Python 谷歌应用程序引擎的损失并没有减少，但Jupyter笔记本电脑的损失却减少了

我在Google App Engine和Jupyter笔记本上运行相同的代码行和相同的源文件：

model = load_model("test.h5")
model.compile(optimizer=Adam(lr=1e-2, decay=0), loss="binary_crossentropy", metrics=['accuracy'])   
with open("data.json", 'r') as f:
    data = json.load(f)
X = data[0]
y = data[1]
history = model.fit(X, y, validation_split=0, epochs=50, batch_size=10)

GAE的输出如下：

Epoch 1/50
2/2 [==============================] - 1s 316ms/step - loss: 8.0590 - acc: 0.5000
Epoch 2/50
2/2 [==============================] - 0s 50ms/step - loss: 8.0590 - acc: 0.5000
Epoch 3/50
2/2 [==============================] - 0s 40ms/step - loss: 8.0590 - acc: 0.5000
Epoch 4/50
2/2 [==============================] - 0s 37ms/step - loss: 8.0590 - acc: 0.5000
Epoch 5/50
2/2 [==============================] - 0s 34ms/step - loss: 8.0590 - acc: 0.5000
Epoch 6/50
2/2 [==============================] - 0s 40ms/step - loss: 8.0590 - acc: 0.5000
Epoch 7/50
2/2 [==============================] - 0s 44ms/step - loss: 8.0590 - acc: 0.5000
Epoch 8/50
2/2 [==============================] - 0s 40ms/step - loss: 8.0590 - acc: 0.5000
Epoch 9/50
2/2 [==============================] - 0s 31ms/step - loss: 8.0590 - acc: 0.5000
Epoch 10/50
2/2 [==============================] - 0s 40ms/step - loss: 8.0590 - acc: 0.5000
...
Epoch 50/50
2/2 [==============================] - 0s 45ms/step - loss: 8.0590 - acc: 0.5000

而Jupyter笔记本是：

Epoch 1/50
2/2 [==============================] - 0s 164ms/step - loss: 952036.8125 - accuracy: 0.5000
Epoch 2/50
2/2 [==============================] - 0s 39ms/step - loss: 393826.0000 - accuracy: 0.5000
Epoch 3/50
2/2 [==============================] - 0s 38ms/step - loss: 99708.9375 - accuracy: 0.5000
Epoch 4/50
2/2 [==============================] - 0s 39ms/step - loss: 8989.7822 - accuracy: 0.5000
Epoch 5/50
2/2 [==============================] - 0s 39ms/step - loss: 8760.8223 - accuracy: 0.5000
Epoch 6/50
2/2 [==============================] - 0s 40ms/step - loss: 3034.8613 - accuracy: 0.5000
Epoch 7/50
2/2 [==============================] - 0s 40ms/step - loss: 167.2695 - accuracy: 0.0000e+00
Epoch 8/50
2/2 [==============================] - 0s 39ms/step - loss: 0.6670 - accuracy: 1.0000
Epoch 9/50
2/2 [==============================] - 0s 41ms/step - loss: 0.6619 - accuracy: 1.0000
Epoch 10/50
2/2 [==============================] - 0s 40ms/step - loss: 0.6551 - accuracy: 1.0000
...
Epoch 50/50
2/2 [==============================] - 0s 42ms/step - loss: 0.3493 - accuracy: 1.0000

---

为什么会这样？在这一点上我很迷茫。这两台机器都安装了keras==2.2.4和tensorflow==1.14.0。

您运行的是GAE Standard还是GAE Flex？GAE Standard无法运行tensorflow，因此您需要在GAE Flex中使用托管虚拟机您运行的是GAE Standard还是GAE Flex？GAE标准无法运行tensorflow，因此需要在GAE Flex中使用托管虚拟机