训练1D CNN时model.fit()出现形状不兼容错误求助

尝试在表格数据上训练1D CNN，后续用于LRP分析，但在model.fit()环节报错，此前步骤均正常运行。

报错信息 #

model_history = model.fit(X_train_smote, y_train_smote, batch_size=100, epochs=100, validation_split = 0.2)

Epoch 1/100
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [60], in <cell line: 1>()
----> 1 model_history = model.fit(X_train_smote, y_train_smote, batch_size=100, epochs=100, validation_split = 0.2)

File D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\utils\traceback_utils.py:67, in filter_traceback.<locals>.error_handler(*args, **kwargs)
     65 except Exception as e:  # pylint: disable=broad-except
     66   filtered_tb = _process_traceback_frames(e.__traceback__)
---> 67   raise e.with_traceback(filtered_tb) from None
     68 finally:
     69   del filtered_tb

File ~\AppData\Local\Temp\__autograph_generated_filelswqetod.py:15, in outer_factory.<locals>.inner_factory.<locals>.tf__train_function(iterator)
     13 try:
     14     do_return = True
---> 15     retval_ = ag__.converted_call(ag__.ld(step_function), (ag__.ld(self), ag__.ld(iterator)), None, fscope)
     16 except:
     17     do_return = False

ValueError: in user code:

    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\engine\training.py", line 1051, in train_function  *
        return step_function(self, iterator)
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\engine\training.py", line 1040, in step_function  **
        outputs = model.distribute_strategy.run(run_step, args=(data,))
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\engine\training.py", line 1030, in run_step  **
        outputs = model.train_step(data)
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\engine\training.py", line 890, in train_step
        loss = self.compute_loss(x, y, y_pred, sample_weight)
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\engine\training.py", line 948, in compute_loss
        return self.compiled_loss(
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\engine\compile_utils.py", line 201, in __call__
        loss_value = loss_obj(y_t, y_p, sample_weight=sw)
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\losses.py", line 139, in __call__
        losses = call_fn(y_true, y_pred)
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\losses.py", line 243, in call  **
        return ag_fn(y_true, y_pred, **self._fn_kwargs)
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\losses.py", line 1787, in categorical_crossentropy
        return backend.categorical_crossentropy(
    File "D:\Programme\Anaconda\envs\LRP_innvestigate\lib\site-packages\keras\backend.py", line 5119, in categorical_crossentropy
        target.shape.assert_is_compatible_with(output.shape)

    ValueError: Shapes (None,) and (None, 6, 2) are incompatible

完整实现代码 #

import matplotlib.pyplot as plt
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from matplotlib.colors import ListedColormap
from sklearn.metrics import plot_confusion_matrix
from scipy.stats import norm, boxcox
from sklearn.metrics import confusion_matrix, classification_report, accuracy_score
from collections import Counter
from scipy import stats
import tensorflow as tf
import warnings
warnings.simplefilter(action='ignore', category=Warning)

dataset = pd.read_csv('F:/Programmieren/this_one/data/Churn_Modelling.csv')

# split into variables and target
X = dataset.iloc[:, 3:-1].values
y = dataset.iloc[:, -1].values

# here gender is encoded
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
X[:, 2] = le.fit_transform(X[:, 2])

# one hot encoding the country (as explained in Ullah et al.)
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
ct = ColumnTransformer(transformers=[('encoder', OneHotEncoder(), [1])], remainder='passthrough')
X = np.array(ct.fit_transform(X))

# split the data
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size= .2)

# upsample minority class with SMOTE 
import imblearn
from imblearn.over_sampling import SMOTENC

#get original class distribution
counter = Counter(y)
print(counter)

#SMOTENC is used instead of SMOTE because there are multiple categorical features in the dataset
oversample = SMOTENC(categorical_features=[0, 1, 2, 4, 9, 10])
X_train_smote, y_train_smote = oversample.fit_resample(X_train, y_train)

#get new class distribution
counter = Counter(y_train_smote)
print(counter)

# normalize values to range 0-1 as explained in Ullah et al.

from sklearn.preprocessing import MinMaxScaler
mms = MinMaxScaler()
X_train_smote = mms.fit_transform(X_train_smote)
X_test = mms.transform(X_test)

#record-wise normalization for relative value comparison as stated by one of the authors I was in contact with
from sklearn.preprocessing import normalize

