Keras model loss function returning nan

Question

I am training an autoencoder in keras which is defined like this:

model = Sequential()
model.add(LSTM(100, activation='relu', input_shape=(430, 3)))
model.add(RepeatVector(430))
model.add(LSTM(100, activation='relu', return_sequences=True))
model.add(TimeDistributed(Dense(3)))
model.compile(optimizer='adam', loss='mean_squared_error', metrics=['accuracy'])
print(model.summary())

context_paths = loadFile()
X_train, X_test = train_test_split(context_paths, test_size=0.20)

print('Fitting model.')
history = model.fit(X_train, X_train, epochs=1, batch_size=8, verbose=1, shuffle=True, validation_data=(X_test, X_test))

predict_sample = X_train[0].reshape((1, 430, 3))

predict_output = model.predict(predict_sample, verbose=0)
print(predict_output[0, :, 0])

This code doesn't give any errors, but when I run it, the loss is nan. I have checked some questions on SO and found that this problem occurs when:

• nan or infinite values are present --> I checked my input data with numpy.isnan(myarray).any() , which returned False , so I also did numpy.isfinite(myarray).any() which returned True , so I assume my data is allright
• batch size is too big --> I reduced from 32 to 8, didn't help much
• layer size is too big --> I reduced from 100 to 24, didn't help much

Here is picture of the first few batches:

Here the loss is gigantic, but I am not sure what is causing it. The range of number in my dataset are reaching the limits of int32. Also my data is padded with 0's.

Answer 1

You clearly have huge range data. You're overflowing everthing, as you yourself observed in your range:

The range of number in my dataset are reaching the limits of int32

Normalize your data before using it in a model.

The correct verification for infinite values shoudl be:

numpy.isfinite(myarray).all()

You can try a transform for a 0 to 1 range (needs to convert to float first):

xMax = x_train.max()
xMin = x_train.min()
xRange = xMax - xMin

x_train = (x_train - xMin) / xRange
x_test = (x_test - xMin) / xRange

Do the same with y.

You could try a Z-transform too:

xMean = x_train.mean()
xStd = x_train.std()

x_train = (x_train - xMean) / xStd
x_test = (x_test - xMean) / xStd