Model training diverges after some level ? about keras HOT 6 CLOSED

Could you post the model? This may be a hyperparameter issue.

from keras.

Yes, could you post a script to repro this (I assume you used the Kaggle Otto challenge data). It looks like an exploding gradient, and there are many things that could trigger one (most likely a learning rate issue?).

In general when faced with this type of problem, you should try different learning rates and learning rate decay, different optimizers, and try to introduce regularization techniques such as L1 regularization (supported by all optimizers in Keras), gradient clipping, BatchNormalization, etc.

from keras.

model = Sequential()
model.add(Dense(20, 64, init='uniform', activation='tanh'))
model.add(Dropout(0.5))
model.add(Dense(64, 64, init='uniform', activation='tanh'))
model.add(Dropout(0.5))
model.add(Dense(64, 1, init='uniform', activation='softmax')

sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='mean_squared_error', optimizer=sgd)

I cannt post whole the script since I overwrote it but this is the same model you give as an example. Only I use the data from some where else.

However, the strange thing is, it goes smoothly to some degree but suddenly it diverges. It is the first time I observed such a thing. I also used Adagrad and it works nice but this is strange anyway.

from keras.

Also I see that time to time for Adagrad , model stucks to very high loss from the beginning but it goes well for different runs.

from keras.

I guess the learning rate becomes negative.
https://github.com/fchollet/keras/blob/master/keras/optimizers.py#L48

from keras.

yeah perfectly rational then it should be corrected.

from keras.