python - from_logits=True keras model auc=0.9013 当应用于model.predict tf.activation.sigmoid auc=0.8183 为什么？

Question

为什么 AUC 不同？以及如何将 model.predict 转换为 AUC = 0.9013 的标签

model = keras.Sequential([
    keras.layers.Dense(4096, activation='relu'),
    keras.layers.Reshape((16,256)),
    keras.layers.Conv1D(64,2),
    keras.layers.Flatten(),
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(1)
])
        
    model.compile(optimizer='RMSprop',
                    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
                    metrics=tf.keras.metrics.AUC(from_logits=True))
    
    model.fit(X, Y, epochs=20, batch_size=64, validation_split=0.15)

Epoch 20/20
129/129 [==============================] - 3s 21ms/step - loss: 0.2583 - auc_1:  0.9585 - val_loss: 0.4556 - val_auc_1: 0.9075
    model.evaluate(Xt,Yt,batch_size=64)

75/75 [==============================] - 0s 6ms/step - loss: 0.4977 - auc_1: 0.9013

[0.4976920783519745, 0.9013347029685974]

pred = model.predict(Xt, batch_size=32, verbose=1)
print(pred)

149/149 [==============================] - 1s 8ms/step
[[-13.761446 ]
 [  1.0361784]
 [  0.439698 ]
 ...
 [  9.102252 ]
 [  2.4544861]
 [  0.8873279]]
    
     

    m = tf.keras.metrics.AUC(from_logits=True)  
    m.update_state(y_test, pred)
    m.result()

  

<tf.Tensor: shape=(), dtype=float32, numpy=0.9013347>

    b = tf.keras.activations.sigmoid(pred)
    b.numpy()
    x = []
    for xser in b.numpy():
      if xser > 0.5: 
        print (xser, '        1')
        x.append(1)
      else:
        print (xser, '        0')
        x.append(0)
    k = tf.keras.metrics.AUC(from_logits=False)
    k.update_state(y_test, x)
    k.result()

<tf.Tensor: shape=(), dtype=float32, numpy=0.81831676>

为什么 AUC 不同？以及如何将 model.predict 转换为 AUC = 0.9013 的标签