tensorflow加入Dropout后loss计算不准确？是不是踩坑了啊，被整蒙圈了啊。如下文，我在全连接里加了一层layers.Dropout, 经过40次Epoch后，loss已经收敛得很小。但是此时无论是predict、还是直接call出来的结果都还是很真实值相差挺多，相差比例1-dropout_rate有那么一点相近，这个按理是不应该的啊，tf默认是inverted dropout啊传播时自动做了scale。只能怀疑时loss计算不准确了？

运行结果及报错内容

>>> model0 = tf.keras.Sequential()
>>> model0.add(layers.Dense(
...     32,
...     activation='relu',
...     kernel_initializer=initializers_v2.HeUniform()))
>>> model0.add(layers.Dense(
...     512,
...     activation='relu',
...     kernel_initializer=initializers_v2.HeUniform()))
>>> model0.add(layers.Dropout(0.5))
>>> model0.add(layers.Dense(
...     512,
...     activation='relu',
...     kernel_initializer=initializers_v2.HeUniform()))
>>> model0.add(layers.Dense(
...     1,
...     kernel_initializer=initializers_v2.HeUniform()))
>>> model0.compile(
...     optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005),
...     loss=tf.keras.losses.MeanAbsoluteError(),
...     metrics=['mae'])
>>> model0.fit(
...     X_a, y_value_a,
...     epochs=40)
Epoch 40/40
625/625 [==============================] - 1s 2ms/step - loss: 0.2190 - mae: 0.2190

预测值/真实值 相差比例跟1-dropout_rate有那么一点相近

>>> y_pred_a = model0.predict(X_a)
625/625 [==============================] - 1s 826us/step
>>> y_call_a = model0.call(tf.convert_to_tensor(X_a))
>>> print(y_pred_a)
[[-13.342887 ]
 [ -5.501849 ]
 [ -6.9179006]
 ...
 [ 19.558184 ]
 [  1.6202304]
 [ -6.4136333]]
>>> print(y_call_a)
tf.Tensor(
[[-13.342106 ]
 [ -5.502385 ]
 [ -6.9180675]
 ...
 [ 19.557985 ]
 [  1.6198621]
 [ -6.413244 ]], shape=(20000, 1), dtype=float32)
>>> print(y_value_a)
[[-26.56738428]
 [-10.90707993]
 [-13.71867984]
 ...
 [ 38.34531367]
 [  3.16246066]
 [-12.71669086]]

这里手动计算mae、调用model的loss\evaluate计算的loss相同，但是跟metrics里的loss相差很多，更奇怪的是当evaluate完之后，再次metrics得到loss又变了，是不是踩到大坑了啊

```python
>>> model0.metrics[0].result().numpy()
0.21902849
>>> loss_a = y_value_a - y_pred_a
>>> abs(loss_a).mean()
7.838831122853866
>>> model0.metrics[0].result().numpy()
0.21902849
>>> model0.loss(y_value_a, y_pred_a)
.Tensor: shape=(), dtype=float32, numpy=7.8388314>
>>> model0.metrics[0].result().numpy()
0.21902849
>>> model0.evaluate(X_a,y_value_a)
625/625 [==============================] - 1s 961us/step - loss: 7.8388 - mae: 7.8388
[7.838833808898926, 7.838833808898926]
>>> model0.metrics[0].result().numpy()
7.838834

```