Lstm With Varying K-hot Encoded Vector

Followup question from: LSTM with keras In this example a one hot encoded vector is used to perform classification using an LSTM. How could this LSTM be used to perform k-hot encod

Solution 1:

This is a multiclass classification task. In order to solve that you need to:

1. Set your output activation to sigmoid:

   model.add(Dense(150, activation='sigmoid'))
   

2. Set your targets to indicator encoding:

   If you e.g. 4 classes and for a given example set classes 0 and 2 your output should be [1, 0, 1, 0]

3. Use the following loss:

   import keras.backend as K
   
   def multiclass_loss(y_true, y_pred):
       EPS =1e-5
       y_pred = K.clip(y_pred, EPS, 1 - EPS)
       return -K.mean((1 - y_true) * K.log(1 - y_pred) + y_true * K.log(y_pred))
   
   model.compile(optimizer=..., loss=multiclass_loss)