Keras model 总是预测 0

Question

所以我有类似(10000, 178, 178, 3)形状的数据，其中我有 10000 个样本，每个都有 3 个不同的颜色通道（不是 RGB 通道），其中我有大约8k个样本，其中 label 0和 Z65E8800B5C6800AAD896F888B2AkA62 为label 1 。 这是我的示例数据之一：

array([[[[1.79844797e-01, 1.73587397e-01, 1.73587397e-01, ...,
          4.84393053e-02, 5.15680127e-02, 5.46967126e-02],
         [1.76716089e-01, 1.79844797e-01, 1.82973504e-01, ...,
          5.15680127e-02, 5.31323589e-02, 5.15680127e-02],
         [1.81409150e-01, 1.86102197e-01, 1.81409150e-01, ...,
          5.15680127e-02, 5.31323589e-02, 5.15680127e-02]]],


       [[[2.51065755e+00, 2.53197193e+00, 2.53197193e+00, ...,
          1.88543844e+00, 1.89964795e+00, 1.90675282e+00],
         [2.51776242e+00, 2.52486706e+00, 2.53197193e+00, ...,
          1.89964795e+00, 1.90675282e+00, 1.90675282e+00],
         [2.53197193e+00, 2.51776242e+00, 2.52486706e+00, ...,
          1.91385746e+00, 1.90675282e+00, 1.90675282e+00]]],


       [[[7.13270283e+00, 7.11016369e+00, 7.13270283e+00, ...,
          4.85625362e+00, 4.90133190e+00, 4.94641018e+00],
         [7.08762503e+00, 7.08762503e+00, 7.08762503e+00, ...,
          4.92387104e+00, 4.96894932e+00, 4.96894932e+00],
         [7.08762503e+00, 7.08762503e+00, 7.06508589e+00, ...,
          4.99148846e+00, 4.96894932e+00, 4.96894932e+00]]],
      dtype=float32)

现在首先我试图通过颜色通道进行标准化。 由于每个颜色通道完全不同，所以我按如下方式按颜色通道进行归一化， dara_array是我的整个数据集：

def nan(index):
    data_array[:, :, :, index] = (data_array[:, :, :, index] - np.min(data_array[:, :, :, index]))/(np.max(data_array[:, :, :, index]) - np.min(data_array[:, :, : ,index]))
    

拆分训练和测试：

rand_indices = np.random.permutation(len(data))
train_indices = rand_indices[0:7460]
valid_indices = rand_indices[7460:len(data)]

x_test = data_array[valid_indices, :]
y_test = EDR[[valid_indices]].astype('float')

x_train = data_array[train_indices, :]
y_train = EDR[[train_indices]].astype('float')

然后我使用这个神经网络来训练这个数据集：

weight_decay = 1e-4
model = Sequential()
model.add(Conv2D(32, (20,20), padding='same', kernel_regularizer=regularizers.l2(weight_decay), input_shape=x_tr.shape[1:]))
model.add(LeakyReLU(alpha=0.01))
model.add(BatchNormalization())
model.add(Conv2D(32, (30,30), padding='same', kernel_regularizer=regularizers.l2(weight_decay)))
model.add(LeakyReLU(alpha=0.01))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Dropout(0.2))

model.add(Flatten())
model.add(Dense(1, activation='sigmoid'))

model.summary()

然后我在这里训练它：

def lr_schedule(epoch):
    lrate = 0.001
    if epoch > 75:
        lrate = 0.0005
    elif epoch > 100:
        lrate = 0.0003        
    return lrate

batch_size = 128

opt_rms = tf.keras.optimizers.Adam()

model.compile(loss= 'binary_crossentropy', optimizer = opt_rms, metrics=['accuracy', tf.keras.metrics.Precision(), tf.keras.metrics.Recall()])

history = model.fit(x_train, y_train, batch_size, epochs=10, verbose=1,
                   callbacks=[LearningRateScheduler(lr_schedule)])

这是我所有时代的结果：

Epoch 1/10
59/59 [==============================] - 420s 7s/step - loss: 6.7227 - accuracy: 0.7263 - precision_2: 0.2697 - recall_2: 0.2846 - lr: 0.0010
Epoch 2/10
59/59 [==============================] - 399s 7s/step - loss: 2.7919 - accuracy: 0.7440 - precision_2: 0.3027 - recall_2: 0.2991 - lr: 0.0010
Epoch 3/10
59/59 [==============================] - 399s 7s/step - loss: 2.9244 - accuracy: 0.7484 - precision_2: 0.3210 - recall_2: 0.3282 - lr: 0.0010
Epoch 4/10
59/59 [==============================] - 399s 7s/step - loss: 3.5013 - accuracy: 0.7509 - precision_2: 0.3246 - recall_2: 0.3261 - lr: 0.0010
Epoch 5/10
59/59 [==============================] - 398s 7s/step - loss: 3.1829 - accuracy: 0.7413 - precision_2: 0.3137 - recall_2: 0.3406 - lr: 0.0010
Epoch 6/10
59/59 [==============================] - 398s 7s/step - loss: 4.9515 - accuracy: 0.7592 - precision_2: 0.3307 - recall_2: 0.2999 - lr: 0.0010
Epoch 7/10
59/59 [==============================] - 398s 7s/step - loss: 2.3082 - accuracy: 0.7613 - precision_2: 0.3539 - recall_2: 0.3588 - lr: 0.0010
Epoch 8/10
59/59 [==============================] - 399s 7s/step - loss: 1.8624 - accuracy: 0.7520 - precision_2: 0.3273 - recall_2: 0.3282 - lr: 0.0010
Epoch 9/10
59/59 [==============================] - 398s 7s/step - loss: 2.7749 - accuracy: 0.7579 - precision_2: 0.3344 - recall_2: 0.3173 - lr: 0.0010
Epoch 10/10
59/59 [==============================] - 399s 7s/step - loss: 2.5800 - accuracy: 0.7513 - precision_2: 0.3288 - recall_2: 0.3362 - lr: 0.0010

现在，当我打印分类报告时，一切都将变为0 ：

y_pred = model.predict(x_test)
y_pred_bool = np.argmax(y_pred, axis=1)

print(classification_report(y_test, y_pred_bool))

Output：

    precision    recall  f1-score   support

         0.0       0.82      1.00      0.90      2030
         1.0       0.00      0.00      0.00       453

    accuracy                           0.82      2483
   macro avg       0.41      0.50      0.45      2483
weighted avg       0.67      0.82      0.74      2483

有人可以告诉我我做错了什么或我错过了什么，我在规范化数据或训练时做错了什么，或者我的 model 有什么问题吗？

这是我的数据中的一张示例图片：

Answer 1

您在单个 output 上使用 argmax，因此它只会预测 class 0。您可以通过分析 y_pred 变量或通过切换 ZA2F2ED4F8EBC2CBB4C21A29DC40 标签并看到 it1DZ6 标签来看到这种情况。

要解决此问题，请使用阈值而不是 argmax 或更改 model 以预测每个 class 1 output。 有一些方法可以优化这一点，但对于初始测试，您可以使用 0.5 左右的阈值，如下所示：

T=0.5
y_pred = model.predict(x_test)
y_pred_bool = y_pred>=T

print(classification_report(y_test, y_pred_bool))