X_train_smote = normalize(X_train_smote, axis=1, norm='l1')
X_test = normalize(X_test, axis=1, norm='l1')

#reshape data for CNN
sample_size = X_train_smote.shape[0] # number of samples in train set
time_steps  = X_train_smote.shape[1] # number of features in train set
input_dimension = 1 

train_data_reshaped = X_train_smote.reshape(sample_size,time_steps,input_dimension)
print("After reshape train data set shape:\n", train_data_reshaped.shape)
print("1 Sample shape:\n",train_data_reshaped[0].shape)
print("An example sample:\n", train_data_reshaped[0])

#reshape test data as well
test_data_reshaped = X_test.reshape(X_test.shape[0],X_test.shape[1],1)

import keras
from keras.models import Sequential
from keras.layers import Dense, Conv1D

#create model as explained in the paper
model = Sequential()
model.add(Conv1D(filters=25, kernel_size=3, activation='relu', input_shape=(12,1)))
model.add(Conv1D(50, 3))
model.add(Conv1D(100, 3))
model.add(Dense(2200, activation='relu'))
model.add(Dense(2, activation='relu'))
model.add(Dense(2, activation='softmax'))

model.compile(optimizer = 'adam', loss = 'categorical_crossentropy', metrics = ['accuracy'])
model.summary()

#output of  model.summary()
Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv1d (Conv1D)             (None, 10, 25)            100       
                                                                 
 conv1d_1 (Conv1D)           (None, 8, 50)             3800      
                                                                 
 conv1d_2 (Conv1D)           (None, 6, 100)            15100     
                                                                 
 dense (Dense)               (None, 6, 2200)           222200    
                                                                 
 dense_1 (Dense)             (None, 6, 2)              4402      
                                                                 
 dense_2 (Dense)             (None, 6, 2)              6         
                                                                 
=================================================================
Total params: 245,608
Trainable params: 245,608
Non-trainable params: 0
_________________________________________________________________

请问当前方法是否存在未察觉的重大缺陷？

核心错误原因 #

报错Shapes (None,) and (None, 6, 2) are incompatible明确指出两个关键问题：

1. 模型输出形状不符合二分类任务要求：从模型summary可见，最终输出是(None,6,2)，即每个样本输出6组二元分类结果，但你的任务是单样本二分类，仅需1组结果。
2. 标签格式不匹配损失函数要求：y_train_smote是一维数组(None,)，而categorical_crossentropy损失要求标签为独热编码后的二维数组(None,2)。

具体修正方案 #

1. 调整模型结构，压缩序列维度 #

在Conv1D层之后添加全局池化层或Flatten层，将序列维度转为一维，再连接全连接层输出最终分类结果：

model = Sequential()
model.add(Conv1D(filters=25, kernel_size=3, activation='relu', input_shape=(12,1)))
model.add(Conv1D(50, 3, activation='relu'))
model.add(Conv1D(100, 3, activation='relu'))
# 全局平均池化，将(None,6,100)转为(None,100)
model.add(tf.keras.layers.GlobalAveragePooling1D())
# 或使用Flatten层：model.add(tf.keras.layers.Flatten())
model.add(Dense(2200, activation='relu'))
model.add(Dense(2, activation='relu'))
model.add(Dense(2, activation='softmax'))

调整后模型最终输出变为(None,2)，符合二分类任务要求。

2. 对标签进行独热编码 #

使用tf.keras.utils.to_categorical处理训练和测试标签：

from tensorflow.keras.utils import to_categorical

y_train_smote = to_categorical(y_train_smote, num_classes=2)
y_test = to_categorical(y_test, num_classes=2)

3. 传入正确的训练数据 #

训练时需传入reshape后的train_data_reshaped，而非原始的X_train_smote：

model_history = model.fit(train_data_reshaped, y_train_smote, batch_size=100, epochs=100, validation_split = 0.2)

其他潜在问题 #

1. Conv1D层缺少激活函数：后两个Conv1D层未指定激活函数，默认线性激活会降低模型表达能力，建议添加activation='relu'。
2. 全连接层参数冗余：Dense(2200)参数数量过大，易导致过拟合，可根据样本量调整为256或512。
3. LRP适配性：调整模型结构后，需确认后续LRP工具兼容修改后的标准CNN+全连接结构（主流LRP库均支持此类结构）。

内容的提问来源于stack exchange，提问作者user19534996