CNN二分类模型全预测为类别0问题排查求助

我正在为大学课程编写首个CNN模型，用于图像二分类任务（类别0或1）。图像存储于data/data路径，训练标签在train_labels.txt中，对应前15000张图像；接下来2000张为验证集，标签在validation_labels.txt；最后5149张图像需预测类别。目前模型输出全部为类别0（该类别为优势类别），已尝试调整epoch数、批次大小、学习率等参数，无任何改善，推测问题出在模型本身，寻求解决方向。

代码示例 #

import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import os
from glob import glob
from sklearn.utils import class_weight
import tensorflow.keras.callbacks as callbacks

train_labels = pd.read_csv("data/train_labels.txt", delimiter=",", header=None, names=["id", "class"], skiprows=1)
val_labels = pd.read_csv("data/validation_labels.txt", delimiter=",", header=None, names=["id", "class"], skiprows=1)

train_labels['id'] = train_labels['id'].apply(lambda x: '{0:0>6}.png'.format(x))
val_labels['id'] = val_labels['id'].apply(lambda x: '{0:0>6}.png'.format(x))


train_labels["class"] = train_labels["class"].astype(str)
val_labels["class"] = val_labels["class"].astype(str)

train_datagen = ImageDataGenerator(rescale=1./255)
val_datagen = ImageDataGenerator(rescale=1./255)
test_datagen = ImageDataGenerator(rescale=1./255)

test_dir = "data/data/"
test_ids = range(17001, 22150)

train_generator = train_datagen.flow_from_dataframe(
    train_labels,
    directory="data/data/",
    x_col="id",
    y_col="class",
    target_size=(224, 224),
    batch_size=32,
    class_mode="binary")

val_generator = val_datagen.flow_from_dataframe(
    val_labels,
    directory="data/data/",
    x_col="id",
    y_col="class",
    target_size=(224, 224),
    batch_size=32,
    class_mode="binary")

test_generator = test_datagen.flow_from_dataframe(
    dataframe=pd.DataFrame({"id": [f"{i:06d}.png" for i in test_ids]}),
    directory=test_dir,
    x_col="id",
    y_col=None,
    target_size=(224, 224),
    batch_size=1,
    class_mode=None,
    shuffle=False)

model = tf.keras.Sequential([
    layers.experimental.preprocessing.Rescaling(scale=1./255),
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(224, 224, 3)),
    layers.MaxPooling2D((2, 2)),
    layers.Conv2D(64, (3, 3), activation='relu'),
    layers.MaxPooling2D((2, 2)),
    layers.Conv2D(128, (3, 3), activation='relu'),
    layers.MaxPooling2D((2, 2)),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dense(1, activation='sigmoid')
])

optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)

model.compile(optimizer=optimizer,
              loss='binary_crossentropy',
              metrics=['accuracy'])

class_weights = class_weight.compute_class_weight(class_weight='balanced', classes=np.unique(train_labels['class']), y=train_labels['class'])
class_weights_dict = dict(enumerate(class_weights))

history = model.fit(train_generator,
                    steps_per_epoch=train_generator.samples//train_generator.batch_size,
                    epochs=10,
                    validation_data=val_generator,
                    validation_steps=val_generator.samples//val_generator.batch_size,
                    class_weight=class_weights_dict,
                    )


model.evaluate(val_generator)

preds = model.predict(test_generator)
preds = np.where(preds > 0.5, 1, 0)
output_df = pd.DataFrame({'id': range(17001, 22150), 'class': preds.flatten()})
output_df.to_csv("data/submission.csv", sep=',', index=False, header=False)

1. 修复重复归一化问题 #

代码中存在两次图像归一化：

• ImageDataGenerator已设置rescale=1./255，将像素从0-255缩放到0-1
• 模型第一层又添加layers.experimental.preprocessing.Rescaling(scale=1./255)，导致像素被二次缩放至0-0.0039，严重破坏特征分布
  建议移除模型中的Rescaling层，保留ImageDataGenerator的归一化即可。

2. 确认类别权重映射正确性 #

class_weight.compute_class_weight返回的权重顺序基于np.unique(train_labels['class'])，而flow_from_dataframe会按类别字符串字典序分配索引（比如"0"对应索引0，"1"对应索引1）。需：

• 打印train_generator.class_indices查看类别-索引映射
• 确保class_weights_dict的键与该映射一致，避免权重分配错误

3. 强化模型特征学习能力 #

当前模型结构较浅，不足以提取有效图像特征：

• 添加更多卷积层（比如增加Conv2D(256, (3,3), activation='relu')+MaxPooling2D）
• 在全连接层后添加Dropout(0.5)抑制过拟合，增强泛化能力
• 考虑使用预训练模型（如MobileNetV2、VGG16）进行迁移学习，利用预训练图像特征提升分类性能

4. 优化训练策略与监控 #

• 增加训练轮数（30-50轮），同时添加EarlyStopping回调，当验证集性能不再提升时自动停止，避免欠拟合
• 监控精确率、召回率、F1分数（而非仅准确率），类别不平衡场景下准确率无法反映真实性能：

  from sklearn.metrics import classification_report
  val_preds = model.predict(val_generator)
  val_preds = np.where(val_preds > 0.5, 1, 0)
  print(classification_report(val_labels['class'].astype(int), val_preds))
  

• 降低学习率至1e-4，避免权重更新幅度过大导致模型无法收敛到少数类

5. 验证数据加载正确性 #

• 随机抽取少量训练图像，检查train_generator加载的图像与标签是否匹配，排除数据加载错误
• 确认验证集和训练集的类别分布是否符合预期，避免数据划分错误

内容的提问来源于stack exchange，提问作者Scooby