NaN loss when training regression network
I have a data matrix in "one-hot encoding" (all ones and zeros) with 260,000 rows and 35 columns. I am using Keras to train a simple neural network to predict a continuous variable. The code to make the network is the following:
model = Sequential() model.add(Dense(1024, input_shape=(n_train,))) model.add(Activation('relu')) model.add(Dropout(0.1)) model.add(Dense(512)) model.add(Activation('relu')) model.add(Dropout(0.1)) model.add(Dense(256)) model.add(Activation('relu')) model.add(Dropout(0.1)) model.add(Dense(1)) sgd = SGD(lr=0.01, nesterov=True); #rms = RMSprop() #model.compile(loss='categorical_crossentropy', optimizer=rms, metrics=['accuracy']) model.compile(loss='mean_absolute_error', optimizer=sgd) model.fit(X_train, Y_train, batch_size=32, nb_epoch=3, verbose=1, validation_data=(X_test,Y_test), callbacks=[EarlyStopping(monitor='val_loss', patience=4)] )
However, during the training process, I see the loss decrease nicely, but during the middle of the second epoch, it goes to nan:
Train on 260000 samples, validate on 64905 samples Epoch 1/3 260000/260000 [==============================] - 254s - loss: 16.2775 - val_loss: 13.4925 Epoch 2/3 88448/260000 [=========>....................] - ETA: 161s - loss: nan
I tried using
RMSProp instead of
SGD, I tried
tanh instead of
relu, I tried with and without dropout, all to no avail. I tried with a smaller model, i.e. with only one hidden layer, and same issue (it becomes nan at a different point). However, it does work with less features, i.e. if there are only 5 columns, and gives quite good predictions. It seems to be there is some kind of overflow, but I can't imagine why--the loss is not unreasonably large at all.
Python version 2.7.11, running on a linux machine, CPU only. I tested it with the latest version of Theano, and I also get Nans, so I tried going to Theano 0.8.2 and have the same problem. With the latest version of Keras has the same problem, and also with the 0.3.2 version.
Regression with neural networks is hard to get working because the output is unbounded, so you are especially prone to the exploding gradients problem (the likely cause of the nans).
Historically, one key solution to exploding gradients was to reduce the learning rate, but with the advent of per-parameter adaptive learning rate algorithms like Adam, you no longer need to set a learning rate to get good performance. There is very little reason to use SGD with momentum anymore unless you're a neural network fiend and know how to tune the learning schedule.
Here are some things you could potentially try:
Normalize your outputs by quantile normalizing or z scoring. To be rigorous, compute this transformation on the training data, not on the entire dataset. For example, with quantile normalization, if an example is in the 60th percentile of the training set, it gets a value of 0.6. (You can also shift the quantile normalized values down by 0.5 so that the 0th percentile is -0.5 and the 100th percentile is +0.5).
Add regularization, either by increasing the dropout rate or adding L1 and L2 penalties to the weights. L1 regularization is analogous to feature selection, and since you said that reducing the number of features to 5 gives good performance, L1 may also.
If these still don't help, reduce the size of your network. This is not always the best idea since it can harm performance, but in your case you have a large number of first-layer neurons (1024) relative to input features (35) so it may help.
Increase the batch size from 32 to 128. 128 is fairly standard and could potentially increase the stability of the optimization.