Keras 回归模型损失：nan。如何解决？答案

【问题标题】：Keras Regression model loss: nan. How to fix it?Keras 回归模型损失：nan。如何解决？
【发布时间】：2020-12-30 05:31:53
【问题描述】：

我按照“使用 scikit-learn 和 tensorflow 的动手机器学习”一书中的代码在 Keras 中构建了一个多输出神经网络。但是，我一直在亏损：nan 输出。如何解决这个问题？

from sklearn.datasets import fetch_california_housing

housing = fetch_cawwwlifornia_housing()

X_train_full, X_test, y_train_full, y_test = train_test_split(
    housing.data, housing.target)

X_train, X_valid, y_train, y_valid = train_test_split(
    X_train_full, y_train_full)

scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_valid_scaled = scaler.transform(X_valid)
X_test_scaled = scaler.transform(X_test)

X_train_A, X_train_B = X_train[:, :5], X_train[:, 2:]
X_valid_A, X_valid_B = X_valid[:, :5], X_valid[:, 2:]
X_test_A, X_test_B = X_test[:, :5], X_test[:, 2:]
X_new_A, X_new_B = X_test_A[:3], X_test_B[:3]

input_A = keras.layers.Input(shape=[5], name="wide_input")
input_B = keras.layers.Input(shape=[6], name="deep_input")
hidden1 = keras.layers.Dense(30, activation="relu")(input_B)
hidden2 = keras.layers.Dense(30, activation="relu")(hidden1)
concat = keras.layers.concatenate([input_A, hidden2])
output = keras.layers.Dense(1, name="main_output")(concat)
aux_output = keras.layers.Dense(1, name="aux_output")(hidden2)
model = keras.models.Model(inputs=[input_A, input_B],
                           outputs=[output, aux_output])
model.compile(loss=["mse", "mse"], loss_weights=[0.9, 0.1], optimizer="sgd")

history = model.fit(
    [X_train_A, X_train_B], [y_train, y_train], epochs=20,
    validation_data=([X_valid_A, X_valid_B], [y_valid, y_valid]))

输出

Train on 11610 samples, validate on 3870 samples
Epoch 1/20
11610/11610 [==============================] - 6s 525us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 2/20
11610/11610 [==============================] - 4s 336us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 3/20
11610/11610 [==============================] - 5s 428us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 4/20
11610/11610 [==============================] - 5s 424us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 5/20
11610/11610 [==============================] - 5s 414us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 6/20
11610/11610 [==============================] - 5s 400us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 7/20
11610/11610 [==============================] - 5s 392us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 8/20
11610/11610 [==============================] - 5s 405us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 9/20
11610/11610 [==============================] - 4s 369us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 10/20
11610/11610 [==============================] - 5s 405us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 11/20
11610/11610 [==============================] - 5s 423us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 12/20
11610/11610 [==============================] - 5s 454us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 13/20
11610/11610 [==============================] - 4s 380us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 14/20
11610/11610 [==============================] - 5s 446us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 15/20
11610/11610 [==============================] - 5s 411us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 16/20
11610/11610 [==============================] - 5s 457us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 17/20
11610/11610 [==============================] - 5s 415us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 18/20
11610/11610 [==============================] - 5s 411us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 19/20
11610/11610 [==============================] - 5s 388us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan
Epoch 20/20
11610/11610 [==============================] - 4s 363us/sample - loss: nan - main_output_loss: nan - aux_output_loss: nan - val_loss: nan - val_main_output_loss: nan - val_aux_output_loss: nan

【问题讨论】：

你用的是什么优化器和损失函数？
您遗漏了 model.compile 行，这很有趣，可能表明您的问题的原因。使用具有 LR 0.001 和 MSE 损失的 Adam，我可以很好地训练网络。你的学习率可能太高了。
哦，对不起。我确实使用 model.compile 运行，但我只是忘记复制到帖子。
我确实使用了 model.compile，但我只是忘记复制到我的帖子
我试着和Adam一起跑，但是输的还是nan

标签： python machine-learning keras scikit-learn deep-learning

【解决方案1】：

正如评论中所解释的，NaN 通常是由太高的学习率或优化期间类似的不稳定性引起的，这会导致梯度爆炸。这也可以通过设置clipnorm 来防止。设置具有适当学习率的优化器：

opt = keras.optimizers.Adam(0.001, clipnorm=1.)
model.compile(loss=["mse", "mse"], loss_weights=[0.9, 0.1], optimizer=opt)

在您的笔记本中进行更好的培训：

Epoch 1/20
363/363 [==============================] - 1s 2ms/step - loss: 1547.7197 - main_output_loss: 967.1940 - aux_output_loss: 6772.4609 - val_loss: 19.9807 - val_main_output_loss: 20.0967 - val_aux_output_loss: 18.9365
Epoch 2/20
363/363 [==============================] - 1s 2ms/step - loss: 13.2916 - main_output_loss: 14.0150 - aux_output_loss: 6.7812 - val_loss: 14.6868 - val_main_output_loss: 14.5820 - val_aux_output_loss: 15.6298
Epoch 3/20
363/363 [==============================] - 1s 2ms/step - loss: 11.0539 - main_output_loss: 11.6683 - aux_output_loss: 5.5244 - val_loss: 10.5564 - val_main_output_loss: 10.2116 - val_aux_output_loss: 13.6594
Epoch 4/20
363/363 [==============================] - 1s 1ms/step - loss: 7.4646 - main_output_loss: 7.7688 - aux_output_loss: 4.7269 - val_loss: 13.2672 - val_main_output_loss: 11.5239 - val_aux_output_loss: 28.9570
Epoch 5/20
363/363 [==============================] - 1s 2ms/step - loss: 5.6873 - main_output_loss: 5.8091 - aux_output_loss: 4.5909 - val_loss: 5.0464 - val_main_output_loss: 4.5089 - val_aux_output_loss: 9.8839

它的表现并不令人惊讶，但您必须从这里优化所有超参数以使其满意。

您还可以按照最初的意图使用 SGD 来观察 clipnorm 的效果：

opt = keras.optimizers.SGD(0.001, clipnorm=1.)
model.compile(loss=["mse", "mse"], loss_weights=[0.9, 0.1], optimizer=opt)

这可以正常训练。但是，一旦您删除 clipnorm，您将获得 NaNs。

【讨论】：