我正在使用神经网络解决二元分类问题，但遇到了一些麻烦。有时在运行我的模型时，我的验证准确性根本没有改变，有时它工作得很好。我的数据集有 1200 个样本和 28 个特征，我有一个类不平衡(200 类 a 1000 类 b)。我所有的特征都被归一化并且在 1 和 0 之间。正如我之前所说，这个问题并不总是发生，但我想知道原因并修复它
我曾尝试更改优化函数和激活函数，但这对我没有好处。我还注意到，当我增加网络中的神经元数量时，这个问题发生的频率会降低，但并没有解决。我也尝试增加 epoch 的数量，但问题有时会不断发生
model = Sequential()
model.add(Dense(28, input_dim=28,kernel_initializer='normal', activation='sigmoid'))
model.add(Dense(200, kernel_initializer='normal',activation='sigmoid'))
model.add(Dropout(0.5))
model.add(Dense(300, kernel_initializer='normal',activation='sigmoid'))
model.add(Dropout(0.5))
model.add(Dense(300, kernel_initializer='normal',activation='sigmoid'))
model.add(Dropout(0.5))
model.add(Dense(150, kernel_initializer='normal',activation='sigmoid'))
model.add(Dropout(0.4))
model.add(Dense(1,kernel_initializer='normal'))
model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
history = model.fit(X_train, y_train,
epochs=34,
batch_size=32,
validation_data=(X_val, y_val),
verbose=1)
这是我有时从训练模型中得到的结果
Epoch 1/34
788/788 [==============================] - 1s 2ms/step - loss: 1.5705 - acc: 0.6865 - val_loss: 0.6346 - val_acc: 0.7783
Epoch 2/34
788/788 [==============================] - 0s 211us/step - loss: 1.0262 - acc: 0.6231 - val_loss: 0.5310 - val_acc: 0.7783
Epoch 3/34
788/788 [==============================] - 0s 194us/step - loss: 1.7575 - acc: 0.7221 - val_loss: 0.5431 - val_acc: 0.7783
Epoch 4/34
788/788 [==============================] - 0s 218us/step - loss: 0.9113 - acc: 0.5774 - val_loss: 0.5685 - val_acc: 0.7783
Epoch 5/34
788/788 [==============================] - 0s 199us/step - loss: 1.0987 - acc: 0.6688 - val_loss: 0.6435 - val_acc: 0.7783
Epoch 6/34
788/788 [==============================] - 0s 201us/step - loss: 0.9777 - acc: 0.5343 - val_loss: 0.5643 - val_acc: 0.7783
Epoch 7/34
788/788 [==============================] - 0s 204us/step - loss: 1.0603 - acc: 0.5914 - val_loss: 0.6266 - val_acc: 0.7783
Epoch 8/34
788/788 [==============================] - 0s 197us/step - loss: 0.7580 - acc: 0.5939 - val_loss: 0.6615 - val_acc: 0.7783
Epoch 9/34
788/788 [==============================] - 0s 206us/step - loss: 0.8950 - acc: 0.6650 - val_loss: 0.5291 - val_acc: 0.7783
Epoch 10/34
788/788 [==============================] - 0s 230us/step - loss: 0.8114 - acc: 0.6701 - val_loss: 0.5428 - val_acc: 0.7783
Epoch 11/34
788/788 [==============================] - 0s 281us/step - loss: 0.7235 - acc: 0.6624 - val_loss: 0.5275 - val_acc: 0.7783
Epoch 12/34
788/788 [==============================] - 0s 264us/step - loss: 0.7237 - acc: 0.6485 - val_loss: 0.5473 - val_acc: 0.7783
Epoch 13/34
788/788 [==============================] - 0s 213us/step - loss: 0.6902 - acc: 0.7056 - val_loss: 0.5265 - val_acc: 0.7783
Epoch 14/34
788/788 [==============================] - 0s 217us/step - loss: 0.6726 - acc: 0.7145 - val_loss: 0.5285 - val_acc: 0.7783
Epoch 15/34
788/788 [==============================] - 0s 197us/step - loss: 0.6656 - acc: 0.7132 - val_loss: 0.5354 - val_acc: 0.7783
Epoch 16/34
788/788 [==============================] - 0s 216us/step - loss: 0.6083 - acc: 0.7259 - val_loss: 0.5262 - val_acc: 0.7783
Epoch 17/34
788/788 [==============================] - 0s 218us/step - loss: 0.6188 - acc: 0.7310 - val_loss: 0.5271 - val_acc: 0.7783
