamic weights (e.g., based on input features), you can calculate weights directly within the model graph instead of passing them as a separate input.

内容的提问来源于stack exchange，提问作者ascripter