TensorFlow医学二输入ML模型数据基数模糊错误求助

触发异常 #

ValueError: Data cardinality is ambiguous:
x sizes: 1, 1
y sizes: 23
Make sure all arrays contain the same number of samples.

背景说明 #

用TensorFlow构建二输入模型，基于MRI光谱的**lactate（乳酸）和lipids（脂质）**指标预测高等级脑肿瘤概率，运行时触发上述维度异常，调整数据塑形后模型准确率仍极差。

相关代码 #

import tensorflow as tf
import numpy as np

# 定义lipidos和lactato输入层
lipidos_input = tf.keras.layers.Input(shape=(23,), name="lipidos_input")
lactato_input = tf.keras.layers.Input(shape=(23,), name="lactato_input")

# lipidos分支隐藏层
lipidos_hidden1 = tf.keras.layers.Dense(10, activation='relu')(lipidos_input)
lipidos_hidden2 = tf.keras.layers.Dense(10, activation='relu')(lipidos_hidden1)

# lactato分支隐藏层
lactato_hidden1 = tf.keras.layers.Dense(10, activation='relu')(lactato_input)
lactato_hidden2 = tf.keras.layers.Dense(10, activation='relu')(lactato_hidden1)

# 合并两个分支输出
merged = tf.keras.layers.concatenate([lipidos_hidden2, lactato_hidden2])

# 额外隐藏层
hidden_layer = tf.keras.layers.Dense(10, activation='relu')(merged)

# 二分类输出层（sigmoid激活）
output_layer = tf.keras.layers.Dense(1, activation='sigmoid', name="output")(hidden_layer)

# 创建模型
model = tf.keras.models.Model(inputs=[lipidos_input, lactato_input], outputs=output_layer)

# 编译模型
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

# 输入数据
lipidos_data = np.array([
    0.016, 0.03, 0.05, 0.0008, 0.009, 0.01, 0.015,
    0.15, 0.004, 0.016, 0.09, 0.003, 0.001, 0.003,
    0.012, 0.01, 0.011, 0.01, 0.003, 0.02, 0.03,
    0.001, 0
], dtype=float)

lactato_data = np.array([
    0.015, 0.021, 0.043, 0.0014, 0.011, 0.01, 0.015,
    0.03, 0.02, 0.012, 0.075, 0.0025, 0.005, 0.01,
    0.028, 0.011, 0.11, 0.032, 0.0045, 0.006, 0.005,
    0, 0.002
], dtype=float)

alto_grado_data = np.array([
    1,1,1,1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,0,1,1,0,0
], dtype=int)

# 将输入数据转为2D数组
lipidos_data = lipidos_data.reshape(-1, 23)
lactato_data = lactato_data.reshape(-1, 23)

# 训练模型
model.fit([lipidos_data, lactato_data], alto_grado_data, epochs=100, batch_size=8)

# 预测示例
lipidos_value = np.array([[0.015]], dtype=float)
lactato_value = np.array([[0.015]], dtype=float)

prediction = model.predict([lipidos_value, lactato_value])
predicted_alto_grado = round(prediction[0][0])
print("Predicted alto_grado:", predicted_alto_grado)

1. 修复数据维度不匹配 #

当前核心错误是样本数不匹配：

• 你把lipidos_data和lactato_datareshape成了(1,23)（1个样本，23个特征），但标签alto_grado_data是23个样本，两者数量必须一致。
• 正确逻辑是：23个样本，每个样本包含1个lipid特征和1个lactate特征。

修正代码：

# 调整输入数据维度：23个样本，每个样本1个特征
lipidos_data = lipidos_data.reshape(-1, 1)
lactato_data = lactato_data.reshape(-1, 1)

# 同步调整输入层shape为(1,)
lipidos_input = tf.keras.layers.Input(shape=(1,), name="lipidos_input")
lactato_input = tf.keras.layers.Input(shape=(1,), name="lactato_input")

2. 解决类别不平衡问题 #

标签数据中18个正样本、5个负样本，严重的类别不平衡会导致模型偏向预测多数类，看似准确率高但泛化能力差：

• 给少数类设置更高权重：

class_weight = {0: 18/5, 1: 1}  # 负样本权重是正样本的3.6倍
model.fit([lipidos_data, lactato_data], alto_grado_data, epochs=100, batch_size=8, class_weight=class_weight)

• 改用更适合不平衡数据的评估指标：

from tensorflow.keras.metrics import AUC, Precision, Recall
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy', AUC(name='auc'), Precision(), Recall()])

3. 优化模型与训练流程 #

样本量仅23个，原模型结构过复杂易过拟合：

• 简化模型结构：

lipidos_input = tf.keras.layers.Input(shape=(1,), name="lipidos_input")
lactato_input = tf.keras.layers.Input(shape=(1,), name="lactato_input")
merged = tf.keras.layers.concatenate([lipidos_input, lactato_input])
hidden = tf.keras.layers.Dense(4, activation='relu')(merged)
output = tf.keras.layers.Dense(1, activation='sigmoid')(hidden)
model = tf.keras.models.Model(inputs=[lipidos_input, lactato_input], outputs=output)

• 加入正则化防止过拟合：

from tensorflow.keras.layers import Dropout
from tensorflow.keras.regularizers import l2

hidden = tf.keras.layers.Dense(4, activation='relu', kernel_regularizer=l2(0.01))(merged)
hidden = Dropout(0.2)(hidden)

• 划分训练/验证集监控泛化能力：

from sklearn.model_selection import train_test_split

X_lipid_train, X_lipid_val, X_lactate_train, X_lactate_val, y_train, y_val = train_test_split(
    lipidos_data, lactato_data, alto_grado_data, test_size=0.2, random_state=42, stratify=alto_grado_data
)

history = model.fit(
    [X_lipid_train, X_lactate_train], y_train,
    epochs=100, batch_size=4,
    validation_data=([X_lipid_val, X_lactate_val], y_val),
    class_weight=class_weight
)

4. 修正预测输入 #

当前预测输入的(1,1)维度符合修正后的输入层要求，无需调整，但要确保输入值对应单个样本的特征。

内容的提问来源于stack exchange，提问作者FFalco