Keras 回归模型损失：nan。如何修复它？

Question

我一直遭受损失：nan 输出。如何解决这个问题？from sklearn.datasets import fetch_california_housinghousing = 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 samplesEpoch 1/2011610/11610 [==============================] - 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: nanEpoch 2/2011610/11610 [==============================] - 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: nanEpoch 3/20

Answer 1

NaN 通常是由学习率过高或优化过程中类似的不稳定性引起的，从而导致梯度爆炸。这也可以通过设置来防止clipnorm。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)

可以在笔记本上进行更好的训练：

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才能将其调整到满意的程度。

您还可以按照您最初的预期使用 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)

这样训练得当。但是，一旦删除clipnorm，您就会得到NaNs。