python - 使用自定义 mape 损失时的 keras nan 损失

Question

我构建了一个简单的 lstm 网络并使用了 costom mape loss，如下所示：

def custom_mape(y_true, y_pred):
    mapes = K.switch(K.equal(y_true, 0), y_true, 100*K.abs(y_true - y_pred)/y_true)
    return K.mean(mapes, axis=-1)

而损失一开始就变成了nan：

Model: "sequential_93"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
lstm_163 (LSTM)              (None, 14, 1)             296       
=================================================================
Total params: 296
Trainable params: 296
Non-trainable params: 0
_________________________________________________________________
Epoch 1/50
410/410 [==============================] - 3s 7ms/step - loss: nan - val_loss: nan
Epoch 2/50
410/410 [==============================] - 2s 6ms/step - loss: nan - val_loss: nan
Epoch 3/50
410/410 [==============================] - 2s 6ms/step - loss: nan - val_loss: nan
Epoch 4/50
410/410 [==============================] - 2s 6ms/step - loss: nan - val_loss: nan
Epoch 5/50
410/410 [==============================] - 2s 6ms/step - loss: nan - val_loss: nan
Epoch 6/50
410/410 [==============================] - 2s 5ms/step - loss: nan - val_loss: nan
Epoch 7/50
410/410 [==============================] - 3s 6ms/step - loss: nan - val_loss: nan
Epoch 8/50
410/410 [==============================] - 2s 5ms/step - loss: nan - val_loss: nan
Epoch 9/50
410/410 [==============================] - 2s 5ms/step - loss: nan - val_loss: nan
Epoch 10/50
410/410 [==============================] - 2s 5ms/step - loss: nan - val_loss: nan

以下是我尝试的一些方法：

1. 当我将 K.abs(y_true - y_pred)/y_true 更改为 K.abs(y_true - y_pred) 时，网络正常工作。
2. 为了判断是不是梯度爆炸，我分别尝试了clipvalue=1、lr=0和batchsize=1。损失仍然是nan。

此外，我对 y 使用了 min-max 归一化，y 的样本如下：

[[1.84368752e-05],
[9.86574098e-04],
[8.09853832e-04]]

标签： pythontensorflowkeras

Answer 1

K.abs(y_true - y_pred)/y_true

在这里， if y_trueis 0，你会得到 nan 因为你试图除以0。