Epoch 18/34
788/788 [==============================] - 0s 210us/step - loss: 0.6642 - acc: 0.6142 - val_loss: 0.5676 - val_acc: 0.7783
Epoch 19/34
788/788 [==============================] - 0s 200us/step - loss: 0.6017 - acc: 0.7221 - val_loss: 0.5256 - val_acc: 0.7783
Epoch 20/34
788/788 [==============================] - 0s 209us/step - loss: 0.6188 - acc: 0.7157 - val_loss: 0.8090 - val_acc: 0.2217
Epoch 21/34
788/788 [==============================] - 0s 201us/step - loss: 1.1724 - acc: 0.4061 - val_loss: 0.5448 - val_acc: 0.7783
Epoch 22/34
788/788 [==============================] - 0s 205us/step - loss: 0.5724 - acc: 0.7424 - val_loss: 0.5293 - val_acc: 0.7783
Epoch 23/34
788/788 [==============================] - 0s 234us/step - loss: 0.5829 - acc: 0.7538 - val_loss: 0.5274 - val_acc: 0.7783
Epoch 24/34
788/788 [==============================] - 0s 209us/step - loss: 0.5815 - acc: 0.7525 - val_loss: 0.5274 - val_acc: 0.7783
Epoch 25/34
788/788 [==============================] - 0s 220us/step - loss: 0.5688 - acc: 0.7576 - val_loss: 0.5274 - val_acc: 0.7783
Epoch 26/34
788/788 [==============================] - 0s 210us/step - loss: 0.5715 - acc: 0.7525 - val_loss: 0.5273 - val_acc: 0.7783
Epoch 27/34
788/788 [==============================] - 0s 206us/step - loss: 0.5584 - acc: 0.7576 - val_loss: 0.5274 - val_acc: 0.7783
Epoch 28/34
788/788 [==============================] - 0s 215us/step - loss: 0.5728 - acc: 0.7563 - val_loss: 0.5272 - val_acc: 0.7783
Epoch 29/34
788/788 [==============================] - 0s 281us/step - loss: 0.5735 - acc: 0.7576 - val_loss: 0.5275 - val_acc: 0.7783
Epoch 30/34
788/788 [==============================] - 0s 272us/step - loss: 0.5773 - acc: 0.7614 - val_loss: 0.5272 - val_acc: 0.7783
Epoch 31/34
788/788 [==============================] - 0s 225us/step - loss: 0.5847 - acc: 0.7525 - val_loss: 0.5272 - val_acc: 0.7783
Epoch 32/34
788/788 [==============================] - 0s 239us/step - loss: 0.5739 - acc: 0.7551 - val_loss: 0.5272 - val_acc: 0.7783
Epoch 33/34
788/788 [==============================] - 0s 216us/step - loss: 0.5632 - acc: 0.7525 - val_loss: 0.5269 - val_acc: 0.7783
Epoch 34/34
788/788 [==============================] - 0s 240us/step - loss: 0.5672 - acc: 0.7576 - val_loss: 0.5267 - val_acc: 0.7783
请您参考如下方法:
鉴于您报告的类别不平衡，您的模型似乎没有学到任何东西(报告的准确性似乎与仅将所有内容预测为多数类别一致)。但是，您的代码存在问题；对于初学者:
- 将除输出层之外的所有激活函数替换为
activation = 'relu'。 - 在最后一层
activation='sigmoid'添加一个 sigmoid 激活函数；事实上，你的是一个回归网络(最后一层的默认线性输出)并且不是分类网络。 - 从所有层中删除所有
kernel_initializer='normal'参数，即将其留给 default一个kernel_initializer='glorot_uniform'，众所周知，它可以实现(好得多)更好的性能。
另外，不清楚为什么要使用 28 个单元的输入密集层 - 不。这里的单位与输入维度无关；请参阅Keras Sequential model input layer .
默认情况下，Dropout 不应进入网络 - 先尝试不使用它，然后在必要时添加。
总而言之，这是您的模型应该如何寻找初学者:
model = Sequential()
model.add(Dense(200, input_dim=28, activation='relu'))
# model.add(Dropout(0.5))
model.add(Dense(300, activation='relu'))
# model.add(Dropout(0.5))
model.add(Dense(300, activation='relu'))
# model.add(Dropout(0.5))
model.add(Dense(150, activation='relu'))
# model.add(Dropout(0.4))
model.add(Dense(1, activation='sigmoid'))
model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
并且，如前所述，根据您的实验结果取消注释/调整丢失层。
