实现文本数据ResNet遇TypeError与ValueError，寻求解决方法

错误详情 #

1. 初始错误
   运行代码触发 TypeError: __init__() missing 1 required positional argument: 'kernel_size'，报错位置在ResNet50函数的Stage 1中X = Conv1D(64)(X_input)行。

2. 修改后新错误
   将该行改为X = Conv1D(64, kernel_size=(3))(X)后，出现ValueError: Input 0 of layer "conv1d" is incompatible with the layer: expected min_ndim=3, found ndim=2. Full shape received: (None, 53)。

完整实现代码 #

导入依赖 #

from scipy.stats import zscore
import pandas as pd
from sklearn.model_selection import KFold
from keras.layers import Input, Add, Dense, Activation, BatchNormalization, Flatten, Conv1D
from keras.models import Model

CLASSES = 2

def read_pdf():
    df = pd.read_csv("https://data.heatonresearch.com/data/t81-558/jh-simple-dataset.csv", na_values=['NA', '?'])
    return df

def k_fold():
    # 交叉验证
    no_of_folds = int(input('输入K折数量: '))
    kf = KFold(no_of_folds, shuffle=True, random_state=42)  # 用于分类任务的KFold
    return kf

def epochs_fun():
    EPOCHS = int(input('输入训练轮数: '))
    return EPOCHS

数据清洗与K折交叉验证 #

def prepare_data():
    kf = k_fold()
    df = read_pdf()

    # 为job生成独热编码
    df = pd.concat([df, pd.get_dummies(df['job'], prefix="job")], axis=1)
    df.drop('job', axis=1, inplace=True)
    # 为area生成独热编码
    df = pd.concat([df, pd.get_dummies(df['area'], prefix="area")], axis=1)
    df.drop('area', axis=1, inplace=True)
    # 为product生成独热编码
    df = pd.concat([df, pd.get_dummies(df['product'], prefix="product")], axis=1)
    df.drop('product', axis=1, inplace=True)
    # 填充income的缺失值
    med = df['income'].median()
    df['income'] = df['income'].fillna(med)
    # 标准化数值特征
    df['income'] = zscore(df['income'])
    df['aspect'] = zscore(df['aspect'])
    df['save_rate'] = zscore(df['save_rate'])
    df['subscriptions'] = zscore(df['subscriptions'])
    # 转换为numpy数组 - 分类任务
    x_columns = df.columns.drop('age').drop('id')
    x = df[x_columns].values
    y = df['age'].values
    fold = 0
    for train, test in kf.split(x):
        fold += 1
        print(f"第{fold}折")
        x_train = x[train]
        y_train = y[train]
        x_test = x[test]
        y_test = y[test]
        print(x_train)
        print(y_train)
        print(x_test)
        print(y_test)
        return [x_train,y_train,x_test,y_test,x]

恒等模块（Identity Block） #

def identity_block(X, filters):

    # 获取滤波器数量
    F1 = filters

    # 保存输入值，后续需要加回到主路径
    X_shortcut = X

    # 主路径第一部分
    X = Conv1D(filters = F1)(X)
    X = BatchNormalization()(X)
    X = Activation('relu')(X)

    # 主路径第二部分
    X = Conv1D(filters = F1, kernel_size = 3)(X)
    X = BatchNormalization()(X)
    X = Activation('relu')(X)

    # 主路径第三部分
    X = Conv1D()(X)
    X = BatchNormalization()(X)

    # 最后一步：将 shortcut 值与主路径相加，然后通过RELU激活
    X = Add()([X, X_shortcut])
    X = Activation('relu')(X)

    return X

卷积模块（convolutional_block） #

def convolutional_block(X, filters):
    # 获取滤波器数量
    F1 = filters
    ##### 主路径 #####
    # 主路径第一部分
    X = Conv1D(F1)(X)
    X = BatchNormalization()(X)
    X = Activation('relu')(X)

    # 主路径第二部分
    X = Conv1D(filters=F1, kernel_size=(3))(X)
    X = BatchNormalization()(X)
    X = Activation('relu')(X)

    # 主路径第三部分
    X = Conv1D(filters=F1, kernel_size=(3))(X)
    X = BatchNormalization()(X)

    ##### 捷径路径 ####
    X_shortcut = BatchNormalization()

    # 最后一步：将 shortcut 值与主路径相加，然后通过RELU激活
    X = Add()([X, X_shortcut])
    X = Activation('relu')(X)
    return X

ResNet模型定义 #

def ResNet50(input_shape, classes=2):
    # 定义输入张量，形状为input_shape
    X_input = Input(input_shape)

    # Zero-Padding
   # X = ZeroPadding2D()(X_input)

    # Stage 1
    X = Conv1D(64)(X_input)
    X = BatchNormalization()(X)
    X = Activation('relu')(X)

    # Stage 2
    X = convolutional_block(X)
    X = identity_block(X)

    # Stage 3
    X = convolutional_block(X)
    X = identity_block(X)

    # 输出层
    X = Flatten()(X)
    X = Dense(classes, activation='softmax')(X)

    # 创建模型
    model = Model(inputs=X_input, outputs=X, name='ResNet50')

    return model

模型训练 #

def model_check():
    data = prepare_data()
    X_train = data[0]
    Y_train = data[1]
    x = data[4]
    epochs = epochs_fun()
    model = ResNet50(input_shape = x.shape[1], classes = CLASSES)
    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
    model.fit(X_train, Y_train, epochs = epochs, batch_size = 64)

model_check()

错误原因及修复建议 #

1. 第一个错误（缺少kernel_size）
   Conv1D必须指定kernel_size参数，所有未传该参数的Conv1D调用都需要补充，比如：

   • identity_block中X = Conv1D(filters = F1)(X)改为X = Conv1D(filters=F1, kernel_size=3)(X)
   • identity_block中X = Conv1D()(X)改为X = Conv1D(filters=F1, kernel_size=3)(X)
   • convolutional_block中X = Conv1D(F1)(X)改为X = Conv1D(filters=F1, kernel_size=3)(X)

2. 第二个错误（输入维度不匹配）
   Conv1D要求输入为3维张量：(batch_size, timesteps, input_dim)，当前输入是2维(None,53)，需增加维度：

   • 在prepare_data中，将x = df[x_columns].values改为x = df[x_columns].values.reshape(-1, x_columns.shape[0], 1)
   • 同时将ResNet50的input_shape参数改为(x.shape[1], 1)

3. 其他潜在问题

   • convolutional_block中X_shortcut = BatchNormalization()错误，需对输入X做卷积+BN处理，保证维度与主路径输出一致才能相加
   • 当前标签y = df['age'].values是连续值，若为分类任务需将其转为类别，或改用回归损失函数
   • categorical_crossentropy要求标签为独热编码，若使用原始类别标签需改用sparse_categorical_crossentropy

内容的提问来源于stack exchange，提问作者Python